Aiming at the current problem of unsatisfactory vehicle detection in complex scenes, an improved vehicle target detection network model is proposed. First, Res2Net residual network is fused in SCP, and the CSP_R structure is proposed, so that the model can extract deeper feature information and strengthen the ability to characterize small-scale targets; the attention mechanism is introduced, and the C3_CBAM module is designed to strengthen the attention to the detection targets while avoiding the increase of the model's computational volume; the loss function of the MPDIoU regression optimization is introduced, and the loss function is optimized by combining the prediction frame with the real frame length, width and area loss, and quantitative indicators to improve the convergence speed and robustness of the model. Finally, the model is validated on the SODA10M dataset, and the experimental results show that the model detection speed reaches 32 frames per second. The average detection accuracy reaches 83.7%, which is an improvement of 7.8 percentage points compared with YOLOV5s.
|