An Optimal Road Network Extraction Methodology for an Autonomous Driving-Based Demand-Responsive Transit Service Considering Operational Design Domains

Abstract

In addition to addressing the labor shortage due to an aging population, the transition to autonomous vehicle (AV)-based mobility services offers enhanced efficiency and operational flexibility for public transportation. However, much of the existing focus has been on improving AV safety without fully considering road conditions and real-world service demand. This study contributes to the literature by proposing a comprehensive framework for efficiently integrating AV-based mobility services at the network level, addressing these gaps. The framework analyzes and optimizes service networks by incorporating actual demand patterns, quantifying road segment difficulty from an AV perspective, and developing an optimization model based on these factors. The framework begins by quantifying the operational difficulty of road segments through an evaluation of Operational Design Domains (ODDs), providing a precise measure of AV suitability under varying road conditions. It then introduces a quantitative metric to assess operational feasibility, considering factors such as the service margin, costs, and safety risks. Using these metrics alongside Genetic Algorithms (GAs), the framework identifies an optimal service network that balances safety, efficiency, and profitability. By analyzing real-world data from different mobility services, such as taxis, Demand-Responsive Transport (DRT), and Special Transportation Services (STSs), this study highlights the need for service-specific strategies to optimize AV deployment. The findings show that optimal networks vary with demand patterns and road difficulty, demonstrating the importance of tailored network designs. This research provides a scalable, data-driven approach for integrating AV services into public transportation systems and lays the foundation for further improvements by incorporating dynamic factors and broader urban contexts.

1. Introduction

Autonomous driving technology has advanced rapidly, improving road safety, reducing traffic congestion, and making transportation more efficient. These changes are transforming industries such as logistics and public transportation. In the field of public transport, AVs (autonomous vehicles) are expected to increase operational efficiency and reduce reliance on human drivers [1,2,3]. However, AVs require an understanding of the Operational Design Domain (ODD), which defines the specific road conditions under which AVs can operate safely.

AVs can only operate on roads that meet certain conditions [4]. Before implementing AV services, road segments must be tested for their suitability for safe autonomous driving. Experts conduct repeated tests to confirm that AVs can operate safely on these sections [5]. AV services begin with small pilot networks. As they expand to larger networks, evaluating only specific routes becomes impractical. For large-scale operations, a shift is needed from route-based evaluations to assessments that cover entire networks with multiple vehicles and routes [6].

As mobility patterns become more complex, there is increasing demand for flexible transport options [7]. These services need to go beyond fixed routes and schedules [8,9,10,11,12,13,14,15]. Demand-responsive mobility services can meet this need by offering flexible routes. Network evaluation methods must evolve to accommodate these flexible routes and various vehicle types [16,17,18]. This study focuses on optimizing AV networks for three transport modes: taxis, Demand-Responsive Transit (DRT), and Special Transportation Services (STSs). Taxis offer peer-to-peer (P2P) services, providing flexibility. DRT pools multiple passengers, improving efficiency. STSs operate within a limited ODD, mainly serving passengers with mobility challenges. The aim is to develop an AV network that meets the demand of these services. Current AV services do not meet real-world urban mobility patterns, causing inefficiencies. Low service utilization rates are common for taxis, DRT, and STSs. To solve this issue, AV services must shift from fixed routes to flexible, demand-driven approaches [18].

In this study, Genetic Algorithms (GAs) are chosen due to their effectiveness at solving complex optimization problems [19]. GAs utilize a population-based approach that allows for simultaneous exploration of diverse solutions, reducing the risk of becoming trapped in local optima. Additionally, GAs continuously improve solutions through evolutionary mechanisms such as selection, crossover, and mutation. By comparison, other metaheuristics such as Simulated Annealing may require careful tuning of temperature schedules, and Particle Swarm Optimization can converge quickly to local optima [20]. Ant Colony Optimization, although beneficial for path optimization, often faces high computational costs, limiting its application to large-scale problems.

Ensuring both safety and operational efficiency is key [15,21,22]. While AVs are constrained by ODD limitations, advancements in technology could expand their operational range. When extracting AV mobility service networks, it is essential to consider demand-reflected ODD operation levels. GAs can assist in optimizing networks by factoring in safety, cost, and demand.

The method we develop follows a three-step approach. First, we create a metric to assess road segment difficulty from an autonomous driving (AD) perspective, considering the relationship between the ODD and the driving environment. This metric quantifies how challenging each road segment is for AV operations, factoring in elements such as road geometry and environmental constraints, which influence AV performance. Second, we develop an evaluation metric for introducing AV-based services to each road segment, incorporating factors such as safety, difficulty levels, costs of implementation and operation, and potential revenue. This phase ensures that not only the technical feasibility, but also economic and safety considerations are accounted for in selecting suitable segments for AV deployment. Finally, we design an algorithm based on these metrics to derive an optimized AV mobility service network. This algorithm identifies a sub-network within the overall urban traffic system, selecting road sections that maximize operational efficiency and profitability for AV-based services, particularly for demand-responsive services, while addressing diverse urban travel patterns.

2. Literature Review

Implementing autonomous mobility services requires understanding the route design and operations of public transport systems. Analyzing their strengths and limitations is crucial for integrating autonomous driving technology [4,5,6,7,23]. Traditional public transportation systems operate primarily on fixed routes and schedules, designed based on predicted traffic demand [24,25,26,27,28,29,30]. Fixed routes enhance operational efficiency and facilitate movement between key locations but fail to meet the individual needs of diverse users [31]. Fixed-route public transportation is difficult to adapt in quickly changing cities with complex traffic. Optimizing routes to reflect actual demand is essential for improving the quality and efficiency of mobility services [8,9,10,11,12,13,14,15,32,33]. However, fixed-route systems lack the flexibility to meet these needs, making new solutions necessary [30].

