Computer vision-based identification of motorcycle helmets using a TPU-based platform

The use of helmets can significantly reduce fatal injuries in motorcycle accidents. How ever, there are still many instances where helmets are not used. In this paper, we present a computer-vision system for the automatic identification of helmet use by motorbike riders, that can be used to encourage the use of helmets. The system is developed on a Google Coral Dev Board with an embedded EdgeTPU, connected to a display and a buzzer that are used to provide feedback to the people it detects. Firstly, we present the training of the specific model with several optimizations and then we evaluate two versions of the SSD-MobileNetV2 model in terms of performance, accuracy and energy consumption. The proposed system can achieve high accuracy, low-latency and it consumes less than 4.2 Watts on a TPU board.