Reducing Safety Risks in Construction Tower Crane Operations: A Dynamic Path Planning Model

Abstract

Tower cranes are the most used equipment in construction projects, and the path planning of tower crane operations directly affects the safety performance of construction projects. Traditional tower crane operations rely on only the driving experience and manual path planning of crane operators. Poor judgement and bad path planning may increase safety risks and even cause severe construction safety accidents. To reduce safety risks in construction tower crane operations, this research proposes a dynamic path planning model for tower crane operations based on computer vision technology and dynamic path planning algorithms. The proposed model consists of three modules: first, a path information collection module preprocessing the video data to capture relevant operational path information; second, a path safety risk evaluation module employing You Only Look Once version 8 (YOLOv8) instance segmentation to identify potential risk factors along the operational path, e.g., potential drop zones and the positions of nearby workers; and finally, a path planning module utilizing an improved Dynamic Window Approach for tower cranes (TC-DWA) to avoid risky areas and optimize the operational path for enhanced safety. A prototype based on the theoretical model was constructed and tested on actual construction projects. Through experimental scenarios, it was found that each tower crane operation poses safety risks to 3–4 workers on average, and the proposed prototype can significantly reduce the safety risks of dropped loads from tower crane operations affecting ground workers and important equipment. A comparison between the proposed model and other regular algorithms was also conducted, and the results show that compared with traditional RRT and APF algorithms, the proposed model reduces the average maximum collision times by 50. This research provides a theoretical model and a preliminary prototype to provide dynamic path planning and reduce safety risks in tower crane operations. Future research will be conducted from the aspects of multiple device monitoring and system optimization to increase the analysis speed and accuracy, as well as on human–computer interactions between tower crane operators and the path planning guidance model.

1. Introduction

Tower cranes are among of the most used equipment in construction projects, playing a critical role in transporting materials and building components on construction sites [1]. The operations of tower cranes largely depend on the experience and manual path planning of crane operators. Poor judgment and inadequate path planning by operators as a result of human inaccurate distance calculation and restricted view on the construction sites can significantly increase safety risks, potentially leading to severe construction accidents, such as lifted objects striking already constructed parts of a building or falling on workers. Several severe crane accidents have happened globally during the past several years, e.g., in August 2023, a devastating crane collapse occurred in Thane, a city in Western India, resulting in 20 fatalities and three injuries [2]. In December 2021, a tower crane collapse in Turin, Italy, led to three deaths and three injuries [3]. Between 2011 and 2020, the United States recorded 129 construction crane accidents [4]. In Australia, crane-related accidents in the construction sector resulted in 47 fatalities between 2003 and 2015, accounting for 8.5% of all construction industry fatalities [4,5]. In China, tower crane accidents result in an average of 1.17 fatalities per year, and the consequences of such accidents are often more severe than typical construction incidents [6]. Many fatalities involving cranes and ground workers are caused by contact with objects and equipment, particularly crane loads and parts [7]. Due to the nature of high-altitude operations, it is difficult for ground workers to consistently monitor the lifted loads above them, making the risk of falling objects a critical safety issue that needs to be addressed.

Currently, ensuring the safety of tower crane operations relies heavily on manual monitoring, which is both time-consuming and labor-intensive [8]. To address these challenges, researchers have explored various solutions. In the field of sensor-based monitoring, Zhou et al. [9] developed a digital twin-based monitoring system for port cranes that integrates multi-source heterogeneous data from various sensors, enabling real-time operation monitoring and control program simulation. Similarly, Lai et al. [10] proposed a shape-performance integrated digital twin framework that combines multiple sensors with analytical, numerical, and AI models to enable real-time structural analysis and predictive maintenance of cranes. Meanwhile, with the increasing availability and affordability of surveillance cameras on construction sites, there has been increasing interest in leveraging computer vision technology to monitor and assess operational safety [11,12,13]. Yang et al. [14] were the first to utilize video data from cameras mounted above cranes to identify hazardous areas using Mask Region-based Convolutional Neural Network (Mask R-CNN) and pixel coordinate distance conversion methods. Cao et al. [15] employed a convolutional neural network (CNN) to detect whether workers on construction sites were wearing safety helmets in surveillance videos, thus reducing the risk of injury. Wang et al. [16] developed a system to automatically detect and three-dimensionally (3D) locate the position of tower crane hooks using cameras. Although these studies have contributed to the identification of potential safety risks during tower crane operations, they have seldom provided solutions for mitigating those risks. For example, while operators may be informed of the presence of workers near a load, there is no guidance on how to adjust crane movements to avoid endangering those workers.