To overcome the limitations of fixed-route public transportation systems, Demand-Responsive Transit (DRT) was introduced. DRT adjusts routes and schedules based on real-time demand, making the service more flexible. This improves efficiency and ensures coverage even in low-demand areas [17,18]. Unlike traditional systems, which are difficult to adapt to dynamic and diverse transportation needs, DRT offers a more flexible solution. Initially, DRT operated on small networks using mathematical modeling and optimization techniques. However, as cities grew larger and more complex, these simple mathematical methods proved insufficient for optimizing DRT on a large scale [34,35,36]. Optimizing DRT in larger urban networks involves multiple variables and constraints, increasing complexity [17,36]. As such, more advanced optimization methods became necessary.

To address the increasing complexity of DRT networks, Genetic Algorithms (GAs) were introduced as a more powerful alternative to other metaheuristic methods, such as simulated annealing and tabu search [37,38,39,40]. GAs use natural selection and genetics to find optimal solutions, especially in complex networks where other methods falter [40]. GAs can handle many variables and constraints in large networks, making them effective for optimizing DRT routes [20]. The introduction of GA into DRT optimization allows for a more flexible and scalable approach to managing large networks, addressing the limitations of previous mathematical models.

With recent advancements in autonomous driving technology, there has been growing interest in autonomous public transportation services [41]. Autonomous vehicles are expected to bring significant innovation to the transport industry, especially by combining with DRT to improve both service efficiency and safety [42]. However, the current deployment of autonomous vehicles is largely limited to predefined areas [43,44]. This is due to the need to ensure functional safety [45], requiring autonomous vehicles to operate only within their ODD [42]. As a result, services are often confined to areas that have undergone extensive testing, which sets limits on the expansion of autonomous vehicle operations to new regions [41,46].

Developing methods to assess the driving feasibility of autonomous vehicles in advance is essential for expanding service areas and ensuring safe operations [5,46,47,48,49,50,51]. By evaluating road difficulty, it becomes possible to determine whether autonomous vehicles can safely navigate specific roads, aiding in selecting efficient service areas and planning safe operations [48]. Currently, regions are chosen through repeated testing [5,46,50], but this approach is inefficient in terms of time and cost. Sun et al. [52] emphasized the importance of creating a framework for quantitatively evaluating ODDs, which would allow operators to identify suitable regions for autonomous driving in advance and optimize service strategies accordingly.

Despite the growing interest in autonomous mobility services, much of the existing research has been confined to specific road sections, limiting the broader applicability of these findings. Efforts have been made to define the ODD for autonomous driving based on predefined scenarios [52], but research integrating formalized ODD evaluation with traffic demand analysis for mobility services is lacking. Existing studies tend to focus on isolated factors, such as safety metrics or operational constraints. However, they have not considered the complexity of real-world networks and their effect on the feasibility of autonomous driving concerning ODDs. This gap indicates the need for an approach that can address autonomous driving safety, operational efficiency, and traffic demand simultaneously.

This study aims to derive an optimal network for an autonomous mobility service considering traffic demand by using GAs to quantify road difficulty for autonomous vehicles. By combining these factors in GAs, this study allows for a more holistic evaluation of road segments. This approach facilitates the extraction of optimal road networks that ensure safety while also reflecting traffic demand and operational efficiency. As a result, the proposed framework represents a significant advancement in optimizing autonomous mobility services for complex urban environments. Unlike existing studies that are primarily limited to the safety assessment of individual road segments, this study provides a methodology that considers both safety and demand.

3. Methodology

This section outlines a systematic approach to enhance the implementation of AV-based mobility services through detailed analysis and strategic planning. The three subsections each focus on a distinct aspect of network analysis and optimization to ensure that AV systems are deployed efficiently and safely.

Figure 1 outlines the process used in this study. First, the difficulty levels of road segments were measured through expert surveys and classified by ODD. Mobility service characteristics were analyzed to define networks and distribute traffic demand. Next, road suitability was evaluated using formulas that calculate revenue, costs, and safety risks. Finally, GAs were used to find the optimal road network by considering safety, efficiency, and demand. The analysis was conducted using RStudio Version 2024.04.2 Build 764 and ArcGIS Pro Version 2.9.5 to ensure accurate results.

3.1. Measurement of Link Difficulty Levels for Implementation

The ODD defines the limits within which the autonomous driving system is designed to operate and, as such, will only operate when the parameters described within the ODD are satisfied [44]. A total of 282 ODDs were identified and classified into five categories, as defined in [44], encompassing the physical infrastructure, environmental conditions, operational constraints, objects, and zones. The details of the 282 ODDs are given in Appendix A.

We assessed the difficulty level of each of the 282 ODDs from the perspective of safe and stable autonomous driving. To do this, a survey was conducted with professionals from both autonomous driving research institutions and industry sectors. The survey utilized a four-point Likert scale to evaluate the difficulty levels of each ODD. The Likert scale, providing multiple-choice questions to which respondents can record varying degrees of agreement, was considered suitable for this evaluation.

Table 1. Example of assessment outcomes for difficulty level of ODDs (details of the 282 ODDs are given in Appendix A).

