Traffic Rules Violation Detection using Deep Learning
Aniruddha Tonge, Shashank Chandak, Renuka Khiste, Usman Khan, Laxmi Bewoor
Abstract
In order to ensure safety measures on roads of India, the identification of traffic rule violators is highly desirable but challenging job due to numerous difficulties such as occlusion, illumination, etc. In this paper we propose an end to end framework for detection of violations, notifying violators, and also storing them for analyzing and generating statistics for better decision making regarding traffic rules policy. In the proposed approach, we first detect vehicles using object detection which is performed using YOLO, and then accordingly each vehicle is checked against appropriate violations viz. not wearing a helmet, violation of crosswalks. Helmet violation is detected using a CNN (Convolutional neural network) based classifier. Crosswalk violation is detected using Instance Segmentation by Mask R-CNN architecture. After violations are detected, vehicle numbers are obtained of respective violators using OCR, and violators are notified. Thus an end to end autonomous system will help enforcing strong regulation of traffic rules.