Path planning technology has been widely adopted across various crane applications, demonstrating its maturity and versatility in the industry. In offshore applications, time-optimal trajectory planning methods have been developed specifically for offshore cranes operating on non-inertial platforms. These methods incorporate multiple physical constraints to achieve safe and efficient material handling in challenging offshore environments [17,18,19]. In construction sites, autonomous tower cranes have implemented offline trajectory planning algorithms to enable precise payload movement along straight lines while efficiently navigating through obstacles [20,21]. To further enhance safety and efficiency in tower crane operations at construction sites, researchers have developed various intelligent algorithms for path planning optimization [22,23,24]. Zhu et al. [25] compared and summarized three intelligent algorithms, including the node generation-based A-star (A*) search algorithm, the sampling-based Rapidly Exploring Random Tree (RRT), and the heuristic Genetic Algorithm (GA). Du et al. [26] studied the optimization of tower crane path points based on construction crane operation standards and proposed an improved RRT Connect algorithm. By introducing heuristic information into the path search, the efficiency and quality of the algorithm were enhanced. Lin et al. [27] proposed an automated lift path planning method for tower cranes based on environmental point clouds, utilizing an optimized A* algorithm to identify executable paths. However, in real-world crane operations, the dynamic environment is complicated by the unpredictable movement of ground workers. Traditional path planning methods may not provide effective decision support to mitigate the risk of falling loads.

To address these limitations in tower crane operations, this study aims to develop an effective method for mitigating the risk of falling loads to ground workers during tower crane operations, thereby providing operational support to tower crane operators and ensuring safety. We propose a dynamic path planning model for tower crane operations that integrates computer vision and path planning algorithms. This model includes three key modules: a path information collection module, a path safety risk evaluation module, and a path planning module. It identifies on-site information necessary for the dynamic operation of tower cranes, perceives safety risks, and proposes path planning solutions. The instance segmentation capability of the You Only Look Once version 8 (YOLOv8) framework is demonstrated through a case study, and the effectiveness of the improved Dynamic Window Approach for tower cranes (TC-DWA) algorithm for path planning is validated via simulation experiments. The study also involves the construction of a safety monitoring platform for tower crane operations, which revealed a significant number of unsafe conditions on construction sites. This platform provides driving reminders and suggestions for adjusting the operation paths of tower crane operators.

2. A Dynamic Path Planning Model for Tower Crane Operations

During the lifting and lowering phases of tower crane operations, operators often engage in tasks that reduce hazards for surrounding workers. The risk of dropped loads is greatest during horizontal operations in the air, where the area of potential impact is the widest. Therefore, this study focuses on the horizontal operational process of the tower crane to develop path planning strategies. In tower crane operations, dangers arise when lifted objects pass over the heads of ground workers due to disengagement. To prevent such accidents, lifted loads can be projected onto the ground to determine whether there is a risk of collision in the horizontal direction with the workers.

The dynamic path planning model framework for tower crane operations proposed in this study is illustrated in Figure 1. First, a camera mounted above the hook monitors the tower crane operations, collecting real-time path information from close range. The path information collection module preprocesses the video data to extract relevant operational path information and generate a dataset of construction path images. Secondly, a safety risk evaluation module for tower crane operations is developed based on the YOLOv8 instance segmentation framework, which defines potential risk factors along the operational path, such as drop zones and the positions of nearby workers. The YOLOv8-seg model is employed to segment potential drop zones and identify protected objects in the video. The pixel coordinate information obtained is then converted into real-world coordinates to determine the actual state of each object, including its size and position as scene information, thereby enabling dynamic risk evaluation during crane operations. Finally, the path planning module applies an improved TC-DWA algorithm to avoid risky areas and optimize the operational path for enhanced safety.

2.1. Path Information Collection Module

The tower crane operation process involves transporting lifted objects from the starting point to the unloading point, often spanning multiple buildings during construction. Using long-range cameras to monitor tower crane operations can lead to line-of-sight obstructions [28]. To obtain more accurate data, a close-range camera is installed above the hook to capture video of the tower crane operations. In this study, a Hikvision camera is used, employing the Real-Time Streaming Protocol (RSTP) video streaming protocol to acquire real-time monitoring video and collect path information.

The collected tower crane operation videos are converted into an image dataset. Due to the complex construction site environment, some images may have suboptimal lighting conditions, necessitating data preprocessing. Blurry, duplicate, and other invalid images are filtered out, and the image sizes are standardized for input into the network. To enrich the dataset, data augmentation techniques such as translation, rotation, flipping, cropping, and brightness adjustment are applied to the construction images.

2.2. Path Safety Risk Evaluation Module

To estimate the risk of falling objects during tower crane operations, it is necessary to define the potential drop zones of the operations [29]. The Occupational Safety and Health Administration (OSHA) in the United States has established health and safety requirements for construction cranes in its Federal Regulations (CFR) [30]. CFR 1926.1425 defines the drop zone as “an area where materials that are reasonably foreseeable to be partially or fully suspended during an accident may fall (including, but not limited to, the area directly beneath the load being lifted)”. Based on the CFR’s definition, in this research, it is assumed that the dropped load follows a parabolic trajectory, as shown in Figure 2, and the operational hazard radius can be set around the tower crane hook to define the potential drop zone. In construction projects, workers, critical machinery, and important work areas are significant protected objects. This study focuses on mobile workers rather than static objects.