Classification of ODD	Attribution	Conditions (Difficulty Level)
Physical Infrastructure	Road type	Central divider (1), One-way street (2), Merge section (1)
	Intersection with flat surface	4-way signal intersection (2), 3-way signal intersection (1), Pocket lane (1), Crosswalk (1)
	Intersection	Trumpet type (2), Diamond type (2), Clover type (2)
	Road pavement	Asphalt, concrete (1); Potholes (3)
	Lane markings	Solid line (1), No lane markings (2)
	Road edge	Bicycle lane (2), Shoulder (1)
	Road structure	Vertical curve (2), Horizontal curve (2), Radius of curvature of 600 m or more (3)
Environmental Conditions	Weather	Beaufort Wind Scale of 8 or higher (3), Light snow (2), Heavy snow (3)
	Road conditions according to weather	Stagnant water on roads (4), Submerged roads (4), Icy roads (4)
	Air pollution and conditions	Clear (1), Yellow dust (2)
	Illumination environment	Sunlight (1), Cloudy night (3)
Operational Constraints	Traffic Conditions	Traffic jam (1)
Objects	Road Users	Ambulances (3), Electric wheelchairs (1), Parked cars (1)
Zones	Geo-fenced Area	Tunnels, bridges (1); Bus stops (2); Hospital entrances (3); Underground, indoor, outdoor parking lots (4)

presents a portion of the assessment outcomes for the difficulty level of ODDs (the comprehensive assessment results are presented in Appendix A). As shown in Table 1, the ODDs are categorized into four levels: Level 1 (Easy), Level 2 (Normal), Level 3 (Hard), and Level 4 (Very Hard). Level 1 (Easy) ODDs represent road conditions that can be adequately dealt with by current autonomous driving technology. This level of ODD encompasses conditions in which autonomous vehicles are capable of effectively tracking lanes marked with solid lines. Level 2 (Normal) ODDs encompass road conditions that are encountered less frequently compared to Level 1, including situations where there are no lane markings. In these conditions, autonomous driving systems utilize more advanced technologies, such as edge-based lane detection, to accurately identify lane boundaries. Level 3 (Hard) ODDs involve dynamic conditions and irregular road features, including the identification of objects such as potholes with a non-uniform structure. Level 4 (Very Hard) ODDs are challenging due to extreme road and environmental conditions, including flooded, icy, and snow-covered roads. For AVs, the ability to navigate a road safely and efficiently is significantly compromised during adverse weather conditions, particularly on snow-covered roads, flooded thoroughfares, and icy surfaces. These environmental factors lead to a pronounced degradation in sensor performance, which is crucial for the navigation and control systems of autonomous vehicles. Sensors such as LiDAR, radar, and cameras, which are integral for detecting objects, lane markings, and road surfaces, often suffer from reduced visibility and accuracy in such conditions. The accumulation of snow or ice on sensors can obscure their lenses and distort the data received, while water from flooded roads can cause refraction errors in laser and radar systems. Consequently, the accuracy of sensor data is markedly reduced, posing challenges for maintaining operational safety and efficacy.

The distribution of ODD levels is crucial for understanding the operational capabilities and limitations of autonomous vehicles. As illustrated in Figure 2, approximately 50% of ODDs are classified as Level 1, while about 10% are Level 4. Despite the high proportion of Level 1 ODDs, the roads feature a mixture of various ODD levels, necessitating a methodology that can effectively evaluate the safety and efficiency of autonomous vehicle operations. Integrating various ODD levels into the evaluation framework will provide a more robust understanding of autonomous vehicle performance and operation. This approach is essential for advancing the safety, efficiency, and reliability of autonomous systems.

As previously mentioned, considering the diverse ODD levels on roads, a formula is necessary to evaluate the levels of autonomous operations for road segments, as outlined in Equation (1):

$$a_i^{link}={\displaystyle \frac{1}{K}}{\displaystyle \sum }_{j=1}^K{a}_{i,j}^{ODD}={\displaystyle \frac{a_{i,1}^{ODD}+a_{i,2}^{ODD}+a_{i,3}^{ODD}+\cdots +a_{i,K}^{ODD}}{K}}$$

(1)

$$x_{operation}^{i,level}=\left\{\begin{array}{c}1, a_i^{link}\le 1 \\ 2, 1<a_i^{link}\le 2 \\ 3, 2<a_i^{link}\le 3\\ 4, 3<a_i^{link}\le 4\end{array}\right.,$$

where $a_i^{link}$ denotes the difficulty level of road segment $i$ for autonomous driving operation; $a_{i,j}^{ODD}$ represents the level of the j-th ODD in road segment i; $K$ represents the total number of ODDs in road segment i; and $x_{operation}^{i,level}$ is defined by an ODD level of road segment i, calculated across all ODD levels within that segment. This operational level facilitates the assessment of autonomous systems’ suitability for specific road segments. The quantification enables strategic adjustments to accommodate the road conditions encountered in AV-based mobility services. This approach ensures that the operational levels, derived from the ODD distribution, align precisely with the requirements for safe and efficient autonomous vehicle operation.

Figure 3 shows a sample road section of the study site with ODD features and difficulty levels. The road includes a sidewalk, an asphalt surface, dotted lane markings, and no median strip. These features are classified as difficulty Level 1. Two-way traffic and a four-way signalized intersection are classified as difficulty Level 2. Using this ODD information, the operational level for each road section is calculated as shown in Equation (1). The section in Figure 3 is classified as operational Level 2 based on the road infrastructure.

3.2. Measurement of Road Suitability for AV-Based Mobility Service

This study introduces a quantitative measure designed to extract the optimal road network to facilitate the introduction and operation of the AV-based mobility service. The proposed measure is defined as the maximization of operational suitability within the context of AV mobility services, as represented by Equation (2).

$$\begin{array}{c}max\\ n\in R^N\end{array}{C}^{suitability}\left(n\right)=C_{ income}^{expected}\left(n\right)-C_{expense}^{expected}\left(n\right)$$

