Multiscale feature enhancement algorithm for target detection in traffic scenarios

To address and resolve the issue of visual detection errors due to the different scale of the target and more blockage in complex traffic environments, a multi-scale enhanced traffic scenario target detection algorithm based on YOLOv5 has been proposed. A multi-branch stacking module MSB is designed to replace the C3 module in the backbone network to enhance the feature extraction capability of the backbone network. Based on the structural design of the CSP, the spatial pooling pyramid module is reconstructed, and the SPPFCSP module is designed to replace the SPPF module to obtain more feature information while ensuring the same receptive field. An additional small-scale target detection layer is constructed by fusing the shallow information of the backbone network and the neck network information. A network with shallow information improves the detection capability of small targets. The experimental results show that the improved algorithm`s mAP@0.5 and mAP@0.75 are 91.5% and 61.5% in the KITTI dataset, respectively, which are improved by 3.8% and 6.2% compared with YOLOv5s. The improved algorithm is executed on NVIDIA Jetson AGX Xavier with a real-time detection rate of more than 30 fps, which meets the speed and accuracy requirements of detection algorithms for in-vehicle scenarios.