CNN, R-CNN, Mask R-CNN, and other models used in deep learning have shown significant effects in object detection, image segmentation, trajectory tracking, and other fields. However, two-stage region-based neural networks such as R-CNN have high computational costs in image recognition and classification. To achieve real-time operation, a cutting-edge object detector called YOLO emerged in 2016, employing a single-stage detection method that directly predicts bounding boxes and classes within a grid on the image, significantly speeding up the detection process compared to existing frameworks [31]. Therefore, YOLO has been widely applied and continually updated [32,33]. He et al. [34] used YOLOv5 for automatic identification of reflective vests, achieving an average accuracy of over 80%, meeting practical needs. Demetriou [35] examined real-time localization and classification models for construction site debris using various deep learning models, finding that the YOLOv7 model demonstrated stronger inference speed and accuracy. To further calculate the distance between cranes and workers, mere object detection is insufficient; instance segmentation is needed to provide accurate pixel-level segmentation for each object instance in the image. YOLOv8 leads in accuracy and precision over YOLOv7, with YOLOv8-seg featuring instance segmentation capabilities [36,37,38]. This study employs YOLOv8-seg as an advanced instance segmentation algorithm, which offers higher accuracy and faster speed compared to previous YOLO versions (YOLOv5, YOLOv6, YOLOv7). The YOLO model, once trained, can perform instance segmentation in tower crane operation videos, outputting classes, shapes, coordinates, and confidence scores. The YOLOv8-seg model mainly consists of three modules: the input module, the backbone network module, and the task head module, as illustrated in Figure 3.

• Input Module: This module is used for preprocessing images to ensure they can be correctly handled by the model. Generally, the main tasks of the input module include resizing, normalizing, and padding the images. All YOLO algorithms require input images to be converted to a fixed size before being fed into the detection model for training. The standardized image size designed for this study is 640 × 640 × 3.
• Backbone Module: This module is built based on CNN. The convolutional layers (Conv) are used to extract and process features from the input feature map, capturing richer semantic information. CSP Bottleneck with 2 Convolutions (C2f) consists of 1 × 1 convolution layers and multiple Bottleneck layers, focusing on processing and merging features to extract more detailed feature information. Spatial Pyramid Pooling—Fast (SPPF) is a variant of spatial pyramid pooling used to perform pooling operations at different scales and integrate the results to extract multi-scale feature information, thereby enhancing the understanding and processing of targets of varying sizes.
• Task Head Module: This module is constructed based on the Path Aggregation Network (PAN) framework, which integrates concepts from the Feature Pyramid Network (FPN). Up-sampling and contact modules for up-sampling and connection are utilized, which enhances the effective use of multi-scale feature information in instance segmentation tasks, thereby improving detection accuracy and robustness.

Images are preprocessed to a standardized resolution of 640 × 640 × 3 pixels within the input module. The preprocessed images are then fed into the backbone module, where a series of CNN operations extract features. This process involves convolutional operations, with 1 × 1 convolutional layers in the C2f component and Bottleneck layers working in concert to process and merge features, thereby extracting semantic information from the input images. The features extracted by the backbone are passed to the task head module for further refinement, specifically for instance segmentation. Up-sampling and contact modules are used to integrate multi-scale feature information, which is critical for accurately detecting and segmenting objects of varying sizes. After feature enhancement in the task head module, the model generates the final output. This output includes the classification of instances into predefined classes, the determination of their shapes, the specification of their bounding box coordinates, and the assignment of confidence scores for each detection.

Through instance segmentation, the potential drop zone (the operational hazard radius around the hook) and the protected objects (workers) can be represented in pixel coordinates within the plane projection. This study employs a method to determine the actual coordinates of the objects based on a simple proportional relationship between pixel coordinates and real-world coordinates [14,39,40]. The conversion between pixel coordinates and actual coordinates is described by the following Equation (1) [41]:

$$\left(x_{actual},y_{actual}\right)={\displaystyle \frac{L_a}{L_p}}\left(x_{pixel},y_{pixel}\right)$$

(1)

where $\left(x_{actual},y_{actual}\right)$ represents the real coordinates, $\left(x_{pixel},y_{pixel}\right)$ represents the pixel coordinates, $L_a$ is the real length of the object, and $L_p$ is the measured length of the object in the image. In this study, the objects are projected onto a horizontal plane, and thus the Z-axis coordinate is not considered. Previous experiments [14] have shown that this method yields a positioning error of approximately 3%. As illustrated in Figure 4, based on the coordinate conversion, the distance between the two closest green grid points that surround the tower crane hook’s center blue point is 0.5 m. By real-time monitoring of the potential drop zone and the environment of protected objects, the study identifies the actual distance between the drop zone and the workers during tower crane operations, thereby dynamically perceiving the existing safety risks.

