Artificial Intelligence (AI) System for Traffic Flows Monitoring. Evidences from the Interreg Mimosa Project
- Open Access
- 2026
- OriginalPaper
- Buchkapitel
Erschienen in:
Aktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by (Link öffnet in neuem Fenster)
Abstract
1 Introduction
Monitoring traffic flows in urban areas is often an expansive and complicated activity for a public administration. Nevertheless, these data are fundamental for improving the urban and transport planning processes at local and regional levels. In the Italy-Croatia Interreg Mimosa project, an innovative open-source traffic flows monitoring system based on Artificial Intelligence (AI) was created and tested. The collected results showed that the system is able to correctly recognize different typologies of vehicles (cars, bicycles, motorbikes, persons, light freight vehicles and heavy freight vehicles) using only open-source libraries.
The output of the project included the design of an AI tool based on computer vision and it was designed in order to be easily replicable in other contexts.
Anzeige
The pilot action has been designed to reply to the need to simplify and automize the traffic flows data collection for local administrations, that usually have limited budgets for traffic monitoring activities. Traffic flows data collection often requires relatively high-cost technologies (traffic radars, dedicated sensors, etc.) or a large staff involved in monitoring tasks spending time to collect the required data from the streets. Compared for example to the inductive loop detectors technology, the Mimosa AI solution has the advantage of not requiring infrastructure works on the roads. Moreover, AI allows the detection of bicycles and pedestrians more accurately compared to inductive loop detectors. Finally, thanks to AI it is possible to reduce the costs and the human resources needed to collect data on traffic flows.
The paper focuses on a new AI tool, that uses open-source algorithms and models to monitor vehicular flow. The pilot test conducted in Bologna (Italy) by the Institute for Transport and Logistics (ITL Foundation) in collaboration with “GoatAI SRL” and Bologna Airport, as part of the Interreg Italy-Croatia Mimosa project is documented.
A beta version of this AI tool for traffic monitoring is available on an online repository on GitHub.1
2 Artificial Intelligence Open-Source Approach
In recent years, AI has emerged as a powerful tool for analyzing vehicular flows, as it can process and analyze large amounts of data and images quickly and accurately [1]. The open-source tool developed in the Mimosa project uses AI system to analyze vehicular flows on a specific road. The key goals of the tool are:
1.
Accurately count the number of vehicles in a flow.
2.
Accurately classify the types of vehicles that go through an identified road.
3.
Extract and analyze aggregate statistics related to vehicular flow.
Anzeige
Two key open-source AI key tools were adopted for the development of the AI tool for traffic monitoring:
-
YOLOx for the Object Detection and Classification Models activities.
-
Pascal VOC and COCO as datasets for training and testing object detection and classification models.
For the specific case of the Mimosa project, detection and classification were limited to a small subset of objects and sporadic errors do not significantly affect the final aggregate statistics derived from the analysis of several hours of footage. For these reasons, after careful consideration, we have determined that it would be more appropriate to adopt a one-stage detector as the first step for the vehicular flow analysis system. In particular, it is chosen to take advantage of one of the latest and best-performing variants of the famous open-source YOLO model, namely YOLOx [2]. YOLO (You Only Look Once) is one of the most famous open-source and real-time object detection and classification systems. Because of the state-of-the-art performance in terms of both accuracy and processing speed, YOLOx has been selected by the Mimosa team as most suitable option.
In relation to the training and testing object detection and classification models, two free datasets are used: Pascal VOC and COCO. Pascal VOC (Visual Object Classes) is a benchmark on which several tools are tested [3]. Pascal VOC was surpassed by what now is probably the most important dataset when it comes to object detection and classification, i.e. Microsoft Common Objects in COntext (COCO). This dataset considers 80 different object categories, including animals, people and a variety of vehicle types. It has been used in many research papers and has inspired the development of various object detection algorithms. For these reasons, and because it contains almost all the classes of interest for the project, COCO dataset was selected for training the model used in the Mimosa project.
3 Mimosa AI Tool Architecture
This section provides the basis for understanding the artificial intelligence (AI) informatic architecture adopted for the implementation of the open-source Mimosa vehicular flows analysis system. The vehicular flows analysis system developed in the Mimosa project consists of three main modules:
-
The video stream processing module.
-
The short-term tracking module.
-
The flow statistics generation module.
3.1 Video Stream Processing Module
The video processing module is a system designed to allow the offline analysis of one or more video streams. This analysis was conducted by utilizing YOLOx object detection model. To begin the process, each frame of the video stream is normalized and transferred to the graphics processing unit (GPU) for further processing. Once on the GPU, the frame is passed as input to the YOLOx model. The model analyzes the input image for all object categories covered by the COCO Dataset (on which the Mimosa model was trained). Note that the number of classes covered by the dataset (80 classes) is much larger than those of interest to the Mimosa project, which was limited to the following 7 classes: (I) light commercial vehicles, (II) heavy commercial vehicles, (III) passenger cars, (IV) motorcycles, (V) buses, (VI) bicycles, and (VII) pedestrians.
