Comprehensive Approach to the Evaluation of Off-Line License Plate Recognition Data

Abstract

The aim of the article is to present a comprehensive procedure for processing and evaluating directional data from vehicle license plates, focusing on the specific challenges of areas that are sparsely or not at all equipped with permanently located standard license plate recognition systems. This remains a current issue, especially in smaller towns; it leads to the implementation of short-term directional traffic surveys, often using inexpensive measurement devices, in order to obtain directional traffic data. In this research, a procedure for evaluating license plate recognition data is proposed with a primary focus on its simple adaptability and automation for any subsequent use. The data sources considered are primarily the above-mentioned traffic surveys; however, the proposed evaluation procedure is theoretically transferable to any off-line data obtained from license plate recognition systems. Identifying potential inaccuracies in the data is also an integral part of the evaluation process. The design of the proposed procedure follows the Checkland soft systems methodology and the functionality of the resulting procedure was validated through a case study of a directional survey in Prague. The proposed procedure contributes to greater accuracy of the conclusions drawn from evaluated traffic engineering parameters under non-ideal, but common conditions of smaller cities, not only in the Czech Republic.

1. Introduction

Traffic data are a key component of the traffic management process—high-quality, data-driven decisions are vital for both on-line and off-line traffic management, traffic control and planning [1,2,3,4,5]. Nowadays, there are numerous data sources—from conventional traffic sensors [6,7] to mobile phone data [8] and other modern sources, which all offer great possibilities of data utilization [6,9]. However, in the actual practice of governments and municipalities in charge of transport systems, proper data management practices are still often lacking, leading to underutilization of available data sources [10]. In a broader context, deficiencies in data management processes hinder cities from successfully transforming into smart cities and data-driven societies.

This article is focused on a specific part of traffic data—license plate recognition data (LPR data), because it is a widely available but often underutilized traffic data source [3,6] (in this article, the term LPR is used not only for data from robust LPR/ANPR systems, but generally for data from any measurement devices or sources that record vehicle passage based on their license plates). The article takes into account the current conditions in the Czech Republic [11] and aims to fill the gap in the field of LPR data evaluation processes in order to maximize the potential use of this data source, including ensuring compliance with General Data Protection Regulation (GDPR). The main objective of the research is to propose a unified and comprehensive procedure for the processing and evaluating of LPR data that is broadly applicable across various applications. In particular, this article asks the following research questions:

RQ1:

How can a procedure for the processing and evaluating of LPR data be designed as a general and comprehensive methodology?

RQ2:

How can the proposed procedure be applied to the evaluation of real directional traffic survey data?

RQ3:

How can the proposed procedure be designed to support the evaluation of the widest possible range of traffic engineering parameters related to traffic directionality?

RQ4:

How can the quality issues of the source data be effectively identified during the data processing and evaluation process?

RQ5:

What is the impact of LPR data quality on the evaluated traffic parameters?

This research is based on the existing literature, in particular on [6,12,13]. The authors of [6] present the methodology for LPR data processing with an emphasis on freight vehicle movements. However, the proposed methodology is based on anonymized raw data, which limits the ability to evaluate data quality. This is where our research makes a major contribution by emphasizing quality factors and presenting an approach for identification of lower-quality data in non-anonymized LPR datasets. The authors of [12] describe a traffic survey application of LPR data measurement. We build on this work because traffic survey applications are common in the Czech Republic. There are many areas that are not sufficiently covered with LPR systems, especially rural areas, smaller towns and villages. In order to evaluate important parameters for transport planning, management and infrastructure design, such as traffic directionality or the proportion of through traffic, short-term traffic surveys are still needed. These kinds of measurements are usually conducted with portable cameras, often installed mainly on exits/entrances on the border of an investigated area, and complemented with measurement points inside the area, if needed. Our research extends the traffic survey scheme proposed in [12] and utilizes low-cost measurement devices such as smartphones. In [6,12], the authors present a variety of possible traffic engineering parameters, which can be derived from LPR data and which we incorporate in our research.

At the same time, this article develops the findings of the final thesis methodology of evaluation of directional surveys [13], where the author addressed a similar issue. In the area of data measurement, the authors also follow the current knowledge and best practices, taking into account the methodology of the Ministry of Transport of the Czech Republic [11].

The article follows the structure below: Section 2 presents the thorough literature background on which this research is based. Section 3 provides description of the methodology used; the research mainly follows Checkland’s soft systems methodology. Section 4 contains the results; the main result is the proposed LPR data evaluation procedure. The presented procedure was verified using data from a case study of directional traffic survey. The traffic survey was selected for two reasons: Firstly, acquiring raw LPR data from robust camera systems is problematic due to GDPR and related personal data protection legislation. Secondly, the aim was to use a low-cost, widely available data measurement method that that would be easily adaptable for potential clients such as municipalities. Section 5 presents a discussion that specifically address the fulfilment of the previously defined research questions and declares possible limitations of the research and future work possibilities. The article is concluded in Section 6.

2. Literature Background

The Literature Background Section provides an insight into the state of the art in the measurement and use of LPR data. The general topic of the use of LPR data is presented, as well as detailed current practices and technologies used to measure and process LPR data and the quality factors of LPR data. All of this is then contextualized with respect to the conditions of the Czech Republic, including compliance with legal rights and GDPR. Finally, the section is concluded with a summary relevant to the research topic.