2.3. Path Planning Module

The Dynamic Window Approach (DWA) is a commonly used dynamic path planning algorithm in robotics [42]. The core idea of DWA is to use the current state of the robot (position, speed, etc.) and surrounding obstacles to generate a dynamic feasible region within a velocity space, known as the velocity window. By dynamically adjusting the velocity window and selecting optimal velocity commands, mobile robots can safely avoid obstacles while moving quickly toward their targets. Due to its excellent performance, Hou et al. [43] have continuously improved DWA for dynamic path planning. Yao et al. [44] proposed a fuzzy logic improved dynamic windows approach for automatic navigation of mobile robots. Sun et al. [45] dynamically modified the weight coefficient combination of the dynamic window method to cope with the constantly changing dynamic environment and suboptimal global planning path problems that arise after local obstacle avoidance. The dynamic window method is continuously improving to adapt to more complex path planning problems [46,47,48]. Therefore, this study attempts to apply an improved version of DWA to tower crane operation path planning for the first time.

In solving the path planning problem for horizontal tower crane operations, the real-world environment can be simplified to a two-dimensional space. A working hazard radius is defined around the hook’s center, creating a circular area, while workers in the environment are modeled as rectangular obstacles, with their top-left and bottom-right corners marked, as shown in Figure 5.

Assuming the tower crane’s base coordinate is at the origin $O\left(0,0\right)$, the starting point of the construction task is denoted as S, and the endpoint as E. The solution space for path planning is defined as shown in Equation (2):

$$P=\left[0,X\right]\times \left[0,Y\right]\cup [-X,0]\times \left[0,Y\right]$$

(2)

where $X,Y\in {ℝ}^{+},P\subset {ℝ}^2$.

$X,Y$ denotes the upper boundary of the construction environment. Each obstacle in the environment is represented as $o{b}_i$, with the set of obstacles denoted as $ob=\left\{o{b}_1,o{b}_2,\dots ,o{b}_n\right\}$, where n is the total number of obstacles. $P_{obs}=o{b}_1\cup o{b}_2\cup \dots \cup o{b}_n$ represent the infeasible regions in the environment; thus, each obstacle area satisfies the following condition: $P_{obs}=\{(x,y)\mid x_{i_1}\le x\le x_{i_2} ,y_{i_1} \le y\le y_{i_2} ,i=1,2,\dots ,N\}$. The feasible solution region can be expressed as follows: $P_{free}=\{(x,y)\mid 0\le x\le X,0\le y\le Y ,(x,y)\notin o{b}_i,i=1,2,\dots ,n\}$. If there exists a feasible path point represented as $P_j(x_j,y_j)$, where $P_j\in P_{free}$, then a safe working path for the tower crane will consist of $P_T=\{P_0,P_1,P_2,\dots ,P_n,P_{n+1}\}$. The tower crane operation can be divided into two motion modes: the rotation of the boom and the trolley’s amplitude adjustment. Therefore, the following physical constraints must also be satisfied during the motion:

• Rotation Radius:$r$, $r_{min} \le r\le r_{max}$, where $r_j=\sqrt{{x_j}^2+{y_j}^2},j=0,1,2,\dots ,n+1$.
• Velocity Constraints: $V_s$, $V_s=\left\{(v,\omega )|v\in \left[v_{\min },v_{\max }\right],\omega \in [{\omega }_{\min },{\omega }_{\max }]\right\}$, where $v$, $\omega$ are linear and angular velocities, respectively.
• Acceleration Constraints: $V_a$, $Va=\left\{\begin{array}{l}(v,\omega )|v\in \left[v_c-{\dot{v}}_m\times \Delta t,v_c+{\dot{v}}_m\times \Delta t\right]\\ \omega \in \left[{\omega }_c-{\dot{\omega }}_m\times \Delta t,{\omega }_c+{\dot{\omega }}_m\times \Delta t\right]\end{array}\right\}$, where $v_c,{\omega }_c$ are the current linear and angular accelerations and are their respective maximum values.
• Braking Distance Constraint: $V_{ob}=\left\{\begin{array}{l}\left(v,\omega \right)|v\le \sqrt{2\times dist(v,\omega )\times \dot{v}},\\ \omega \le \sqrt{2\times dist(v,\omega )\times \dot{\omega }}\end{array}\right\}$, where $dist(v,\omega )$ is the distance from the simulated trajectory to the nearest obstacle.