(2)

$$s\mathrm{u}bject to C_{expense}^{expected}\left(n\right)\le B_{Budget}^{suitability}, \forall n\in R^N,$$

where $N$ is the total number of road segments within the study area where AV-based mobility services are to be implemented; $n$ refers to a specific road segment, representing a network link where the feasibility of AV service deployment is being evaluated; $C^{suitability}\left(n\right)$ represents the suitability of introducing AV-based mobility services for road segment $n$, measured in units of Korean won; $C_{ income}^{expected}\left(n\right)$ is the total expected revenue generated from road segment $n$; $C_{expense}^{expected}\left(n\right)$ represents the total expected costs incurred for road segment n, which includes the operational costs, safety costs, and collision risk costs; and $B_{Budget}^{suitability}$ represents the total budget available for the introduction of AV-based mobility services. The calculation method for the expected income, $C_{ income}^{expected}\left(n\right)$, is given in Equation (3):

$$C_{ income}^{excepted}\left(n\right)={{\displaystyle \sum }}_{i=1}^n{{\displaystyle \sum }}_{j=1}^m(({\displaystyle \frac{z_{suitability margin,j}^{service, i}·d^i}{d_j^i}}·{\displaystyle \frac{d_j^i}{d_j^{excluding,i}}})),$$

(3)

where $m$ represents the number of trips that pass through road segment $i$ and $z_{suitability margin,j}^{service, i}$ refers to the net profit per kilometer for trip $j$ passing through road segment $i$. This margin not only incorporates the profitability per kilometer but also accounts for associated costs such as fuel and labor. $d^i$ represents the length of road segment $i$, $d_j^i$ is the shortest total length of trip $j$ that encompasses road segment $i$, and $d_j^{excluding,i}$ represents the total length of trip $j$ when it does not pass through road segment $i$.

$C_{expense}^{excepted}\left(n\right)$ is defined as in Equation (4):

$$C_{expense}^{excepted}\left(n\right)={{\displaystyle \sum }}_{i=1}^n(K_{cost per}^{operation }\left(x_{operation}^{i,level}\right)·d^i)+{{\displaystyle \sum }}_{i=1}^n(P_{safety}^{collision}\left(x_{safety}^{i,pro}\right)·C_{safety}^{sev}\left(x_{safety}^{i,collision}\right)·d^i·S_j^i),$$

(4)

where $C_{expense}^{excepted}\left(n\right)$ is determined for each road segment through the operation cost and risk cost. $K_{cost per}^{implementation}$ represents the operational cost per kilometer, taking into account the difficulty level as previously discussed for $x_{operation}^{i,level}$. $K_{cost per}^{operation}\left(x_{operation}^{i,level}\right)$ details the cost associated with operations at different levels of $x_{operation}^{i,level}$, and this relationship is formalized in Equation (5):

$$K_{cost per}^{operation}\left(x_{operation}^{i,level}\right) =\left\{\begin{array}{c}653,000,000\left(₩/\mathrm{km}\right), x_{operation}^{i,level}=4 \\ 52,860,000\left(₩/\mathrm{km}\right), x_{operation}^{i,level}=3 \\ 35,720,000\left(₩/\mathrm{km}\right), x_{operation}^{i,level}=2\\ 22,860,000\left(₩/\mathrm{km}\right), x_{operation}^{i,level}=1 \end{array}\right.,$$

(5)

where $x_{operation}^{i,level}=1$ is set as a baseline reference point. When the operation levels for AV-based mobility services reach 2, 3, and 4, the costs increase by factors of 2-, 3.3-, and 50-fold, respectively. However, the actual costs can vary depending on the advancement level of autonomous driving technologies and specific regional characteristics. The unit cost is flexible, varying according to the different levels of technological advancement across countries.

$P_{safety}^{collision}\left(x_{safety}^{i,pro}\right)$ represents the probability of a collision and/or accident occurring on road segment $i$, quantified as in Equation (6):

$$P_{safety}^{collision}\left(x_{safety}^{i,pro}\right)=\left\{\begin{array}{c}0.00000910\left({\displaystyle \frac{1}{km·trip}}\right), x_{safety}^{i,pro}=3\\ 0.00000296\left({\displaystyle \frac{1}{km·trip}}\right), x_{safety}^{i,pro}=2\\ 0.00000061\left({\displaystyle \frac{1}{km·trip}}\right), x_{safety}^{i,pro}=1\end{array}\right.,$$

(6)

where $P_{safety}^{collision}\left(x_{safety}^{i,pro}\right)$ is defined as the probability function of collisions and/or accidents occurring on road segment $i$ that offers AV-based mobility services. The probability of collision, $P_{safety}^{collision}$, is derived from data collected by the Department of Motor Vehicles (DMV) in California, USA. This dataset includes autonomous vehicle accidents and the distance traveled per disengagement by each manufacturer, covering the period from 2021 to August 2022. $x_{safety}^{i,pro}$ is defined based on the speed limit of road segment $i$. It is categorized into safety probabilities of 1, 2, and 3, corresponding to the speed limit of [0, 30), [30, 50), and [50, 70) km/h [21,22].

$C_{safety}^{sev}\left(x_{safety}^{i,collision}\right)$ represents the severity of collisions and/or accidents on road segment $i$, where such collisions and/or accidents occur as detailed in Equation (7):

$$C_{safety}^{sev}\left(x_{safety}^{i,collision}\right)=\left\{\begin{array}{c}13,472,000,000\left(₩\right), x_{safety}^{i,collision}=3\\ 76,000,000\left(₩\right), x_{safety}^{i,collision}=2\\ 24,000,000\left(₩\right), x_{safety}^{i,collision}=1\end{array}\right.,$$

(7)

where $C_{safety}^{sev}$ is calculated based on the comprehensive crash unit cost, which quantifies the severity of accidents [15]. $x_{safety}^{i,collision}$ is also defined based on the speed limit of road segment $i$. It is categorized into safety levels of 1, 2, and 3, corresponding to the speed limit ranges of [0, 30), [30, 50), and [50, 70) km/h.

The proposed method enables service providers to assess the profitability of introducing AV-based mobility services across a specific subset of a road network. Consequently, it aims to identify the optimal combination of road segments to maximize overall profits and to derive the effects of introducing AV-based mobility services for each road segment.