2.1. Application of LPR Data

Nowadays, with the constant development of road transport and the increasing degree of motorization in Europe and worldwide [14], the sustainability of transport is receiving growing emphasis, and methods for its effective management and control are being actively explored [15,16]. The basis for successful traffic management is the availability of sufficiently high-quality data sources that provide on-line and historical information on the state of traffic [1,2,17,18]. Currently, due to the constant technological development, a variety of traffic data sources are available, e.g., sensors permanently installed on road infrastructure [7,19], vehicle sensors directly providing data from individual vehicles [9] or indirect sources of a population’s mobility information as location data from mobile phones [8].

In this article, the authors focus specifically on LPR data—data obtained by recording vehicle license plates. The main advantages of this data source include wide availability and the ability to uniquely identify each vehicle. These factors provide the potential for highly accurate traffic status information that can be utilized for traffic management [3,4,20,21]. The authors focus on LPR data collected in usual traffic flow conditions (driving vehicles), in defined measuring profiles on the road network. Nowadays, LPR data measured in this way are commonly used in security surveillance, section speed measurement and in other sub-systems [22,23]. However, the potential is much wider, as illustrated by recent research; for example, the article [4] explores the use of LPR data for traffic control optimization. Ref. [24] investigates directionality research, specifically vehicle commuting patterns, and [25] investigates reconstruction of vehicle routes. In [26], various research trends are reviewed, including origin–destination relations, travel times and commute times. While travel times and commute times are nowadays also commonly determined based on other traffic engineering data sources—e.g., FCD data [27]—information about specific vehicle routes, traffic directionality in an area and transit traffic share is currently quite difficult to obtain from commonly used traffic data sources other than LPR. FCD data is an option, but only for the sections for which this data is available [28]. Another possible source, cellular phone data, shows limited localization accuracy on the road network due to the dependency on BTS cells [29]. LPR data in this area represent a promising contribution and therefore are addressed in this research. This article does not address the use of LPR in parking systems, which is important; however, different approaches are used due to detection of motionless or low-speed vehicles [30].

The potential of LPR data is not fully exploited in practice today. Possible applications are often limited by the relatively sparse coverage of the road network with appropriate sensors (cameras). A number of articles demonstrate the applicability of macroscopic traffic analysis in larger areas [3,6,24,31,32]. However, use of more detailed traffic analysis in small areas is only possible with much denser road network coverage [4,20,21,33,34]. Another factor affecting the use of LPR data is the need to ensure data privacy, which is a global issue [35,36] and in the European Union is governed by the GDPR regulation [37]. The quality of LPR data sources represents another factor; despite current technological advances, satisfactory accuracy, completeness and therefore reliability of data is not always achieved due to the used technologies, financial resources, etc. Scientific sources approach this issue in two ways: Some automatically assume the accuracy and completeness of LPR data and do not (or only marginally) address the issue of quality [3,38], because it is not essential for their research purpose. In contrast, other scientific sources admit a reduced real quality of LPR data [4,33,39].

For the purpose of this research, data-quality-related terminology was defined in accordance with [40], which defines data quality as a concept that encompasses several dimensions and determines how well a dataset matches the real state. This research mainly works with the dimensions of completeness and accuracy. Completeness compares the dataset to reality and deals with completely missing data. Accuracy addresses errors and ambiguities in the existing (non-missing) data.

2.2. Measurement and Processing of LPR Data, Limiting Factors

The first prerequisite for the successful use of LPR data is the existence of a sufficient number of measurement profiles according to the purpose measurement. This issue is addressed in sources [41,42].

The second prerequisite is a sufficiently high-quality record of license plates [43,44]. LPR data are commonly collected by video recognition technology, which consists of capturing video footage with cameras and using software tools that recognize the license plates from the imagery [44]. Image processing and license plate recognition can be performed on-line during the measurement [3,45] or off-line using stored video recordings [46,47].

The quality of the software LPR process is influenced by a number of external factors related to both the technical parameters of the recording equipment used (especially image resolution, fps and failure rate) [23,26] and situational aspects at the measurement profile locations [26,48]. The situational aspects include the angle of the camera; the vehicle time in the frame and vehicle speed; the position and clarity of the license plates; and the traffic density and associated occlusions. Specific layout and shape of characters on the license plate can be problematic as well, especially in traffic flow with multi-type license plates [49]. The lighting conditions and weather also have a significant influence [26,48]; for example, infrared illumination is required at night. Overexposure from sunlight, sharp light–shadow transitions and reduced visibility due to rain, fog or snowfall also have a negative influence.

Current research focused on increasing the quality of LPR data by improving the reliability of recording devices and eliminating license plate recognition errors [45,50]. Robust software algorithms capable of processing lower-quality video (from common surveillance cameras, etc.) are also developed [48,51,52]. Currently, a number of sources declare a license plate recognition success rate of over 95% [44,45,50]; however, these values, obtained under test conditions, may not reflect actual performance in real-world operations [4,53,54].

Therefore, a prerequisite for the successful practical use of LPR data is their appropriate processing and evaluation, using procedures that avoid introducing new errors into the data and minimize the influence of existing errors on the interpretation of the results [26]. A specific challenge is the evaluation of vehicle directionality [25,55]. An erroneously recognized license plate cannot be matched with records of the same vehicle in other measuring profiles without applying correction methods. This leads to misdetermination of the vehicle routes and affects the overall quality of the evaluation results. The methods of correction of incorrectly recognized license plates in relation to the determination of the vehicle route are discussed in the following articles: [25,33,55,56]. Without access to ground-truth data, the scope for effective error correction is significantly constrained [33,56].