The trajectory evaluation function of the TC-DWA algorithm consists of three sub-functions:

$$G(v,\omega )=\sigma [\alpha \times heading(v,\omega )+\beta \times dist(v,\omega )+\gamma \times vel(v,\omega )]$$

(3)

where $\alpha ,\beta ,\gamma$ are the weights corresponding to each sub-function; $heading(v,\omega )$ indicates the deviation between the endpoint of the simulated trajectory and the target point, with smaller values signifying closer trajectories; $dist(v,\omega )$ represents the distance between the simulated trajectory and protected objects, with larger values indicating safer trajectories; $vel(v,\omega )$ signifies the current motion speed, where higher values allow for quicker approaches to the endpoint. A higher value of the total trajectory evaluation function $G(v,\omega )$ signifies better overall trajectory performance. To balance the impact of various indicators on the evaluation results, normalization is applied to the three sub-evaluation functions, as shown in Equation (3) [49]. The motion modes of the tower crane are categorized into boom rotation and trolley amplitude adjustment. Each mode requires a corresponding kinematic model.

The kinematic equation for the boom rotation model is shown in Equation (4).

$$\{\begin{array}{l}x=r\times \cos (yaw+d\theta )\\ y=r\times \sin (yaw+d\theta )\\ d\theta =w\times dt\end{array}$$

(4)

The kinematic equation for the trolley amplitude adjustment is shown in Equation (5).

$$\{\begin{array}{l}x=x+v\times \cos (yaw)\times dt\\ y=y+v\times \sin (yaw)\times dt\end{array}$$

(5)

where $yaw$ is the current yaw angle, and $d\theta$ is equal to the angular velocity $w$ multiplied by the time increment $dt$.

The framework of the path optimization algorithm is illustrated in Figure 6. It outlines the steps involved in the TC-DWA algorithm for optimizing the tower crane’s operational path. The process begins with the collection of real-time data from the crane’s environment, followed by the identification of obstacles and the potential drop zones. Next, the algorithm evaluates the current trajectory based on various constraints, including speed, acceleration, and safety distances from protected objects. The trajectory evaluation function incorporates multiple criteria to ensure a smooth and safe path, adjusting as necessary based on dynamic environmental feedback. Finally, the algorithm generates an optimized path that balances efficiency and safety, allowing the tower crane to perform its tasks effectively while minimizing risks to workers and equipment.

This model aims to optimize the dynamic path of tower cranes during operation, ensuring the safe transportation of lifted loads while minimizing the risk of drops affecting ground workers, and striving to select the shortest path. In certain situations, when the current motion state is unable to execute the planned route, a switch in motion mode will be implemented.

In the long-distance operation scenario, the tower crane is responsible for transporting loads from the material storage area to the worker’s operation area on the main construction site. During this phase, the transportation range is extensive, and the rotation angle is significant, which may necessitate traversing around critical equipment. A common scenario is illustrated in Figure 7.

The path optimization strategy is as follows:

• Initially assess the distance L1 from the tower crane to the material storage area and the distance L2 to the worker’s operation area.
• If L1 > L2, initialize the motion mode to linear movement of the trolley to reduce the rotation radius, adjusting the crane’s rotation radius L1 to L2.
• Throughout the operation, if the forward path encounters critical equipment, switch the motion mode to minimize the radius for obstacle avoidance. Upon successful avoidance, revert to the rotation mode.
• If the path ahead encounters ground workers, increase the radius for obstacle avoidance until the rotation radius aligns with the endpoint.
• Finally, perform trolley extension to safely transport the load to the destination.

In the short-distance operation scenario, the tower crane transports loads to different worker operation areas within the main construction site, characterized by a limited movement range and smaller rotation angles. A common scenario is depicted in Figure 8.

• The path optimization strategy is as follows: Begin by assessing the distance L1 from the tower crane to the starting point of the worker’s operation area and the distance L2 to the endpoint.
• If L1 < L2, directly engage the rotation movement.
• During operation, if the forward path encounters ground workers, switch the motion mode for obstacle avoidance. Upon successful avoidance, revert to the rotation mode.
• Continue rotating until the rotation radius aligns with the endpoint, and then execute trolley extension to safely transport the load to the destination.

In this study, an integrated tower crane construction monitoring program interface was developed based on the experimental platform parameters shown in

Table 1. Configuration parameters of the experimental platform.

Device	Configuration
Computer	Lenovo G5000 IRH8 (Lenovo Corporation, Beijing, China)
Operation system	Windows 11 (64-bit)
CPU	13th Gen Intel® Core™ i7-13700H 2.40 GHz
RAM	16 G
GPU	NVIDIA RTX4060 Laptop GPU, 8 G
GPU accelerator	Cuda 11.7
Framework	Pytorch 1.13.0
Scripting language	Python 3.8

, as depicted in Figure 9, which includes capabilities for risk evaluation and path planning. The interface displays a live video feed of the crane operation on the left side, where instance segmentation is conducted using YOLOv8. This sophisticated algorithm facilitates the identification and tracking of distinct objects within the video frame, thereby enhancing the system’s capacity to monitor the crane’s operation in real time. On the upper right side of the interface, a corresponding diagram is presented, reflecting the monitoring information from the video feed. This diagram offers a visual representation of the distance between the hook center and any workers present, enabling the system to detect risk situations during the operation. The system is designed to indicate a safe status during normal operations and to alert operators when workers enter a dangerous radius, thereby signaling a hazardous state. By setting the base position of the tower crane and the destination point, the system suggests the optimal direction of movement for the crane’s center. The example diagrams included in the study illustrate both safe and hazardous conditions, with the crane’s base positioned at the lower left and moving towards the upper left endpoint.