3.3. Extraction Method for Optimal Service Network for AV-Based Mobility Service

This study proposes an extraction methodology for an optimal service network for AV-based mobility service. As shown in the previous section, the measurement under consideration addresses multiple dimensions pertinent to AV-based mobility services, encompassing safety, operational efficiency, and revenue generation driven by the demand for mobility services. The proposed algorithm finds the optimal service network that maximizes the proposed measurement described in the previous section from the perspective of the introduction and operation of an AV-based mobility service. The proposed algorithm is designed to extract a subset of the road network for optimal operation of AV-based mobility services within the entire road network considered for the application of mobility services.

The utilization of a GA is proposed to address the intricate combinatorial problem of optimizing the subset of a road network for AV-based mobility services. This choice is justified given the complex nature of the problem, which involves multiple challenging factors such as the extensive number of links in the network, varying user demand, safety considerations for autonomous vehicles, and the operational efficiency of mobility services.

The GA is well suited for this application due to its robust ability to handle large-scale and complex optimization problems [39,40]. The evolutionary approach is adept at exploring a vast search space, which is essential when the network encompasses a considerable number of links. This feature is critical for efficiently navigating the combinatorial aspects of network selection to achieve optimal results [19,53]. Moreover, GAs are particularly advantageous in environments characterized by dynamic and multifaceted criteria. They excel at adapting to varied and sometimes conflicting requirements, such as balancing user demand with the imperative of AV safety and operational efficiency. This adaptability is crucial for optimizing mobility services, where each aspect does not exist in isolation but interacts with others, influencing the overall service efficacy and safety. Therefore, employing a GA not only aligns with the technical demands of the task but also enhances the effectiveness of a solution that comprehensively addresses multiple critical dimensions of AV-based mobility services within a complex road network.

Building on the rationale for employing a GA in optimizing road network selection for AV-based mobility services, this section details the specific methodology proposed. The approach adheres to the foundational framework of GA, which provides a structured yet adaptable method for tackling optimization problems.

Input Parameters: The input to the algorithm involves identifying which road links from the entire network should be considered for use. This selection is critical as it sets the parameters within which the GA operates, defining the potential solutions to be evaluated. In this study, a binary encoding GA was used, wherein each road segment is represented as a bit, indicating its inclusion or exclusion from the network. The number of bits corresponds to the total number of road segments being evaluated for the AV network configuration.

Output Metrics: The output from the algorithm is derived using the previously mentioned measurements, which quantify aspects such as safety, operational efficiency, and user demand fulfillment. These metrics are incorporated into the fitness function, which evaluates each candidate network configuration based on its ability to balance safety, cost-effectiveness, and demand satisfaction. Higher fitness scores are assigned to configurations that optimize these factors, guiding the algorithm toward more efficient solutions.

Optimization Process: The GA follows the classic evolutionary steps of selection, crossover, and mutation. The process starts with the initialization of a population of randomly generated network configurations, each encoded as a binary string where each bit represents the presence or absence of a road segment. The selection process identifies the best-performing configurations based on their fitness scores, ensuring that higher-quality solutions have a greater chance of being passed to the next generation.

In the crossover step, pairs of selected parent configurations exchange bits at random crossover points to create new offspring, introducing new combinations of road segments. Mutation is then applied by randomly flipping bits in the offspring solutions, introducing diversity into the population and preventing premature convergence to suboptimal solutions. This process ensures that the algorithm can explore a wide variety of possible configurations before narrowing down to the optimal one.

The GA’s iterative process is designed to converge on a configuration of road links that optimally balances safety, operational efficiency, and user demand. By continuously exploring and refining combinations of road links, the GA can systematically search the solution space and narrow down to the most effective network structure. This methodology leverages the GA’s strengths in handling complex optimization problems, ensuring a thorough exploration of feasible solutions and leading to a robust and efficient road network configuration for AV-based mobility services.

4. Analysis of Mobility Characteristics in the Study Site

Sejong City, a meticulously planned urban center situated in the heart of South Korea, has emerged as a pioneering smart city specializing in AV technologies. Spanning an area of 456 square kilometers and housing approximately 390,000 residents, Sejong is divided into six functional living zones, each designed to optimize the urban experience by using cutting-edge Information and Communication Technology (ICT). The first living zone, which encompasses the administrative districts of Go-un-dong, Areum-dong, Jongchon-dong, Dodam-dong, and Eojin-dong, is a focal point for advanced mobility solutions (see Figure 4). This area is not only central to Sejong’s administrative functions but also serves as a dynamic hub for the trial and deployment of various autonomous mobility services. These services are underpinned by sophisticated infrastructural technologies, including Smart City features and Cooperative Intelligent Transport Systems (C-ITSs), which enhance communication between AVs and road infrastructure.

Sejong’s commitment to innovation is further highlighted by its zero-energy community strategy, which aligns with environmental sustainability goals while supporting high-tech urban management solutions. The city has designated specific geo-fenced zones within these districts where AV technologies are actively tested and used, serving diverse needs such as DRT, Special Transportation Services (STSs), and taxis.

This paper aims to optimize the introduction and operation of AV-based mobility services by considering user demand in Sejong City. To achieve this, this study focuses on analyzing the actual usage patterns of mobility services, drawing on comprehensive mobility data characteristics from the city. Sejong City, recognized for its smart urban planning, provides a rich dataset from various mobility services including taxis, DRT, and STSs. These data are crucial for understanding the travel dynamics and service demand within the city. Our analysis utilized service histories, particularly focusing on origin–destination records, to map out the necessary networks for efficient mobility service. By focusing on Sejong City’s mobility service data, this paper seeks to provide a grounded analysis of the current mobility framework and proposes pathways for the integration of AV technologies.

The examination of travel data from taxis, DRT, and STSs in Sejong City provides crucial insights into the feasibility of autonomous vehicle (AV) deployment in urban mobility services. This analysis, supported by the data in

Table 2. Travel features of three mobility services.

