Traffic Participants Detection and Classification Using YOLO Neural Network

Abstract

One of the most important requirements for the next generation of traffic monitoring systems, autonomous driving technology and Advanced Driving Assistance Systems (ADAS) is the detection and classification of traffic participants. Although in the areas of object detection and classification research, tremendous progress has been made, we focused on a specific task of detecting and classifying traffic participants from traffic scenarios. In our work, we have chosen a Deep Convolutional Neural Networks – YOLOv4 (You Only Look Once Version 4), a object detection algorithm to detect and classify traffic participants accurately with fast speed. The main contribution of our work included: firstly, we generate a custom image dataset of traffic participants (Car, Bus, Truck, Pedestrian, Traffic light, Traffic sign, Vehicle registration plate, Motorcycle, Ambulance, Bicycle wheel). After that, we run K-means clustering on the dataset to design an anchor box that is utilized to adapt to various small and medium scale. Finally, we train the network for the mentioned objects and test it in several driving conditions (including daylight, low light, high traffic, foggy, rainy environment). The results showed cutting-edge performance with a mean Average Precision (mAP) of up to 65.95% and a speed of about 54 ms.