3. Experimental Simulation

3.1. Instance Segmentation Experiment

In this study, the image data used for model training and testing were sourced from a vertically mounted downward-facing camera on a tower crane at a construction site in Fujian, China. Multiple scenarios featuring lifting operations were selected, and 896 images were annotated using LabelMe version 5.3.1 image annotation software. The You Only Look Once version 8 Segmentation (YOLOv8-seg) network model was then trained and tested. The configuration parameters for the software and hardware platform used in this study are presented in Table 1.

Using the YOLOv8-seg network model for training, the dataset was randomly divided into a training set and a validation set in an 8:2 ratio. The batch size was set to 16, the number of epochs to 200, weight decay to 0.0005, and the learning rate to 0.01, with the initial weight model file being yolov8n-seg.pt. The model’s performance was tested using the following methods. Precision, recall, and [email protected] scores were considered as evaluation metrics [50]. TP, FP, and FN refer to True Positive, False Positive, and False Negative, respectively. TP represents the number of instances that the model correctly predicts as positive samples, FP represents the number of instances where the model incorrectly predicts positive samples, and FN represents the number of instances where the model incorrectly predicts negative samples. Precision refers to the proportion of instances correctly predicted as positive by the model to all instances predicted as positive by the model. Recall rate refers to the proportion of instances correctly predicted by the model as positive samples to all instances that are actually positive samples. Average Precision (AP) is the area under the precision–recall curve, which evaluates the model’s performance across different confidence thresholds. Mean Average Precision (mAP) is the average of AP values across all classes, used to evaluate the model’s overall performance in multi-class detection tasks. [email protected] represents the average precision at an Intersection over Union (IoU) threshold of 0.5. The calculation formulas are shown in Equations (6)–(9) [51], respectively.

$$Precision={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}}$$

(6)

$$Recall={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(7)

$$\mathrm{AP}={\displaystyle {\int }_0^1\mathrm{p}(\mathrm{r})\mathrm{d}\mathrm{r}}$$

(8)

$$\mathrm{mAP}={\displaystyle \frac{1}{\mathrm{x}}}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{x}}{\mathrm{AP}}_{\mathrm{i}}}$$

(9)

The experimental results are presented in

Table 2. The instance segmentation training results.

Class	Precision	Recall	[email protected]
Worker	0.985	0.992	0.99
Hook	0.975	0.994	0.99

, showing that the YOLOv8-seg model achieved instance segmentation precision, recall, and [email protected] scores close to 1 for all objects. This indicates that the trained YOLOv8-seg model meets the instance segmentation requirements for the hooks and workers in the experimental videos, as illustrated in Figure 10.

3.2. Path Planning Simulation

The parameters for the improved Dynamic Window Approach for tower cranes (TC-DWA) experimental environment are set as detailed in

Table 3. TC-DWA experimental environment parameters.

Scenario	Number of Workers	Starting Point S	End Point E
E1	6~8	(−36, 10)	(10, 30)
E2	6~8	(18, 2)	(2,30)

, establishing common scenarios E1 for long-distance crane operations and E2 for short-distance crane operations. The number of workers is set to be between 6 and 8, with their positions randomly assigned. Workers move slowly during the construction phase, and their speed is set to range from 0 to 1 m/s. In the long-distance scenario, important machinery remains stationary and can be simulated by using stationary workers. The TC-DWA parameters are outlined in

Table 4. TC-DWA parameter settings.

Parameter	Value
Maximum linear velocity (m/s)	1
Minimum linear velocity (m/s)	−1
Maximum angular velocity (degree/s)	7.98
Minimum angular velocity (degree/s)	−7.98
Linear acceleration (m/s2)	0.5
Angular acceleration (degree/s2)	15
Linear velocity resolution (m/s)	0.1
Angular acceleration resolution (rad/s)	2
Safety distance (m)	1.5
Simulation prediction time (s)	1
Unit time Δt (s)	0.5
Detection window range (m)	2
Maximum rotation radius (m)	55
Minimum rotation radius (m)	10

, based on the specifications of the tower crane model at the construction site. The rotation range of the crane is set between 0 and π, with the rotation radius configured between 10 m and 55 m. Figure 11 illustrates the simulated tower crane operation scenario, where “origin” denotes the base of the tower, “obstacle” represents the moving workers, and the arrows indicate the direction of worker movement. The tower crane operation is modeled as moving from the start point to the end point with the tower body as the origin. The evaluation metrics of TC-DWA include the planned path length and the number of avoidance failures (when the planned route does not maintain a safe distance from obstacles).