Mobility Service		Taxi	DRT	STS
Total Distance of Service (km)		263.855	1102.726	253.353
Average Distance per Service (km)		1.534	2.293	1.964
Travel Frequency by links	max	30	46	58
	average	4.746	9.778	5.955

, details the total travel distances, average distances per service, and frequency of travel per link for each mobility service.

Our findings indicated that DRT services encompass the longest total travel distances, with an average distance per service amounting to approximately 2.3 km. This suggests that DRT operations cover more extensive areas within the city, catering to broader spatial demand. In contrast, both taxi services and STSs exhibit similar average travel distances, which highlights a uniformity in service delivery in terms of distance covered per trip. Moreover, the assignment of demand to specific network links reveals differing usage intensities among the services. The average link usage frequency for DRT stands at 9.778, the highest among the services, which underscores its broad coverage and frequent utilization of numerous links within the transport network. On the other hand, STSs and taxis show lower average frequencies, with values of 5.955 and 4.746, respectively. The data suggest a less intensive use of the network compared to DRT. Notably, the maximum link usage frequency observed is for STSs, peaking at 58. This indicates that, while STSs may have lower average link usage, there are certain routes where its demand is highly concentrated, possibly due to specialized transport needs such as accessibility services. Such concentrated demand in specific sections of the road network for STSs and DRT compared to taxis implies a more targeted and possibly predictable pattern of service deployment, which could be advantageous for integrating AV solutions.

Using mobility service data, the spatial distribution of the travel patterns for taxis, DRT, and STSs was analyzed, as illustrated in Figure 5, Figure 6 and Figure 7. For taxis, out of a total of 1656 links in the study area, 563 were utilized to provide 172 service trips. This shows a moderate use of the available networks, indicating a dispersed pattern of service distribution across the city. In the case of DRT, a more extensive utilization of the network was evident. Specifically, out of the total 1656 links, 1115 were utilized, providing 481 service trips. This extensive coverage demonstrates DRT’s role in catering to a broader range of areas within the city, suggesting its crucial impact on urban mobility. Conversely, STSs utilized only 330 links, representing about 20% of the total networks, to accommodate 129 service trips. The usage patterns depicted in Figure 5 show that STSs are densely concentrated in certain areas, reinforcing the high demand intensity on specific routes. This concentration likely reflects the targeted nature of STSs, which are designed to meet specific mobility needs, such as accessibility for individuals with limited mobility.

The results presented in Figure 5 corroborate the earlier analysis, revealing the distinct travel patterns and network utilizations of taxis, DRT, and STSs. This figure effectively visualizes the differences in how each service type strategically uses the urban network, which is critical for informing future strategies for AV integration and the enhancement of mobility services in Sejong City.

5. Results

Figure 8 provides a detailed visualization of the road link difficulty levels across the study site, illustrating the varying challenges involved in planning an AV-based mobility service. The map categorizes road links into four difficulty levels, ranging from Level 1 (Easy) to Level 4 (Very Hard), based on an assessment derived from Equation (1). Level 1 (Easy) road links comprise 56.5% of the network, indicating that many parts of the network are relatively straightforward for an AV-based mobility service. However, interspersed among these are a significant number of Level 2 (Normal) road links, accounting for 38.8%, which suggests a moderate difficulty that could influence the operational strategies of AV-based mobility services. Although less frequent, there are also discernible pockets of Level 3 (Hard) and Level 4 (Very Hard) road links, representing 4.2% and 0.5% of the network, respectively. These areas represent higher challenges due to factors such as temporarily closed areas or underground parking lots. The presence of these higher difficulty levels, albeit sporadic, adds complexity to the network and highlights the necessity of advanced planning technology to ensure safe and efficient AV operations.

The spatial mixture of different levels, particularly the coexistence of lower- and higher-difficulty road links, complicates the task of seamless AV-based mobility service planning. This complexity necessitates that any AV system implemented must be capable of handling a range of conditions, from the simplest to the most challenging. To maintain continuous and reliable mobility services, overcoming the higher difficulty levels is imperative, requiring robust AV technology and adaptive planning algorithms that can navigate through or around such challenging areas effectively.

The study presented herein evaluates the effectiveness of the proposed methodology at deriving an optimal service network tailored to three distinct mobility services (taxi, DRT, and STSs). This analysis incorporates the unique demand characteristics of each service to facilitate targeted planning using the suggested approach. The methodology’s effectiveness is gauged by comparing the outcomes of human-driven vehicle (HDV)-based mobility services with those of AV-based mobility services. In executing this comparative analysis, the primary objective is to assess how well the proposed planning approach aligns with the operational needs, difficulty level for road links, and efficiency demands of each mobility service type. By integrating and analyzing the specific demand characteristics inherent to each service, this study aims to identify and construct service networks that optimize performance metrics such as safety, operational cost, and number of trips accommodated by the mobility service.

5.1. Taxi

Figure 9 illustrates the levels of road links extracted for use in AV-based taxi mobility services, demonstrating the optimal networks chosen for their operation. This visualization allows for a comparative analysis with Figure 5, which displays the road network utilization of HDV-based taxi mobility services. The spatial differences in network usage between HDV-based and AV-based taxi mobility services can be discerned through this comparison.

Notably, the network for the AV-based taxi mobility service is strategically configured around links where taxi demand is highly concentrated. Despite this concentration of demand, certain links have been deliberately excluded from the AV-based network. These excluded links correspond to sections with high difficulty levels. The strategic exclusion of high-difficulty links is a significant operational decision; it indicates that the service does not cater to demands requiring passage through these challenging areas. Consequently, this selective approach focuses on providing services primarily in regions where the demand is high, yet the road difficulty level is manageable.

