A hierarchical perception decision-making framework for autonomous driving

Self-driving vehicles have attracted significant attention from both industry and academy. Despite the intensive research efforts on the perception model of environment-awareness, it is still challenging to attain accurate decision-making under real-world driving scenarios. Today’s state-of-the-art solutions typically hinge on end-to-end DNN-based perception-control models, which provide a rather direct way of driving decision-making. However, DNN models may fail in dealing with complex driving scenarios that require relational reasoning. This paper proposes a hierarchical perception decision-making framework for autonomous driving by employing hypergraph-based reasoning, which enables fuse multi-perceptual models to integrate multimodal environmental information. The proposed framework utilises the high-order correlations behind driving behaviours, and thus allows better relational reasoning and generalisation to achieve more precise driving decisions. Our work outperforms state-of-the-art results on Udacity, Berkeley DeepDrive Video and DBNet data sets. The proposed techniques can be used to construct a unified driving decision-making framework for modular integration of autonomous driving systems.