Automated detection of helmet on motorcyclists from traffic surveillance videos: a comparative analysis using hand-crafted features and CNN

The higher mortality rate in motorcycle accidents is attributed to negligence in wearing a helmet by two-wheeler riders. Identification of helmetless riders in real-time is an essential task to prevent the occurrence of such events. This paper presents an automated system to identify motorcyclists without a helmet from traffic surveillance videos in real-time. The problem becomes more challenging when computational resources are limited. We have compiled a custom dataset for developing an automated helmet detection algorithm. The proposed system uses a two-stage classifier to extract motorcycles from surveillance videos. Detected motorcycles are further fed to a helmet identification stage. We present two algorithms for classifying riders with and without a helmet, one based on hand-crafted features and the other based on deep convolutional neural network (CNN). Our experiments show that the proposed CNN model gives the best performance in terms of accuracy while the feature-based model gives faster detection. Most importantly, to ensure the light-weightiness of the proposed system all the computations are performed in CPUs only.