2.3. Specifics of LPR Data Usage in the Czech Republic

In Czech conditions, common applications of LPR systems are mainly enforcement systems (section speed measurement, red light violence detection, etc.), surveillance systems or parking systems [32]. However, the possibilities of obtaining data from existing LPR system are usually limited primarily due to insufficient distribution of these systems, as well as due to personal data protection regulations and vendor lock issues. For these reasons, so-called directional traffic surveys are still used nowadays to detect directional traffic characteristics in specific defined areas [11,57]. Traffic surveys in general are a tool for short-term detection of current traffic condition characteristics in locations where the required data are not available from permanently installed sensors or other data sources. They are often commissioned by municipalities as part of transport system development efforts [34,58].

In practice, the measurement and evaluation of LPR data in directional traffic surveys are carried out as described above, with the following specifics: The data are measured only in a limited time range, so the results are to some extent indicative of the inspected area. The methodology, measurement technology and data evaluation are influenced by factors such as the financial budget and other requirements of the contractor, as well as the real-life practice of the hired company [58]. Portable cameras or other low-cost video recording devices are often used, either temporarily installed on the infrastructure [59] or placed on conventional tripods, and are physically supervised by trained personnel throughout the measurement period [57]. However, measurement profiles designed in this way exhibit higher error rates and recording failures; the average license plate recognition rate can drop to 80%, for example [60], which has a negative impact on the results.

Therefore, the practical experience shows that it is both appropriate and necessary to explore approaches on how to achieve the best possible results even under the described conditions, and how to optimize the procedures of processing and evaluating LPR data from traffic surveys.

2.4. Literature Background Summary

The literature review confirms that license plate recognition is an important traffic data source, and research is focused on various possibilities of utilizing this data. However, despite recent technological advances, various limiting factors affect measurement and processing of LPR data and can negatively affect evaluated traffic characteristics. To mitigate the negative impacts of the possible lower quality of measured data and to ensure comprehensive LPR data evaluation, a new data processing and evaluation procedure is proposed in this article. A tabular comparison of the contributions of this article compared to existing published approaches can be found in Appendix A. The comparison table includes a representative sample of state-of-the-art studies that in various ways address the topic of LPR data evaluation.

3. Materials and Methods

As part of the research, Checkland’s soft systems methodology [61] was used to define the goals and design the LPR data evaluation procedure. The initial assumptions and expected outcomes were defined first, and then the procedure itself. The search and description method was used for defining the individual sections and the complete procedure is a result of logical deduction, where the effort was first to derive logical connections from acquired knowledge and experience and then to verify these on a specific model. To verify the entire procedure within the case study, a practical experiment method was performed with the use of LPR data obtained from a directional traffic survey. The authors also utilized their practical experience with the implementation and evaluation of directional surveys in the conditions of the Czech Republic [34,62].

In the first phase of the research, the initial assumptions were defined. In particular, the specific conditions for the use of the proposed LPR data evaluation procedure were clearly defined. It was determined which part of the whole process of measuring and processing license plate data the proposed procedure covers and which inputs it works with. Figure 1 illustrates the scope of the research within the broader context of the directional traffic survey process.

At the input, the authors assume a measured and checked, possibly corrected, set of LPR data in electronic form of one or more files, which are easily processable in tabular form (e.g., csv format). The input dataset contains individual records of vehicle passes: ID, timestamp, license plate number of the vehicle, measurement profile ID and, optionally, vehicle category, which expands the evaluation possibilities. Thanks to this general input data specification, the universality of the proposed procedure is ensured—particularly with respect to varying numbers of vehicle records, as well as measurement profiles and their various locations. In this research, the authors work with the following related terminology (see

Table 1. Terminology specification.

Term	Definition
LPR data	In this article, the term LPR is used not only for data from complex LPR (ANPR) systems, but generally for data from any measurement devices or sources that record vehicle passage based on their license plates.
Measuring profile	Location on the road where license plates in one direction of travel are recorded.
Measuring site	Pair of measuring profiles close to each other; used in the case of two-way traffic, one profile for each direction of travel.
Entry (exit) profile	Measuring profile on the border of the area.
Internal profile	Measuring profile inside the area.
Trip section	Sequence of two consecutive measurement profiles that the vehicle has passed through; trip section is given by license plate matching.
Stay	Significant interruption of the journey due to the destination being in the inspected area.
Trip	Movement of a vehicle through the inspected area without a stay.
Route	Sequence of all consecutive measuring profiles that the vehicle has passed in one trip.
OD matrix	Two-dimensional matrix, where the value of each cell denotes the number of trips between the origin measuring site (row index) and destination measuring site (column index).

):

Table 2. Input data specification.

ID	Timestamp	License Plate	Category	Measuring Profile ID
1	8:00:09	5L913XX	VAN	A1
2	8:00:17	2SJ36XX	VAN	A1
3	8:00:56	7AT45XX	CAR	A1
4	8:01:00	EL113XX	CAR	A1
…	…	…	…	…

presents an example of an input dataset from the case study; the license plates were partially anonymized for the purposes of presentation in the article. The proposed procedure assumes that the anonymization of license plates will occur later as part of data processing. Access to original license plate strings is important for assessing the quality of data and results.