As a result, the network for AV-based taxi mobility services is predominantly concentrated in the lower right area of the map, where the demand is high and the difficulty level of the roads is comparatively low. This area-specific concentration ensures that AV-based services are optimized for efficiency and safety, aligning with the overarching goal of enhancing service reliability in less-challenging traffic environments.

Further numerical insights are presented in Figure 10, which quantifies the changes in the number of road links utilized when integrating and operating AV technologies in taxi mobility services. While 563 links are employed for the existing HDV-based service, only 225 links are selected for the AV-based service. The analysis reveals that only 40% of the current road links are suitable for AV services, primarily due to the strategic avoidance of higher-difficulty road links in the AV-based network selection. Consequently, the introduction of AV-based services necessitates a 17.6% increase in travel distance, suggesting the need for considerable detours to circumvent roads of higher difficulty to effectively meet the widespread demand.

Additionally, Figure 10b shows the estimated costs associated with implementing the AV-based taxi mobility service. The operation of autonomous vehicles is expected to require approximately 77,964,000 KRW. Furthermore, there is an anticipated risk-related cost of 10,227,000 KRW when applying AV-based mobility services. This financial analysis underscores the economic considerations necessary to facilitate the transition from HDV to AV systems within the taxi service framework, reflecting both the operational and risk-related financial implications.

Figure 11 depicts the changes in road link difficulty levels when applying the proposed extraction methodology for AV operations. This analysis shows that approximately 57% of the 128 selected road links for AV services are categorized as difficulty Level 1. About 38% of the links, equating to 86 links, are at difficulty Level 2, while around 5% face difficulty Level 3. It is noted that road links requiring difficulty Level 4 are not utilized for AV-based taxi services, emphasizing a strategic avoidance of the most challenging conditions to ensure smoother AV operation. These figures collectively highlight the strategic adjustments required in road network usage to facilitate the deployment of AV-based taxi mobility services.

5.2. DRT

Figure 12 presents the levels of road links extracted specifically for AV-based DRT mobility services. This visualization enables a direct comparison with Figure 6, which delineates the network utilized by HDV-based DRT services. The comparison highlights the strategic adjustments made to the network to accommodate the unique operational needs of AV technology.

Demand is generally higher for DRT mobility services than for taxi mobility services. Moreover, DRT trips characteristically involve shorter distances between origin and destination compared to taxis. This distinctive trait of DRT trips influences the configuration of the network for AV-based DRT services, as illustrated in Figure 12. The network shows a dispersed pattern of road links, strategically avoiding high-difficulty links while striving to meet the demand for shorter travel distances effectively. This approach ensures that AV-based DRT services cater to frequent, short-distance trips across a wider area without compromising on safety or efficiency due to challenging road conditions.

As a result, while the lower right area of the map still has concentrated areas of AV-based service provision such as taxi mobility services, the distribution for DRT is different. It is characterized by individual trips rather than continuous routes, reflecting the segmented nature of DRT services. This dispersed yet focused deployment allows AV-based DRT services to efficiently cover diverse and sporadic user needs within the urban landscape, optimizing the use of road links to accommodate trip demands efficiently. This strategic dispersal aligns with the operational philosophy of DRT, which is designed to offer flexible and responsive transport solutions, adapting dynamically to varying urban mobility needs.

Figure 13 provides a numerical depiction of the changes in the number of road links used when AV technologies are introduced to DRT mobility services. Originally, DRT services utilized 1115 road links. However, with the application of the optimal network extraction methodology tailored for AV services, only 194 links were selected, which constitutes about 17% of the previous network. It was found that 13% of these selected links overlap with those used in the HDV-based services. The predominant selection of different links for AV services suggests a strategic avoidance of more challenging roads, which inherently requires detouring to meet service demands across the target area. This strategy results in a 16.55% increase in travel distance due to necessary detours around high-difficulty areas. Additionally, Figure 14b presents the estimated costs associated with implementing the AV-based DRT mobility service. The operation of autonomous vehicles for DRT services is expected to require approximately 66,078,000 KRW. Moreover, the application of AV-based mobility services entails an anticipated risk-related cost of 8,979,000 KRW.

Figure 14 further explores the implications of the extraction methodology on the difficulty levels of road links employed for AV-based DRT services. Sixty percent of the 194 chosen road links are of difficulty Level 1, indicating a preference for easier routes to facilitate smoother AV operation. Additionally, about 38% of the links, totaling 73 segments, are classified under difficulty Level 2, and approximately 2% require a difficulty Level 3. Notably, road links categorized as difficulty Level 4 are completely avoided in the AV-based DRT services, underscoring the methodology’s effectiveness at minimizing exposure to the most challenging road conditions.

5.3. STS

Figure 15 displays the extracted road link levels specifically tailored for AV-based STS mobility services. This visualization allows for a spatial comparison with Figure 7, which outlines the road network used by HDV-based STSs. The comparison illustrates how the integration of AV technology influences the strategic selection and utilization of road links for STSs.

Unlike the other two mobility services—taxi and DRT—STS mobility services exhibit a unique characteristic where demand is highly concentrated in specific areas. This concentration of demand has led to a distinctive placement of the network for AV-based STS mobility services, predominantly in the upper part of the map. This location is significantly different from the other services. The focused demand in STSs has also resulted in an increased usage of higher-difficulty roads, specifically those classified as Level 3 or above. Unlike taxi and DRT services, which strategically avoid high-difficulty roads to ensure safer and more efficient routes, STSs are compelled to navigate these challenging links to meet the essential transportation needs of their specific user groups. This necessitates traversing roads with higher difficulty levels, indicating that the provision of the minimum viable mobility service to meet concentrated demand is a higher priority than avoiding difficult road conditions.

