Research on vehicle safety based on multi-sensor feature fusion for autonomous driving task

How to enable autonomous vehicles to sustain high safety and rapid response capability in complex, dynamic environments remains a critical and urgent challenge. This study aims to elevate the learning efficiency of multi-sensor feature fusion in autonomous-driving tasks, thereby strengthening system safety and responsiveness. To this end, we present an innovative multi-modal, multi-sensor feature-fusion model that integrates two highly effective mechanisms (Sparse Channel Pooling and Residual Triplet-Attention) to refine the fusion process. First, the Sparse Channel Pooling mechanism allows the model to adaptively select salient feature channels while suppressing redundant information. Second, the Residual Triplet-Attention mechanism resolves inter-modal feature misalignment, enabling accurate cross-modal alignment of key features and markedly improving computational efficiency. Finally, a reinforcement-learning module is introduced to learn and optimize policies in continuous action spaces. Coupled with the feature-fusion framework, this module enables end-to-end, high-efficiency driving-policy learning within the CARLA autonomous-driving simulator. Experimental results demonstrate that the proposed method not only maintains safe operation in complex dynamic scenarios but also preserves real-time, high-level perception and decision-making accuracy, offering a novel paradigm for multi-sensor data fusion in autonomous driving.