Figure 12 shows that the proposed model can successfully provide multiple tower crane paths in various complex scenarios, validating the effectiveness of the method. In Scenario E1, based on the start and end positions, the system initiates the trolley extension mode to shorten the radius, followed by rotation. When the presence of workers along the rotation path is predicted, the model switches to trolley extension mode to increase the radius and avoid moving workers before resuming rotation. This process is repeated until the rotation path closely aligns with the endpoint, allowing the crane to shorten the radius and deliver the load to the destination. After simulating 20 random movements of workers on the job site, the crane successfully avoids workers in 20 instances of long-distance operation, with an average path length of 71.22 m. In Scenario E2, based on the start and end positions, the crane initiates the boom rotation mode. When workers are detected along the rotation path, it switches to trolley extension mode to increase the radius, avoids the moving workers, and then resumes rotation, repeating this until the load reaches the endpoint. The simulation of 20 random worker movements again results in safe avoidance and successful delivery of the load, with an average path length of 42.49 m.

4. Discussion

It has been demonstrated that the tower crane effectively plans paths using the TC-DWA algorithm across various scenarios, highlighting its adaptability and precision. This section aims to delve deeper into the comparative analysis of the TC-DWA algorithm against traditional path planning methods, specifically the Rapidly Exploring Random Tree (RRT) and the Artificial Potential Field (APF) methods, to establish the ultimate effectiveness of our approach in dynamic environments.

The RRT algorithm is a well-known path planning technique that operates by randomly sampling points in the configuration space from the starting point. It constructs a tree by connecting the nearest point that is free of obstacles to each sampled point until it explores the vicinity of the endpoint. While RRT is capable of exploring large spaces and can find paths in complicated environments, it often generates paths that are not optimal or smooth. This can result in operational inefficiencies, particularly in scenarios requiring precise movements, such as tower crane operations. The inherent randomness in path generation can lead to routes that are unnecessarily convoluted, as evidenced by the simulation results presented in Figure 13 The steep and complex paths generated by RRT highlight its limitations in adhering to the operational constraints of tower cranes, such as avoiding workers and maintaining safe distances from obstacles.

On the other hand, the APF method represents a dynamic path planning approach that utilizes a potential field to guide movement. Figure 14 displays the APF algorithm’s performance in similar scenarios. In this model, the goal point exerts an attractive force, while obstacles produce repulsive forces, directing the movement of the robot along the gradient of the potential field. While this method offers smoother paths, it can fail to respect the physical constraints of tower crane operations. Our simulations indicated that while APF managed to achieve shorter path lengths, it frequently encountered difficulties in avoiding moving obstacles, especially as construction conditions became more complex. The failures observed in the APF method underscore the need for a path planning solution that can dynamically adapt to the changing environment while ensuring safety and efficiency.

The comparative analysis of the three methods includes planned path lengths and the number of avoidance failures (instances where the planned route did not maintain a safe distance from obstacles). All results were run 20 times, and the mean (Mean) and standard deviation (SD) are provided. The best results for each experimental scenario are highlighted in bold. From

Table 5. Experimental results obtained from TC-DWA and other comparative algorithms.

Method	Scenario	Metric	Path Length (m)	Number of Obstacle Avoidance Failures
TC-DWA	E1	Mean	71.22	0
	E1	SD	16.51	0
	E2	Mean	42.49	0
	E2	SD	11.06	0
APF	E1	Mean	49.32	1.38
	E1	SD	0.23	1.52
	E2	Mean	31.19	1.62
	E2	SD	0.44	1.47
RRT	E1	Mean	67.05	33.76
	E1	SD	3.58	26.94
	E2	Mean	39.39	77.27
	E2	SD	1.70	20.26

, it is evident that the traditional static path planning algorithm, RRT, struggles to handle complex and dynamic environments, failing to generate high-quality planned paths and effectively managing moving obstacles. The APF algorithm, as a dynamic path planning method, achieved the shortest path lengths in all three scenarios; however, it encountered situations where it could not avoid moving obstacles, and its success rate in obstacle avoidance decreased with increasing complexity of construction conditions. The TC-DWA algorithm proposed in this study, while resulting in longer operational path lengths, successfully avoided obstacles in all scenarios, ensuring worker safety. Due to the variations in construction environments, the increased path lengths fluctuated between 1.36 and 1.44 times compared to the path lengths planned by the APF algorithm.

The increase in path lengths observed with the TC-DWA method can be attributed to its focus on safety, as the algorithm actively adjusts its path in response to the detection of nearby workers. This aspect of TC-DWA enhances operational safety by prioritizing the avoidance of potential hazards, even if it results in a longer travel distance. In practical applications, the safety of workers on-site is paramount, and the trade-off between path length and safety becomes a critical consideration. The ability of TC-DWA to maintain a safety buffer around moving workers while still achieving successful load delivery positions it as a superior choice for tower crane operations in dynamic environments. Furthermore, the experiments conducted illustrate the dynamic capabilities of the TC-DWA algorithm in real-time scenarios. The algorithm’s ability to adapt to unexpected changes in the environment—such as the sudden movement of workers—demonstrates its practical applicability and robustness. The simulations performed in this study underscore the importance of integrating real-time data from computer vision systems into the path planning process, as this can significantly enhance the algorithm’s situational awareness and responsiveness.