3.2 Short-Term Tracking Module
The short-term tracing module is a crucial component of the Mimosa AI system that enables us to analyze the movement and flow of vehicular traffic. It operates by taking the detection output from the video processing module, which independently analyzes each frame of the video stream, and linking them together over time to create time traces. This is necessary in order to extract meaningful statistics about vehicular flows and understand the patterns and trends of traffic within our system. The Mimosa project team decided to use the SORT [4] (Simple Online and Real-time Tracking) open-source tracking algorithm. It is designed to be used in scenarios where processing speed plays an important role and where accuracy in the short term is more important than overall consistency over the entire video.
3.3 Statistics Generation Module
Having processed the entire video through the two modules previously described, the Mimosa system will finally proceed to the analysis of vehicular and pedestrian flows taking into consideration the classification of the detected vehicles, as well as their coordinates and trajectories. In particular, the system has been created to detect:
-
Count.
-
Direction of travel.
-
Classification of vehicular type.
-
Frequency of passage.
-
Average speed.
Using the statistical generation module, each of the above statistics can be aggregated by vehicle/pedestrian class, and/or motion direction, and/or road lane depending on the user's input query.
4 Testing Activities at the Bologna Airport and AI Tool Quality Check
The pilot test has been conducted within the premises of Bologna Airport. Bologna airport was selected as Mimosa testing area as it has an only access point for both freight and passengers’ flows. This allows us to monitor all the traffic flows by using a limited number of cameras. The Mimosa project testing activities have been carried out using 3 GoPro Hero 10 cameras for video collection. Each recording session has been supervised by operators in order have a human check on the number and typologies of vehicles passing during the monitoring sessions. The position and inclination of the camera have been carefully made to maximize lane visibility. A vehicle traveling on the monitored lane should be visible at least 30 m from the camera location.
During the Mimosa pilot test, two data acquisition sessions were conducted in accordance with the Emilia-Romagna Region and the Bologna Airport key stakeholders. The first recording session took place on 03/08/2022. A total of 80 min of registration from two cams has been collected. A total of 1,557 vehicles and persons were monitored and classified. The second recording session took place on 01/11/2022. A total of 120 min of registration from three cams has been collected. A total of 2,581 vehicles and persons were monitored and classified.
The Mimosa AI solution reliability was tested empirically by selecting some frames of the registered video and checking if the number and typology of the detected vehicles correspond to the video frame and to the human counting activities conducted by ITL Foundation staff. This human check conducted by Mimosa project staff allowed to detect a complete correspondence among the vehicles detected by humans and by the developed AI solution.
5 Problems and Potential Solutions of the MIMOSA AI Tool
The main legal and organizational problems in the AI application on traffic flow analysis were related to:
-
Recorded videos can be considered sensitive under the EU's data protection law.
-
Short recording period due to action camera memory size and battery life.
-
Bad camera positioning due to bureaucratic problems in installing the cameras in the right place.
The main technological problems in the utilization of an open-source AI technology for traffic analysis is:
-
A limited number of vehicle categories available, in particular in relation to freight logistic flows;
-
Requires dedicated and expensive hardware.
Potential technological and organizational solutions are:
-
Privacy can be handled by displaying dedicated video surveillance warning signs that report GDPR General Data Protection Regulation. The warning signs must be positioned in a way as to be clearly visible before the interested party can enter the area.
-
The short recording period could be solved using action cameras with increased battery life and by expanding the memory with SD cards.
-
Limitations about the open-source nature of the tool can be solved by purchasing ad hoc software from a consultancy software house with AI and Computer Vision knowledge and/or by training the open-source tools in order to make the recognition of all the interested vehicular categories.
6 Conclusion and Recommendations
During this pilot action, Fondazione ITL in collaboration with the external technical experts GoatAI SRL, developed a tool for automatic traffic flow estimation based on Artificial Intelligence and Computer Vision able to (i) count the number of vehicles and people, (ii) classify the types of vehicles and (iii) extract aggregate statistics. This project aimed to provide decision-makers with a new tool for traffic monitoring and data-oriented decision-making on the topic of sustainable transport promotion. This information can be used, for instance, to optimize traffic signals, improve road design, and reduce congestion.
The proposed software has been tested at the Bologna Airport on two separate days. During the tests, a total of 02 h:10 m was recorded from three different cameras. The collected videos have been elaborated by the proposed tool. Videos have been promptly deleted after computation to be compliant with privacy regulations.
The data extracted during experimentation is intended to provide insights into urban mobility planning in a cost-efficient way. The solution has been designed to be replicable in multiple contexts where a camera can be easily placed in an elevated position above traffic flows. The software has been open sourced under MIT license and can be downloaded on GitHub.
In order to replicate the tests in other contexts, it is necessary to formulate a series of recommendations:
-
Cameras should be positioned to face the incoming traffic (monitored vehicles should move towards the camera).
-
A vehicle traveling on the monitored lane should be visible at least 30 m from the camera location.
-
The camera should be placed at least 3 m above the ground. The defined position should not exceed 6 m in height.
-
For long recordings that should last for more than 3 h, we advise employing preinstalled cameras, as the battery life of action cameras is usually in the range of 100–200 min.
-
In the video collection activities, it is better to use filters allowing to cover vehicles license plates. This could be an important solution for reducing the needed legal checks.
The AI solution reliability has been measured on the most challenging reference benchmark namely COCO dataset. The good performance obtained by the AI solution shows good generalization capability and demonstrates that the system can be employed in different contexts.
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
