**1. Introduction**

#### *1.1. The Purpose and Significance of This Paper*

The monitoring of traffic in urban networks, whatever their complexity, is a problem that has been tackled for decades. The aim of this monitoring depends on the case and can involve managing the daily traffic flow to perform urban mobility plans. Regarding the techniques and tools to identify and quantify the vehicles on the network, traditional manual recording has been displaced by more sophisticated techniques due to their economy, and also to collect the traffic information with enough performance and quality. Basically, the emerging techniques consist of a sensor or device able to collect a type of information through its interaction with a vehicle or the infrastructure. Therefore, the sensors used for traffic analysis can be classified in different categories according to their physical characteristics, type of collected information, and position with respect to the network among others. In particular, [1] differs between in-vehicle and in-road sensors. The first are those that allow increasing the performance of the driving and the connectivity of the vehicles with their environment. In this, the concepts of communication between vehicle and the vehicular sensing networks (VSN) are called to be important in the improvement of the quality and operability of transportation systems (see [2,3]). The second are those installed in the transportation network and allows the monitoring of the performance of the system and, according to the extracted information, diagnose the problems, improve the resilient and

operational functioning, and inform the users helping them to make better choices. In this paper we mainly focus in this last. Based on the works of [4,5], in-road sensors for traffic network analysis are classified in two main groups according to the characteristics of the data collected (see Figure 1):

**Figure 1.** Classification and examples of sensors according to the characteristics of the collected data, their interaction with the vehicle and position on the road: (**a**) Inductive loop detector; (**b**) Microwave radar; (**c**) Rubber hoses detector; (**d**) Hand electronic counter; (**e**) and (**i**) Automatic Number Plate Recognition (ANPR) fixed sensors; (**f**) Bluetooth sniffer; (**g**) Police ANPR sensor on vehicle; (**h**) Police ANPR portable sensor; (**j**) Bluetooth scanning sensor.

	- - "Passive sensors" do not require any active information provided from a vehicle, i.e., they collect the information when a vehicle is passing in front of the sensor. In particular:
		- - "Passive fixed sensors" have a fixed position on the network. This group includes inductive loop detectors, magnetic detectors, pressure detectors, piezoelectric sensors, microwave radars, among others. These sensors are used to manage the traffic and can also be used to elaborate traffic mobility plans using only the already installed fixed sensors if the available budget is limited.
		- -"Passive portable sensors" have a fixed position on the network, but they are installed for a defined-short period of time. This group includes counters made with rubber hoses or manual counters that are used for example to elaborate traffic mobility plans completing the information provided by fixed sensors.
	- - "Active sensors" require active information from the vehicle to be univocally identified. In fact, these sensors can be included under the term "automatic vehicle identification" (AVI). As well as the passive sensors they can be fixed or portable:
		- - "Active fixed sensors" have a fixed position on the network. This group includes automatic number plate recognition (ANPR) sensors, Bluetooth sniffer of bar-coded

tags. Despite these sensors being designed for other purposes far from the tra ffic network analysis, recent researches have begun to use the data collected by these sensors to estimate tra ffic flows.


The data collected by the sensors can be used for multiple purposes but, since this paper is focused on the topic of tra ffic flow estimation, only those used as inputs for these models are going to be analyzed. These sensors have to satisfy two objectives: accuracy and coverage [6] and, due to their ease installation and capability of data collection, passive sensors (e.g., fixed loop detectors or portable rubber hoses) have been widely used in mobility studies in large urban areas.

As exposed above, sensors as rubber hoses count the number of vehicles that pass over it, obtaining the needed tra ffic counts used by traditional methods to estimate origin–destination (O–D), route and link flows on a network. The quality of the results of this estimate may be enough for some cases, but when the technicians or the authorities look for a better degree of observability (or even full observability) of tra ffic flows to achieve a high quality of estimation, the tra ffic count data has been proved to be not su fficient. For this, it is expected that these sensors are going to be gradually replaced by new active sensors (as ANPR) that, taking advantage of the available technology and the added value provided by the data, allows the development of models to better estimate the non-observed flows.

#### *1.2. State of the Art of Sensors for ANPR*

The automatic number plate recognition (ANPR) system is based on image processing techniques to identify vehicles by their number plates, mainly in real time (for automatic control of tra ffic rules). In [7] or also in [8] a review is made regarding the most significant research work conducted in this area in recent years.

The general process of automatic number plate recognition can be summarized in several well-defined steps [9,10]. Each step involves a di fferent set of algorithms and/or considerations:


phase, one of the most important steps of ANPR, as all subsequent phases depend on it. Another similarity with the plate detection process is that there is a wide range of techniques available, ranging from the analysis of the horizontal projection of a plate, to more sophisticated approaches such as the use of neural networks.

4. Character recognition: The last step in the ANPR process consists in recognizing each of the characters that have been previously segmented. In other words, the goal of this step consists in identifying and converting image text into editable text. A number of techniques, such as artificial neural networks, template matching or optical character recognition, are commonly employed to address this challenge. Since character recognition takes place after character segmentation, the recognizer system should deal with ambiguous, noisy or distorted characters obtained from the previous step.

