Improving traffic management systems by building a framework that is capable of automatically detecting illegal offences in the streets and being able to identify the cars responsible for such offences would be a great feat. In this paper, I design and experiment with the first two phases of the framework that are responsible for vehicle detection, traffic line detection and license plate detection.Those models were built using deep learning approaches using the YOLOv5 pre-trained models and extending those models with extra training using two new data sets. The first data set specifically designed for vehicle detection and traffic line detection while the other data set is customized for the License plate recognition part of the system. I was capable of designing both models and reaching around 0.8 mAP in the vehicle and traffic line detection, and almost 1 mAP in the license plate recognition.
In this project, I used two different data sets to build two different models. The first data set was obtained for detecting vehicles and identifying their types in addition to identifying traffic lines in the street and this data set had around 340 images, while the other UFPR-ALPR data set was obtained from [1] and is only released for academic research purposes and has around 4000 images that were annotated. I tried combining both of the datasets in order to build a more generalized detector but it was not very effective as they had different class annotations; however, I do believe finding a bigger combined dataset would definitely improve the performance of these models. I did some data preprocessing by adjusting the annotations to suit the YOLO model and discard any addition unneeded information.
Future work for this project, would be to combine both detectors using a common much bigger dataset to improve the speed and accuracy of detections, build a framework for detecting the offences based on the output of the detectors and finally extend the LPR to be capable of extracting characters and identifying the license plate numbers automatically.
The specs of the laptop to train both models: Intel i7-8750H CPU @ 2.2 GHZ, 16 GB RAM and NVIDIA GTX 1060. The image size used was 620, the batch number of images were 6. The first experiment was to test the vehicle detection and traffic line detection using the first dataset of 340 images using both YOLOv3 and YOLOv5 and comparing results.
Experiments Precision/Recall mAP@0.
50 Epoch YOLOv3l 0.8844/0.7712 0.
100 Epoch YOLOv5l 0.9411/0.746 0.
The data set for this part of the system was much bigger and thus was much slower when it came to training the model. I was only capable of running the model for 25 epochs over the provided 4000+ images and came up with very satisfying results in very short time using the YOLOv5l model as well.
Experiments Precision/Recall mAP@0.
YOLOv5l 0.963/0.9528 0.
ALP R[1] 1.0/1.0 N/A
References
[1] L. A. Zanlorensi L. S. Oliveira G. R. Gonçalves W. R. Schwartz R. Laroca, E. Severo and D. Menotti. A robust real-time automatic license plate recognition based on the yolo detector. International Joint Conference on Neural Networks (IJCNN), July 2018. [2] A.; Borovec J.; Changyu L.; Hogan A.; Diaconu L.; Ingham F.; Poznanski J.; Fang J.; Yu L.; et al. Jocher, G.; Stoken. Yolov5. 2021. [3] Y.; Li L.; Wang Z.; Zhou Y. Lian, J.; Yin. Small object detection in traffic scenes based on attention feature fusion. 2021. [4] Ruiz-del-Solar Javier Verschae Rodrigo. Object detection: Current and future directions. Frontiers in Robotics and AI. [5] https://github.com/ultralytics/yolov5