Transfer learning based SSD model for helmet and multiple rider detection

Deep learning advancements have enabled the development of practical applications for smart surveillance. In order to improve road safety, a systematic effort is required to identify and punish violators. The present study is one of several deep surveillance applications used for traffic rule violation monitoring of helmetless motorcyclists and multiple pillion riders on motorbikes. Transfer learning is used to extract the multiscale features, and a single-shot multi-box detector (SSD) has been employed to handle crowded traffic scenarios. The proposed motorcycle violation detection (MVD) model combines aspect ratio aware training in the subsequent fine-tuning stage to improve the detection performance. The proposed MVD model is evaluated on the real-world MVD dataset, collected from surveillance cameras, with busy and sparse roads under different views and weather conditions. The proposed method achieves a mean average precision of 78.1% with four classes and a detection speed of 57 frames per second in the challenging image and video dataset. Experimental results show that the MVD model outperforms the state-of-the-art approaches in terms of speed and accuracy. As a result, the proposed system can be used in complex traffic scenarios with an existing traffic surveillance infrastructure.