Abstract: A network model for radar and camera feature fusion was proposed to deal with the impact of low light, rain and fog and other harsh scenes on the detection capability of intelligent driving vision systems. A radar attention mechanism feature module was constructed based on millimeter wave radar information and attention model to provide a priori information and increase the weight of the algorithm in the target candidate region for the feature fusion network. The test results show that, introducing the radar attention mechanism module, the target detection performance of the feature fusion network is significantly better than the detection performance of that relying on computer vision alone, and the target detection is more robust in complex scenes.