Pedestrian-vehicle detection model in road scenes based on improved YOLOv5

The foundation of the autopilot job is the accurate identification of targets in a street scene; however, in complicated street scenes, the detection model performs poorly because of issues with complex backgrounds and occlusion between densely populated targets. This research therefore suggests a target recognition technique based on YOLOv5 for complicated street scene roadways. To enhance the feature extraction of the identified targets, several fusion methods of the SE attention mechanism in the underlying network are initially explored; next, the spatial feature pyramid pooling module is enhanced by using a dense connection approach to address the issue of information loss brought on by maximal pooling; and last, the network's detection performance is enhanced by using decoupled detection headers. To verify the effectiveness of the method, KITTI and Udacity evaluated the method in this paper. The experimental data proves that the method of this paper meets the real-time requirement while improving the accuracy.