Authors: T. Sekar, A. Sangeetha
Abstract: The increased number of road accidents associated with violating two-wheeler helmet usage is very alarming and this situation demands the introduction of smart surveillance systems to maintain safety of the people. In this paper, we present the idea of a new system of helmet detection using deep learning algorithms and image processing to detect whether a person is not wearing a helmet automatically or not. The publicly available Kaggle Helmet Detection dataset that includes 7,500 images having annotations of bounding-boxes of helmet head, and person is used by the system. We transform the annotations to a binary classification task – Helmet and No Helmet and use the YOLOv5 object detection model because of its speed and accuracy of the inference. This was done by training the model using transfer learning and optimizing the model with data augmentation techniques to achieve cross generalization under different kinds of light and environmental conditions. Our system is tenable based on the results of experiments because it took into consideration a real-world scenario. The model based on the YOLOv5 had a generally high accuracy of 95.64%, precision of 94.32%, recall of 91.23% and an F1-score of 92.75. Real-time inference can also be done with the system as it can perform 24.56 ms/frame. This will fit it to be used in surveillance systems in a city environment. Also, Deep SORT tracking has been integrated to provide effective tracking without redundancy. This project will be useful in the development of intelligent traffic systems to automate the process of identifying non-helmets with high precision making it useful to law enforcement and citizen and driver safety on the road. It may be extended in the future to have modules of number plate recognition and fine imposing modules to be able to implement all the traffic rules.