Given the need to navigate more challenging roads, it is apparent that the AV technologies employed in STS mobility services must be of a higher technical standard. This underscores the importance of advanced navigation systems, robust sensory processing, and enhanced decision-making algorithms that are capable of safely and effectively handling complex driving environments. Consequently, the development and implementation of AV solutions for STSs need to focus not only on typical efficiency and safety concerns but also on the ability to cope with higher road difficulty levels to ensure reliable service provision where it is most needed.

Figure 16 further quantifies the differences depicted in Figure 15, detailing the changes in the number of road links utilized with the introduction of AV to STS mobility services. Initially, STSs operated using 330 road links; however, the application of the optimal sub-network extraction methodology for AV services resulted in only 71 links being selected. This substantial reduction means that only 22% of the previous network is suitable for AV-based operations, indicating a critical shift towards selecting road links with lower difficulty levels to enhance AV safety and efficiency.

An analysis of the characteristics of these selected road links reveals that 61% of the links overlap with those used in the HDV-based services, while the remaining 39% represent new routes that avoid higher difficulty levels not suited to AV operation. This strategic selection is predicated on the necessity of minimizing risk and improving the reliability of AV navigation, which typically avoids challenging routes that may compromise safety. Figure 16b presents the estimated costs associated with implementing the AV-based STS mobility service. The operation of autonomous vehicles for STSs is expected to require approximately 25,142, KRW. Moreover, the application of AV-based mobility services entails an anticipated risk-related cost of 5,762,000 KRW.

Figure 17 provides an in-depth look at the difficulty levels of the road links used for the AV-based STS mobility services after applying the extraction methodology. Approximately 56% of the 40 selected road links are classified as difficulty Level 1, reflecting the preference for routes with relatively low difficulty to facilitate smoother AV operations. About 40% of the links, amounting to 23 segments, are assessed at difficulty Level 2, and a small fraction, around 2%, require navigating difficulty Level 3. Notably, routes requiring difficulty Level 4 are completely excluded from the AV service network, ensuring that the most challenging conditions are avoided to maximize safety and operational efficiency.

6. Conclusions

This study developed a method for extracting optimal service networks for AV-based mobility services, considering both the AV driving performance and real-world service demand within a geo-fenced area. Through an analysis of real-world data, this study identified distinct patterns in network utilization across different service types, such as taxis, DRT, and STSs. Additionally, this study assessed the Operational Design Domain (ODD) difficulty for AV deployment, categorizing road links into four levels. In Sejong City, around 50% of the roads were classified as Level 1, suitable for autonomous driving, while 10% fell into the most challenging Level 4 category. Using a meta-heuristic approach, this study revealed that AV services were more suitable in areas with lower ODD difficulty and higher traffic demand. Excluding high-difficulty road links led to detours and increased travel distances, but demand remained concentrated in these challenging areas. Therefore, despite the difficulties, there was a clear trend of overcoming these challenges due to concentrated demand, highlighting the need for alternative methods to reduce AV difficulty in such areas.

The proposed framework can be extended to other urban areas as well. However, to apply the method effectively in different cities, it is essential to have access to Origin–Destination (OD) trip data from existing mobility services to allow for an accurate assessment of service demand patterns. Additionally, the framework requires comprehensive knowledge of ODDs for calculating the difficulty of each road segment in the city. With these datasets in place, the method outlined in this study can be adapted to optimize AV-based mobility networks in a wide range of urban environments, ensuring both profitability and operational safety. It is important to note, however, that the cost associated with navigating road segments of varying difficulty may vary based on local factors such as labor costs and the level of technological advancement. These factors must be considered when extending the framework to new cities, as they can influence the overall economic feasibility of AV deployment. This adaptability enhances the broader applicability and potential impact of our proposed approach.

Despite the positive outcomes demonstrated by the proposed methodology, several areas require further investigation to enhance the framework’s extensibility and applicability. In this study, road difficulty levels were primarily based on expert opinions, which introduced some variability, especially at the boundaries between difficulty levels. To improve accuracy, there is a need for data-driven analysis and quantification of road difficulty levels based on ODD factors. This could be achieved through the use of fuzzy models, which offer a key advantage in handling the inherent uncertainty and gradual transitions between difficulty levels. Fuzzy models allow for more flexible boundaries, reducing the issue of abrupt level changes due to minor fluctuations in data. By applying fuzzy logic, future research could develop a more nuanced, quantitative evaluation of road difficulty, which would help overcome the limitations of purely expert-driven assessments and ensure a smoother transition across difficulty levels. This approach will provide a more reliable and data-based framework for evaluating road complexity in AV deployment.

Secondly, in this study, the methodology was applied and validated primarily based on ODD factors focusing on static road environment characteristics. However, in real-world AV-based mobility services, dynamic factors such as traffic conditions play a significant role in service performance and safety. While this study centered on Sejong City and considered static parameters such as road structure and service history (e.g., taxis, DRT, and STSs), it did not fully account for the impact of fluctuating traffic conditions and other dynamic elements. Therefore, future research should explore the influence of these dynamic factors on AV service operations. As AV services continue to expand from low-traffic areas to more complex urban environments, it is essential to develop a more comprehensive framework that integrates both static and dynamic environmental variables. Addressing these factors will enable a more accurate and adaptable model of the real-world deployment of AV mobility services.

Lastly, this study’s framework was designed with the goal of determining which road segments are suitable for AV deployment and identifying the levels of demand that can be accommodated, while demand in more challenging segments was effectively disregarded. However, limiting service areas based on difficulties faced by autonomous vehicles is not a viable long-term solution. To address the limitations of AV technology, there is a need to explore how human-operated mobility services can be effectively designed and integrated alongside AV services. Future research should focus on creating a hybrid mobility system wherein the limitations of AV technology are acknowledged yet the overall efficiency of mobility services is enhanced by seamlessly incorporating human-operated vehicles. This combined approach would ensure that all demand is met, even in challenging areas, and improve the overall utility and accessibility of mobility services.