Extreme scenario understanding based on causal inference for automatic driving
Ruijie Wang, Lei Tang, Rao Feng, Junchi Ma · 2023
Correctly understanding and identifying rare and extreme driving scenarios is crucial for the safe and reliable operation of autonomous vehicles. Multi-sensor collaborative perception and high-precision visual equipment not only require high configuration and strong computational support, but also pose deployment challenges, which will inevitably increase the cost of autonomous driving systems. Therefore, given the constraints of limited visual equipment performance and computing power, how to enable the system to identify and understand different driving scenarios and make safe decisions is a key issue in developing low-cost lightweight autonomous driving systems and achieving widespread adoption of autonomous vehicles. We proposes a framework for Autonomous Vehicles Understanding Extreme scenario based on Causal Inference(CI-AVUES). Under resource constraints, this method constructs a causal model that conforms to human driver cognitive attention and reasoning mechanisms, in order to infer the behavior and relationships of objects in autonomous driving scenarios and achieve recognition of different extreme driving scenarios. The CI-AVUES framework was validated using the open-source CODA dataset for recognizing extreme driving scenarios, with an AUC value of 0.973, a precision value of 1.0, an accuracy value of 0.95, and an f1 score of 0.909, demonstrating the ability of CI-AVUES to recognize the extreme driving scenarios.