INTEGRATED TRAFFIC VIOLATION TYPE DETECTION AND RECOGNITION SYSTEM USING VIDEO PROCESSING BASED CONVOLUTIONAL NEURAL NETWORK

Based on data from the World Health Organization in 2018, the number of deaths worldwide due to road traffic accidents is 1.35 million people every year. One of the causes is the low level of driver discipline in driving, which is indicated by violating traffic regulations. The solution for this problem is implementing a traffic violation detection system based on computer vision and deep learning. The system designed in this study can detect traffic violations that include running a red light, not wearing a helmet, and being wrong-way. This study implements the YOLOv5 architecture as object detection has 74% mAP50 value performance and uses SORT as object tracking. Vehicle detectors and trackers are then integrated with methods designed to detect traffic violations. The test results using several sample video scenarios show that the running red light detector has an F1-Score value of 0.95. The helmet violation detector based on EfficientNet as a classifier has an F1-Score value of 0.88, and the wrong-way detector has an F1-Score value of 1.00. ICIC International