Autonomous Reckless Driving Detection Using Deep Learning on Embedded GPUs

Reckless driving is dangerous, and must be monitored, detected, and law-enforced to assure road safety. For this purpose, this work presents an embedded system for monitoring and detecting reckless driving activities on the road autonomously in real-time. Using an embedded GPU (eGPU) platform, a camera, and a combination of light-weight deep learning models, we design a system that can identify abnormal vehicle motions on the road. Our system analyzes discrete per-frame images from vehicle detection algorithms, and creates a continuous trace of a vehicle's motion trajectory. While doing so, a virtual grid is generated on the road to obtain positions of vehicles with less overhead and accurately track a vehicle's movement even with low frame rate (5 fps) videos. Vehicle's motion trajectory is then compared against the surrounding to identify abnormal behavior through driving activity classification, which can be provided to law enforcement personnel for final validation. The key challenge is the fundamental resource constraints of embedded platforms, and we design algorithms to overcome their limitations. Evaluation results show that our scheme can well-extract the horizontal and vertical movements of a vehicle (100% recall and 67% precision ) and show the potential for truly autonomous reckless driving activity detection systems.