Improved SAF-FCOS detection algorithm for obstacles with radar and vision fusion

With the rapid development of automated driving technology, it becomes crucial to accurately detect and recognize targets in complex scenes. Camera sensor detection and recognition accuracy is not high enough, has poor stability, and cannot adapt to the target detection of complex scenes. Millimeter wave radar is less affected by the environment and more accurate in speed and distance measurement, but the current mainstream millimeter wave radar sensor point cloud is sparse, so it cannot identify the target feature information. Therefore, the fusion of radar and vision is gradually becoming a mainstream solution for accurate obstacle detection. This paper presents a novel parallel attention fusion module (PAFM) to enhance the performance of the SAF-FCOS target detection network. By incorporating depth-separable convolution and parallel dual attention modules, the RetinaNet network of SAF-FCOS is enhanced to effectively integrate information from upper and lower feature maps of diverse scales. Compared to the direct summation approach, PAFM enables the network to selectively focus on image regions of interest, both in terms of channel and spatial dimensions, without introducing excessive algorithmic complexity. Experimental results on the NuScenes dataset demonstrate that the proposed improved network achieves superior performance compared to the original network.