Through the safety risk monitoring of 20 segments of tower crane operation videos, which range from 1 to 3 min in length and have a frame rate of 24 frames per second (FPS), we observed an average of 286.6 safety hazard warnings per complete crane operation process. Considering the video frame rate, this equates to an average of 11.9 hazardous situations per video, posing safety risks to 3–4 workers. These findings underscore the prevalence of safety hazards in crane operations and the necessity for real-time monitoring systems.

In situations where hazardous conditions are detected, the TC-DWA algorithm provides corresponding directional movement suggestions, effectively reducing the risks associated with tower crane operations. The algorithm’s ability to dynamically adjust the crane’s path in response to detected risks is a significant advancement in ensuring the safety of workers on construction sites. The integration of YOLOv8 for risk evaluation and the TC-DWA algorithm for path planning has demonstrated a synergistic effect in enhancing the safety of tower crane operations. The system not only detects potential hazards but also actively participates in mitigating these risks by providing actionable guidance to operators. The study’s results highlight the importance of integrating advanced computer vision and path planning algorithms into tower crane monitoring systems. The combination of real-time video analysis and dynamic path planning offers a new solution to improve safety standards in construction environments.

5. Conclusions

This study introduces a dynamic path planning model for tower crane operations that leverages computer vision technology and advanced path planning algorithms. By utilizing a camera strategically positioned above the crane’s hook, real-time path information during operational activities is collected. This innovative approach enhances the accuracy of the crane’s movements and integrates a safety risk evaluation module using the YOLOv8 instance segmentation framework. This model effectively identifies potential drop zones for lifted objects and the relative positions of nearby workers, enabling a dynamic awareness of risks during crane operations.

The proposed path planning module, grounded in an improved TC-DWA, autonomously generates paths that prioritize obstacle avoidance. Through comprehensive simulation experiments, effective path planning are successfully demonstrated in both long-distance and short-distance operation scenarios. The results of our evaluations highlighted the algorithm’s robust performance, confirming its ability to adapt to varying operational conditions.

Comparative analyses with traditional path planning algorithms further underscored the superiority of the TC-DWA method. It facilitates efficient route planning and significantly enhances safety measures by effectively avoiding obstacles, thereby reducing the risk of accidents involving ground workers. This dual focus on efficiency and safety positions our method as a valuable contribution to modern construction practices.

As the construction industry increasingly emphasizes the need for improved safety and operational efficiency, dynamic path planning for tower crane operations emerges as an essential area for research and development. Our findings advocate for the integration of real-time data and advanced algorithms into crane operations, which can lead to better decision-making and reduced operator workload.

There are several limitations in the current study that warrant further investigation. Although the TC-DWA algorithm provides effective path planning suggestions, the successful implementation of these suggestions heavily relies on the crane operator’s response time and ability to execute the recommended actions promptly. During crane operations, operators may face challenges in processing and responding to the perceived information in real-time, especially when multiple hazardous situations occur simultaneously. Additionally, the accuracy of vision-based monitoring systems is inherently limited by the capabilities of surveillance cameras alone. Factors such as varying lighting conditions, occlusions, and complex construction site environments may affect the system’s detection accuracy. Future research should focus on developing more automated control mechanisms and integrating multiple types of sensing devices to build a more robust and reliable monitoring system. This multi-modal approach could provide more comprehensive and accurate safety monitoring while reducing the reliance on single-source visual data.

Future work can be conducted to develop a more comprehensive monitoring system that integrates multiple sensing devices to enhance detection accuracy. By combining the proposed model and various data sources such as cameras, LiDAR, and IoT sensors, a more reliable and robust safety monitoring and path planning system can be built. Furthermore, exploring how the proposed model could be integrated with site-wide safety systems, such as collision avoidance or real-time hazard detection, could also enhance its applicability in diverse construction environments. Future studies could also focus on the human–computer interaction aspect, investigating how operators of cranes and other heavy machinery interact with these decision-support systems. This research could help inform the design of more intuitive interfaces, improving usability and increasing adoption across the construction industry. Understanding how operators respond to system-generated suggestions or decisions, and how these interactions affect decision-making efficiency and safety performance, will be crucial for refining the system’s effectiveness in practice. Finally, long-term studies are needed to assess the actual impact of deploying such models in real-world construction sites. Evaluating key metrics such as accident reduction rates, worker injury statistics, and overall operational safety before and after the implementation of the model will provide a roadmap for validating its long-term effectiveness and refining its design.