The second assumed input is a description of the quality of the input dataset. Information about the number of data files, their format and their structure is provided, as well as general information about the inspected area and about the employed method of data measurement. For example, the location of measuring profiles on the road network, the time period of the measurement, the technology used for license plate recording, weather conditions and all significant abnormalities regarding traffic, weather or possible failures of measuring equipment are provided. Information on the method of input data checking and a description of the reference dataset used are also important. Information about identified time segments in the input data set (in connection with measuring profile ID) that are more likely to contain inaccurate or incomplete data, if it was not possible to correct them, is essential.

In the second phase of the research, it was defined which outputs must the proposed LPR data evaluation procedure enable to evaluate. The scope and list of these outputs were based on the existing literature and research in this area, as mentioned in the previous text, and are included in the Section 4. All specific outputs were designed for use in the widest possible portfolio of traffic engineering applications. In the conditions of the Czech Republic, this means mainly medium- and long-term transport planning and implementation of various transport measures.

In the third phase, the research was focused on the definition of the specific LPR data processing procedure, with an emphasis on its comprehensiveness, universality, optimization, clarity and unambiguity. The defined procedure respects practical experience and established general practice in evaluating LPR data. It is based on the following blocks: data preparation; vehicle matching and quality check; determination of vehicle routes; final evaluation; and results. The detailed description is included in the Results Section. The procedure was designed to be user-friendly, facilitating straightforward algorithmization and automated software implementation—thus enabling efficient evaluation of LPR data.

The procedure was designed to work with the information about the input data quality and to identify potentially erroneous data that would negatively affect the evaluation outcomes. In order to achieve the highest quality of evaluation outcomes, the procedure specifies the points at which additional data checks should be performed. For these additional quality checks, the procedure recommends manual verification with the use of original video recordings/vehicle screenshots if these are available. If the errors found in the data cannot be corrected, the proposed procedure ensures their documentation and enables the evaluation of their impact on the reliability of the results. Emphasis is also placed on eliminating secondary errors during data processing and evaluation. The goal is to reconstruct the actual traffic situation in the inspected area as accurately as possible, with an emphasis on the directionality of the vehicles, which is the added value of the LPR data.

The proposed LPR data evaluation procedure was verified in a case study in a selected area in the city of Prague. The goal was to verify the functionality and suitability of the proposed procedure. The case study was conducted as a directional survey (see Figure 2). The inspected area consists of a compact road network where a relatively low number of measuring profiles (8 border and 1 internal) enables the accurate identification of all directional traffic movements.

The method of video recording on cell phones was used, as today cell phones represent low-cost, widely available recording devices with sufficient video-recording quality. Specifically, Samsung A02s with the OpenCamera app were used (Samsung Electronics Czech and Slovak s.r.o., Prague, Czech Republic, recording parameters: FullHD 1920 × 1080, 120 fps, manual focus). Subsequently, license plates were recognized using specialized SW (AnprGUI 2.1.1., Eyedea Recognition s.r.o., Prague, Czech Republic). Each measurement profile was under constant surveillance of a trained worker, who ensured the correct calibration of the video recording (resolution, brightness, angle, zoom). During the measurement, the weather was ideal for recording—cloudy, i.e., without the possible negative effects of uneven sunlight, without rain and with good visibility. There were no extraordinary traffic incidents.

The measured data was prepared according to the required input data format (example in Table 2), checked against a reference set of original video footage, and correctable errors were corrected.

4. Results

The primary result of the research is a comprehensive procedure for evaluating off-line LPR data (e.g., directional traffic surveys). This procedure scheme is shown in Figure 3 and described in detail in the following text.

The proposed procedure was verified in a case study, the results of which are described in the second part of this section.

4.1. Procedure for Evaluating LPR Data

The procedure for evaluating LPR Data consists of four main steps – data preparation, vehicle matching and quality check, determination of vehicle trips, final evaluation and results. All the steps are described in more detail below.

4.1.1. Data Preparation

The input for the processing and evaluation of LPR data is the dataset described in Table 2 and a description of its quality. If the data are contained in several individual data files, they will be merged to facilitate subsequent processing. An important moment is the formal validation of the dataset’s structure and format. The purpose is to check whether the structure of the contained data corresponds to the assumptions in Table 2 and whether the input data corresponds to the accompanying quality description.

4.1.2. Vehicle Matching and Quality Check

Data preparation is followed by a key part of the entire process of vehicle directionality evaluation, the data matching. The purpose is to search for consecutive pairs of records of the same license plate in the merged dataset. Each pair indicates one sub-section between the measuring profiles, where the vehicle provably moved during the measurement period. If we denote the total number of records of one license plate as N_i and the number of pairs of the same license plate as M_i, the following applies: M_i = N_i − 1. In specific cases where the vehicle was recorded only once during the measurement period, M_i = 0. The matching output is then the sole original record of the given license plate and information about non-matching.

