Reinforced learning based algorithm: reducing accidents and increasing road safety

Abstract

Safety on the roads is of utmost importance for both autonomous and humandriven vehicles. State-of-the-art accident anticipation methods, essentially based on supervised learning, show promise but are bounded by their dependence on large labeled datasets and their inability to generalize straightforwardly to new driving environments. This research extends the existing framework of DRIVE, a deep reinforcement learning model that predicts accidents from dashcam videos by mimicking human visual attention. DRIVE proposed a new approach that involves dynamically learned adaptive policies through integrated visual attention and accident prediction. Through this paper, we will be integrating the DRIVE framework tightly with TD3, refining the reward mechanisms of DRIVE in the process; mainly, the dense anticipation rewards and sparse fixation rewards. We would like to explore how such enhancement can further result in early and accurate accident prediction together with robust visual explanations for its decisions while making the computation less hardware intensive. Preliminary insights could also provide the fact that an optimized DRIVE might bring a sea of change in the accuracy and timeliness of accident predictions that will go a long way in ensuring much safer and more reliable autonomous driving systems. This work underlines the imperative of continuous innovation in advanced technologies to check reckless driving and improve road safety globally.