Once the data is collected by the sensors, it has to be properly processed to be used for a grea<sup>t</sup> amount of tra ffic analysis. In particular, focusing on the scope of tra ffic flow analysis, the data allows to


An extra step to complement the aforementioned steps is the error recovery that may occur when recognizing plate numbers. This problem is a very important issue to deal with when plate scanning data is used for tra ffic flow estimation, which some authors have been faced using di fferent approaches [16,19,20].

However, the increasing development of these ANPR systems faces some problems such as: they are fixed sensors and they incur a high cost in terms of hardware [21] (about \$20,000 per camera) and installation and maintenance (about \$4000 per camera). This makes necessary to develop new architectural approaches that allow these types of services to be deployed on a larger scale to face transportation problems such as urban mobility plans. It is worth noting the survey collected in [22], which analyzes the sensors to monitor tra ffic from the point of view of various criteria, including cost. In this study, it is highlighted that the new sensors tend to be of reduced dimensions, of low energy consumption and that, with a certain number of them, it is possible to design and configure a sophisticated wireless sensor network (WSN) that can cover multiple observations in a certain region [23,24].

Regarding existing software libraries and tools focused on automatic plate recognition, "OpenALPR" (2.5.103) [25] stands out. This open source library, written in C++, is able to analyze images and video streams to automatically identify license plates. The generated output is a text representation that comprises the set of characters associated with each one of the identified plate numbers. The hardware required to run OpenALPR depends on the number of frames per second that the system must handle. From a general point of view, a resolution of 5–10 fps is required for low-speed contexts (under 40 km/h), 10–15 fps for medium speed contexts (40–72.5 km/h), and 15–30 fps for high speed contexts (over 72.5 mph). The library requires significant computing power, with the use of several multi-core processors at 3 GHz to process images at 480 p in low-speed contexts. From the point of view of the success rate, OpenALPR represents the software library with the best results on the market (more than 99% success in a first estimation [26]).

"Plate Recognizer" (1.3.8) [27] o ffers cloud-based license plate recognition services for projects with special needs such as di ffuse, low-light, or low-resolution imaging. The cloud processing pricing

plan offers different configurations per processing volume. There are also other specific purpose platforms for automatic license plate identification in the market, such as "SD-Toolkit" (1.2.50) [28], "Anyline" (24) [29], or the framework "Eocortex" (3.1.39) [30].

In recent years, conventional ANPR systems are strengthening their services through the use of AI techniques [31]. "Intelli-Vision" (San Jose, CA, USA) [32], the company that offers intelligent image analysis services using AI and deep learning techniques, has specific license plate recognition services that can be integrated, via an existing SDK, in Intel processors or provided as a web service in the cloud. The Canadian company "Genetec" (Montreal, Quebec, Canada) [33] announced, at the end of 2019 an ANPR camera that includes an Intel chip designed to feed neural networks improving the identification of license plates at high speed or in bad weather conditions.

Finally, it is very important to keep in mind if the ANPR systems can respect the users' privacy rights in the entire process in which the vehicle data is collected according to the different locations all along the network [34]. All this means that, when designing a type of sensor that can be implemented in an architecture that serves to monitor the traffic network, the cost criteria for manufacturing and installation, operability and resilience, and information processing must been taken into account.

#### *1.3. Contributions of This Paper*

It is being seen how the sensors based on the capture of vehicle images constitute an efficient traffic monitoring system for its features. However, there is still a challenge in terms of manufacturing and installation costs, since well-designed equipment and materials are required in terms of performance and functionality to face different network conditions [19,34]. This is a very important challenge because the large number of papers published by researchers in recent years (see [35] or [4] for a good review), stated that in order to achieve good traffic flow estimation results, a large number of sensors has to be installed. Even when trying to minimize this number the model developed in [36] proposed to install 200 ID-sensors to obtain the full observability of a real size city with 2526 links. Depending on the case of study, this can be an unaffordable cost. In addition, the sensor location models have to be designed to take into account the particular characteristics of installation of the type of sensor to be used.

Therefore, all the context exposed in this section motivates the preparation of this original paper, whose main contributions are as follows:


The rest of the paper is organized as follows: in Section 2, the proposed low-cost sensor and its associated system for traffic networks analysis are deeply described. In Section 3 the proposed system is applied in a pilot project in Ciudad Real (Spain). Finally, some conclusions are provided in Section 4.

#### **2. The Proposed Low-Cost ANPR System for Tra** ffi**c Networks Analysis**

This section deals with the description of the proposed system which is composed of three elements: (1) the proposed architecture to deploy the sensor networks, (2) the devised low-cost sensor prototype, and (3) the adopted method to decide the best set of links where the sensors have to be installed.

#### *2.1. Architecture to Deploy Low-Cost Sensor Networks*