The matching procedure can be performed several ways. It is possible to use basic approaches, which match only strictly identical license plate records. The advantage of these basic approaches is their simplicity and the possibility of using them even in the case of semi-anonymized LPR data, which are characterized by the replacement of license plate strings with encrypted IDs, by ensuring that the same license plate is always assigned the same ID. The disadvantage of the basic approaches is the negative impact on the quality of the matched data in case of errors in the license plate recognition algorithm (e.g., in case of character confusion). More advanced approaches, therefore, use an algorithm of matching license plates based on their similarity, not identity. The advantage of these approaches is the possibility of correcting errors introduced into data by erroneous license plate recognition. The disadvantages are higher calibration requirements and non-applicability in case of any kind of anonymized license plate data. The two mentioned approaches are discussed in more detail in [12]. The specific method for erroneous data matching without any ground-truth data, which is based on the edit distance and match probability, is presented in [56]. The authors of [56] also address the problem of determination of threshold value to distinguish correct and false matches.

In the case study presented in the Results Section of this paper, a simple automated algorithm of double sorting was used; all records were sorted according to license plate numbers and time parameter. After sorting, consecutive records of the same license plate determine individual pairs. While this approach belongs to the basic approaches, it fully enables the validation of the proposed data evaluation procedure.

Data matching results are clearly displayed in the matrix of matched data, where each record pair is displayed in the relevant field according to position of the measuring profiles (entry/exit/internal). Each record pair indicates a vehicle trip section. Unpaired records are displayed in “non-matched vehicles” in the relevant row. The generalized form of the matrix is shown in Figure 4. The matrix of matched data shows the location of the trip sections in relation to the investigated area. At the same time, the matrix enables the assessment of the input data quality. The interpretation of matrix of matched data is as follows:

Green cells show pairs representing sections of vehicle trips demonstrably within the inspected area. These trip sections are considered to be correctly determined, except in specific cases in which the outcomes grossly disagree with the assumptions based on the topology of the road network.

Yellow cells show external trip sections that run outside the inspected area. They are considered as correctly determined, but are not subject to evaluation.

Red cells show illogical trip sections, which in most cases indicate the presence of errors. These are repeated entries, repeated exits or actually impossible sequences of vehicle records. These errors are usually caused by an inappropriate distribution of the measuring profiles on the investigated road network, when not all border points were covered. The second possibility is measurement errors that lead to incomplete or inaccurate input data. The third possibility is a repeated recording of the vehicle due to maneuvers like turning in the vicinity of the measuring profile. This case is characterized by a small time difference and this is logically not an error but a case requiring attention in interpretation.

Blue cells show non-matched license plates.

After compiling the matrix of matched data, an evaluation of data quality is carried out. The proportion of illogical trip sections is evaluated. If it is not negligible with regard to the purpose of data evaluation, it is necessary to check them in the reference video recordings with a focus on the reason for their occurrence. Furthermore, the consistency between the number and spatial distribution of non-matched vehicles and trip sections—represented in the green fields of the matrix—and the actual topology of the road network is assessed. In case of a gross discrepancy, to ensure high reliability of the results, it is necessary to re-check the suspicious records in the paired and then the input dataset against the reference data.

4.1.3. Determination of Vehicle Trips

This step covers the determination of complete trips taken by the vehicles without significant interruption in their journey due to the destination being in the inspected area—hereafter referred to as stay. The trips are determined from the matched data (see Figure 4) and kept in a separate dataset, as they are important for the final evaluation and final results determination. Each trip contains information about the route of the trip; the term route is used in this article to refer to the sequence of measuring profiles that the vehicle has passed through during one trip.

Firstly, pairs of records at the same measuring site are assessed. In cases where the time difference is at most a few minutes, turning or other non-standard maneuvers of the vehicle possibly happened in the measuring site vicinity, not an intentional stay. After checking against the reference dataset, such records are assigned to the same trip and remaining pairs are split into two trips. The following criteria do not apply to them.

External trip sections (yellow cells in Figure 4) are divided into two trips—vehicle movement outside the inspected area is not subject to evaluation.

Illogical trip sections (red cells in Figure 4) representing irreparable errors in the data are also divided into two trips—these pairs logically indicate movement of the vehicle outside the inspected area. Trips created in this way, whose routes start at an internal or exit measurement profile, or end at an entry or internal profile, are marked as incomplete.

Non-matched records (blue cells in Figure 4) each form a separate trip with a route that contains only one measuring profile.

The trip sections inside the inspected area (green cells in Figure 4) are evaluated according to the time differences of pairs of records. In order to distinguish stay from direct through travel, a limit time T_limit is set for each relevant pair of measurement profiles. If the time difference of the pair is less than T_limit, it is a through travel, and both vehicle records are assigned to the same route. Otherwise, the records are split into two different routes. Although recent research [57] shows a shift away from expert-determined parameters towards a universal algorithm, T_limit is usually determined using a combination of statistical and expert methods, which consider specifics of the inspected area, including daily traffic variations. Procedures presented in [6] and [12] can be used, where the authors use the average values of the time differences of paired records and determine T_limit by expert estimation. More complex methods used in practice, also based on a combination of statistical and expert approaches, are presented by [63]. In the case study in this research, the median of time differences was used and T_limit was determined expertly in order to filter out outliers that are considered as vehicle’s stay (see Section 4). The use of the median of time difference values ensures the sensitivity of the method to different road section lengths because the median is computed specifically for each road section. At the same time, if data measurement takes place throughout the day, the median needs to be calculated several times, at least separately for peak and off-peak hours. In this way, the median will take into account the travel time of vehicles in free-flowing traffic as well as in congestion. T_limit is determined as a suitable multiple of the median time differences so that the specific characteristics of the road section are taken into account; for example, in the case of a shorter straight section where the variation in travel times is minimal, a lower multiple of the median will be selected. For a longer section that includes, for example, a level crossing with a railway, a higher T_limit will be selected.

