Risky Scenario Identification
Most of the accidents are ascribed to risky circumstances such as negligence, unsafe environment, congestion, human error, and others. These risky scenarios can be identified and rectified using real-time deep learning techniques. Identifying risky scenarios during driving in a real-time manner is crucial for Advanced Driver-Assistance System (ADAS) to make active countermeasures and is also an important component of constructing critical scenarios for evaluating the safety performances of autonomous vehicles (AVs). Recent advancements in intelligent learning methodologies have made a significant improvement in the area of computer vision involving real-time assessments. Hence exploring real-time computer vision and deep learning methods in risk perception and identifying risky scenarios would be an excellent and timely research attempt.
The goal of this pre-study is to determine whether perceived risk, measured by both objective metrics (e.g., jerk changes) and subjective evaluation by a large number of human drivers, could be predicted from vehicle trajectory and basic features obtained using computer vision. We aim to identify possible critical scenarios, evaluate the real driving conditions of AVs, and discover potential unsafe scenarios by using deep learning. An automatic system will be built to find out the relationship between the perceived risk score and the objects extracted from traffic scenes to improve the accuracy and efficiency of identifying critically risky scenarios. Finally, the overall framework can be used to improve the evaluation of implemented safety systems performance of ADAS and AVs in different driving circumstances.