Environmental anomaly detection and intelligent response for autonomous driving based on deep neural networks

To address the real-time and reliability challenges of environmental anomaly detection and response in autonomous driving, this paper proposes an end-to-end solution based on a lightweight graph neural network with multi-modal temporal attention (LGNN-MTA). This framework uses a GNN (Graph Neural Network) to model the interactions between traffic participants, incorporates a channel-spatial attention module to extract anomaly features from multimodal sensor data, and introduces a temporal memory unit to capture the evolution of potential risks. To achieve an intelligent response, this study designs an improved Proximal Policy Optimization (PPO) reinforcement learning algorithm to generate safe obstacle-avoidance strategies. Experiments on the AnoVox simulation dataset show that the proposed model achieves an anomaly detection F1 score of 89.7% while maintaining an inference speed of 25 FPS (Frames Per Second), a 6.2% point improvement over YOLOv7-OBB (You Only Look Once with Oriented Bounding Box). The response decision success rate increases to 93.5%. Compared to traditional MPC (Model Predictive Control) solutions, the false trigger rate for emergency braking is reduced by 31%, significantly improving system safety and driving comfort. This research effectively constructs an integrated “perception-decision-making” intelligent response system that provides solid technical support for the safe deployment of high-level autonomous driving systems.