The dataset of trips, compiled according to the above criteria, is considered as accurate and complete as possible because all possible data corrections have already been conducted. In this step, anonymization is performed by replacing each license plate with a unique identifier. This makes it possible to still identify all movements of the same vehicle in the inspected area during the measurement. However, it is not possible to determine the license plate number or the vehicle operator.

An OD (origin–destination) matrix is created for the dataset of trips. The authors define the OD matrix in this context as a two-dimensional matrix, where the value of each cell denotes the number of trips between the row index (origin measuring site) and column index (destination measuring site). The OD matrix is similarly defined in [64]. It is therefore a representation of the number and directionality of real vehicle trips in relation to the first and last measuring profile where a specific trip was recorded [6]. The general form of the OD matrix is in Figure 5, with the measuring profiles related to the respective measuring sites. If the vehicle trip is determined by only one measuring profile, it is displayed in the appropriate row or column “inside the area”. The sum of the values in all cells of the matrix must equal the total number of trips in the time interval of the OD matrix. The OD matrix enables the direct determination of vehicle trip types and their ratio in relation to the inspected area—through trips, destination trips, origin trips and inner trips. Closer attention should be paid to the trips previously determined as incomplete, as well as trips on the main diagonal, which may be caused by non-standard vehicle movements. In this case, the trip type must be evaluated separately, based on the specific route of the vehicle—measuring profiles where the vehicle was recorded.

To evaluate the reliability of the OD matrix, it is necessary to consider its relationship with the matrix of matched data, particularly in cases where it was not possible to check and correct possibly incorrect pairs of records. Incorrect data matching is reflected in both the OD matrix and the dataset of trips and affects the final results of the data evaluation.

4.1.4. Final Evaluation and Results

Based on the input dataset, the matched data dataset, the trip dataset with vehicle routes and the OD matrix, a final evaluation is performed. The goal is to determine results from the measured data that provide the maximum amount of relevant traffic engineering information for further use. A comprehensive overview of possible results is contained in

Table 3. Overview of results.

Characteristic	A More Detailed Description of the Traffic Engineering Characteristics		Related Literature
Traffic volume	In total; according to the vehicle categories; time variation.	In measuring profiles; in measuring sites; in routes.	[6,65,66]
Traffic flow composition	Ratio of vehicle categories.	In measuring profiles; in measuring sites; in routes.	[66,67]
Traffic directionality	Count; ratio.	In the entire area; in measuring profiles; in measuring sites.	[6,12,66,68]
Route types (through, origin, destination, inner)	Count; ratio.	Throughout the area; in measuring profiles; in measuring sites.	[12,57,68,69]
Matched and non-matched license plates	Count; ratio.	In the entire area; in measuring profiles; in measuring sites.	[6,70,71]
Unique vehicles	Count; ratio.	In the entire area; in measuring profiles; in measuring sites; in routes.	[72]
Stay duration	Count; time duration; average values.	In the entire area; in partial areas.	[6,38]
Travel time	Time; average values.	In routes.	[6,12,20]
Speeds	Speed; average values.	In routes.	[25,33,71,72]

. In practical applications, depending on the purpose of the measurement, it may not be necessary to evaluate all of the listed results. The Related Literature column lists sources that already utilize these traffic engineering characteristics.

The final results include a quality assessment—how well the results reflect the reality and ratio of (possible) errors. Input data quality rate implicates the quality of outputs. For quality evaluation, the interpretation of the following phenomena is essential:

• Illogical pairs of matched records and the trips and routes that contain them;
• Unexpected ratio of matched and unmatched license plates;
• Routes or trips that do not meet expectations in terms of directionality or quantity given the topology of the road network;
• Routes containing subsequent records at the same measuring site;
• Extreme values in driving times, driving speeds and stay durations.

Level of detail in the quality evaluation is given by the existence of a suitable reference dataset that enables checking and correction of the above-mentioned phenomena. Ideally, the quality of the resulting traffic engineering characteristics is evaluated numerically (percentage), not only in summary for the entire measurement period, but in a more detailed time frame.

4.2. Case Study

The input dataset obtained from the conducted measurement follows the structure defined in Table 2 and consists of nine separate csv files—one file for each measurement profile, with a 1 h measuring interval. The reference dataset consists of original video recordings. The input dataset contains a complete number of records; however, some records could not be clearly checked against the reference videos (due to occlusions, deteriorated video quality, etc.).

Table 4. Numbers and quality of records in the input dataset.

Number of Records	Measuring Profile
	A1	A2	B1	B2	C1	C2	D1	D2	E1	Total
Total	354	181	706	474	157	576	384	342	200	3374
Records with (potential) error	8	11	0	3	7	0	0	1	1	31
Proportion of correct records [%]	97.7	93.9	100.0	99.4	95.5	100.0	100.0	99.7	99.5	99.1

presents the total number of records, records with (potential) errors and the total percentage of records clearly determined to be correct after checking. However, following a secondary validation using the matched data, the proportion of correct records in the original input dataset was revised downward to 97% (see the related discussion at page 15).

