Automatic Helmet Violation Detection of Motorcyclists from Surveillance Videos using Deep Learning Approaches of Computer Vision

Automatic detection of helmet for motorcyclists from real-time surveillance videos is a rising application in computer science. Object detection and classification using deep learning was recently well-known over the years. Researchers used these techniques for solving several surveillance-related problems. Several deep learning models adopt for automatic detection of helmet for motorcyclists but they cannot achieve state-of-the-art results due to different difficulties such as low resolution, whether conditions, occlusion and illumination etc. In this paper, we proposed a methodology for surveillance videos that automatically detect the helmet wear by motorcyclist or not. For this purpose, we used the Faster R-CNN model. First, we apply Region Proposal Network (RPN) starting with the input image that has been delivered into the backbone. Then RPN weights are settled and proposals from RPN are utilized to train the Faster RCNN model. For training, we used the self-generated dataset of three different locations in Lahore, Pakistan. The experimental results detect 97.26% accuracy on real-time surveillance videos for the detection of helmet for motorcyclists.