Multi-sensor Fusion Method for Lane and Target Detection Based on Point Cloud Feature Sequence Coding and Cross-attention Mechanism

Integrating the visual camera images,millimeter-wave radar point cloud data and prior navigation maps to realize lane line detection and dynamic target detection in complex scenes is one of the technical challenges faced by current automatic driving environment perception.To solve the problem,a multi-sensor fusion detection framework based on deep learning is proposed.To take the radar point cloud as the query object,the coding mode of point cloud feature sequence and the cross-attention mechanism module are designed.The visual image is used to generate attention weights,and prior navigation map information is fused at the feature level,which can improve the lane detection performance of radar point cloud data and visual image data fusion effectively.The OpenLanev2 and nuScenes public data sets are used to test the proposed method,the results show that the method not only achieves the best lane detection performance,but also has outstanding performance in dynamic target detection.