The process of processing and evaluating the input dataset followed the procedure described in Figure 3. An R algorithm was created and used for automated processing. After matching the data, a matrix of matched data was created (Figure 6).

A total of 3170 records were matched, with 1755 record pairs found. A total of 204 records could not be matched. A more detailed analysis shows that out of the 1755 pairs found, 1662 are green trip sections, demonstrably running inside the inspected area. In contrast, there are a minimum of illogical red trip sections, 15, which is only 0.89% compared to green trip sections. The external yellow trip sections are not relevant to the evaluation of the inspected area.

The case study was conducted primarily for research purposes, and thus a rigorous verification process was performed. All checked pairs are marked in italics in the matrix of matched data. The check covered illogical trips, unpaired vehicles (the rather high amount is suspicious, given the closed area with a small number of parking possibilities) and pairs in unexpected relations given the topology of the road network. For example, there was strong assumption that between measuring profiles A1 and C2 (or D2), vehicles would be recorded on the E1 profile (see the map in Figure 2).

The dataset of matched data and then the input dataset were checked against the reference videos. A total of 100 errors were discovered in the input dataset despite the previously performed check. This outcome is noteworthy, as it suggests that manual checking against the reference recordings can be even more effective when conducted after the initial data matching. All identified errors were corrected. They primarily involved the incorrect recognition of license plates from the reference video recordings.

Table 5. Reflection of the numbers and quality of records in the input dataset.

Number of Records	Measuring Profile
	A1	A2	B1	B2	C1	C2	D1	D2	E1	Total
Original Input Dataset—Original Evaluation
Total	354	181	706	474	157	576	384	342	200	3374
Records with (potential) error	8	11	0	3	7	0	0	1	1	31
Proportion of correct records [%]	97.7	93.9	100.0	99.4	95.5	100.0	100.0	99.7	99.5	99.1
Original Input Dataset—Updated Evaluation
Total	355	186	706	473	156	576	384	343	201	3380
Records with (potential) error	18	52	4	5	13	2	1	2	3	100
Proportion of correct records [%]	94.9	72.0	99.4	98.9	91.7	99.7	99.7	99.4	98.0	97.0

provides a revised evaluation of the input dataset quality, reflecting the newly determined quality of the input dataset originally presented in Table 4.

After the correction, the input dataset was considered to be correct. Data were matched again and the resulting vehicle trips and routes within the inspected area were determined according to the criteria. Figure 7 illustrates an example of using the median time difference of paired records to expertly determine T_limit for distinguishing through travels from stays in the area. Figure 7 depicts a road section B1-D2 (see Figure 2) with a bus stop; therefore, the time difference values show corresponding variance. The evaluated period was 60 min with no significant traffic variations; therefore, a single T_limit value was calculated. Figure 8 presents the final OD matrix.

A total of 1634 trips were determined. The inspected area contains a minimum of parking possibilities; therefore, a more detailed analysis of the origin and destination trips was performed to justify their occurrence: Thirty-nine trips correspond to the impossibility of vehicles matching at the beginning and end of the evaluated time interval. Thirty-nine trips belong to buses, which is logical because the final stop of the bus lines lies in the inspected area. Twenty-six trips are origin/destination trips of other types of vehicles. Four “destination” trips are incomplete trips created by splitting an illogical pair of matched records. In fact, these are through trips, but the input data check showed that these vehicles went around the measuring profiles, so it is not possible to correctly assign exit profile.

The above specifics were taken into account when evaluating the ratio of through, origin and destination trips (internal trips were not found). The first and last 5 min of the measuring interval are not included, so that the result is not distorted by unmatched records. Incomplete “destination” trips were considered to be through. The resulting ratio shows 96% through trips, 2% destination trips and 2% origin trips, and corresponds to the topology of the area.

Furthermore, an evaluation of the selected traffic engineering parameters from Table 3 was performed. The selection was based on an analysis of commonly evaluated characteristics from the practice of traffic engineering studies in the Czech Republic [13]. An illustrative example of evaluated characteristics follows in

Table 6. Illustrative example of the evaluated results—traffic volume according to the vehicle categories.

Vehicle Categories	Measuring Profile
	A1	A2	B1	B2	C1	C2	D1	D2	E1
Passenger cars	297	153	579	401	119	487	329	272	168
Vans	36	14	88	48	21	62	32	43	23
Trucks	8	5	20	6	3	13	9	16	6
Buses	9	8	15	14	9	8	11	12	0
Motorcycles	5	6	4	4	4	6	3	0	0

and Figure 9. Furthermore, it was verified that the output dataset evaluated according to the presented procedure allows for the evaluation of all the characteristics proposed in Table 3. This verified the applicability of the proposed procedure for evaluating LPR data.

5. Discussion

Regarding the research objective and questions, the main goal was fulfilled—the design and verification of a comprehensive LPR data evaluation procedure. The procedure respects conditions of directional traffic surveys, supports software automation and enables a wide scope and high level of detail of evaluated traffic parameters. The procedure is generally applicable to areas of various structure and size and allows for future updates. Given the fact that the procedure describes in detail possible manifestations of errors, it contributes to increasing the reliability of the LPR data evaluation process and increasing the quality of the evaluation outputs.

However, the data quality issue currently represents the main limit of our research. In practice, an accurate and complete reference dataset is often lacking or the evaluation process is burdened by financial and time constraints. Both make a comprehensive data check impossible. Therefore, the proposed procedure does not strictly demand detailed checks. However, the quality of the results should be at least roughly determined (e.g., using probability). The following indicators derived from the research can be used.

5.1. Knowledge of Error Sources in Measurement and Data Preparation

The quality (accuracy and completeness) of vehicle records and license plate recognition is the pitfall of the entire subsequent evaluation of LPR data.

To eliminate errors in this field, methodological procedures for optimizing data measurement should be strictly followed, in Czech conditions mainly [11]. It is also recommended to obtain a reference dataset that allows for the checking and correction of input data, and then for checking possibly incorrect matched data. As the case study showed, if the goal is to achieve high-quality results and the data check is performed manually, it should be repeated after the data matching step.

Implementation of cost-effective processes of correction for already measured data in order to achieve maximum quality of the results is the subject of currently ongoing research.

5.2. Data Matching Results (Interpretation of the Matrix of Matched Data)

It is necessary to take into account the input assumptions based on the topology of the investigated road network. If the data in the matrix of matched data differ significantly from stated assumptions, the presence of incomplete or inaccurate input data can be inferred. Verification and correction are advisable; if this is not possible, it is appropriate to describe the amount, types and possible sources of errors. Based on this, final results of the evaluation are interpreted.

The case study (see Figure 6) provides an example, with the detected suspicious values marked in italics. In particular, a high number of non-matched vehicles was identified, which is considered potentially erroneous given the inspected area topology with a minimum of parking areas. There are also illogical pairs of records—repeated entries and exits, as well as illogical combinations of boundary and internal measuring profiles. Furthermore, direct routes A1-C2 and A1-D2, as well as pairs of entry and exit measuring profiles at the same measuring site (e.g., A1-A2), were also determined as potentially incorrect.

As part of our ongoing research, the aim is to gain detailed knowledge and description of error propagation from an input dataset during data matching and subsequent trip determination. Assessing the correctness of data matched data results across different road network topologies is a non-trivial task and remains an open challenge.

5.3. Route Determination Results (Interpretation of the OD Matrix)

The interpretation of the OD matrix quality is closely related to the interpretation of the matrix of matched data, especially in the case when potentially incorrect pairs were not checked and corrected. Special attention should be given to trips resulting from the splitting of non-logical trip sections. If it was impossible to correct these trips, even though a complete reference dataset was available, the probable reason for their occurrence is that not all boundary points on the road network were covered. Trips on the main diagonal of the OD matrix require attention as well, because these are non-standard and may indicate, for example, an inappropriately chosen time limit for the through traffic differentiation.

If the datasets were checked based on the matrix of matched data, they can be considered as accurate as possible, and it is usually no longer possible to clearly uncover further errors from the OD matrix.

However, all illogicalities in the OD matrix indicate potential errors and will be recorded for the interpretation of the final evaluated results.

As for possible future development of the proposed procedure, three main directions have been identified:

Firstly, greater utilization of the vehicle type parameter and other automatically measurable characteristics for a more detailed evaluation of traffic engineering parameters and for input data quality control.

Secondly, the determination of the time threshold T_limit for distinction of stay from through travel. As current research confirms, this is a complex issue influenced by a number of factors related to the driver’s behavior; the definition of the terms origin, destination and through traffic; and other factors regarding traffic behavior. The characteristics of the inspected area, the structure of the road network and housing development, as well as the time of day and the state of traffic (congestion, etc.) also matter.

Thirdly, adapting the procedure for potential continuous LPR data measurement using permanently installed systems. In this context, it is possible to develop the issue of anonymization methods in order to preserve the maximum range of evaluable traffic parameters, particularly the ability to retrospectively analyze the movement of distinct vehicles in the inspected area, while ensuring full compliance with the personal data protection regulations. Furthermore, the development of real-time quality assessment methods is also a promising direction.

6. Conclusions

This article presents research proposing a comprehensive approach to the processing and evaluation of LPR data, primarily for the purpose of evaluating directional traffic surveys. The proposed procedure identifies potential data errors throughout the evaluation process and enables a straightforward algorithmization and software automation. The functionality of the procedure was successfully verified in the case study of a directional survey in Strahov, Prague. All steps of the proposed procedure were applied to the case study dataset; the OD matrix was determined and common traffic parameters were evaluated. For illustration, the results show 96% through trips, 2% destination trips and 2% origin trips, and correspond to the topology of the area. It has been verified that the presented procedure allows for the evaluation of all important traffic parameters. Thanks to the inclusion of quality evaluation steps, the proposed procedure can contribute to greater accuracy of conclusions drawn from evaluated traffic parameters.

Future research will focus on the limitations associated with input data quality and the manifestation of errors throughout the data evaluation process, and the impact of lower quality input data on the quality of individual outputs. Enhanced evaluation of LPR data can play a key role in supporting cities with informed decisions and rational planning of future transport measures. In the broader context, enhanced LPR data evaluation can contribute to more effective traffic management, reduced congestion, increased road capacity and mitigation of the negative environmental effects of road traffic. As the LPR data are collected via cameras, future research can be significantly enriched by focusing on utilization of single detection system video data for additional purposes, such as detecting the number of occupants in a vehicle, monitoring driving behavior or analyzing vehicle trajectories at intersections to estimate potential collision risks.