Improvement of YOLOv5 lightweight combined with OTA pedestrian detection methods
The pedestrian detection method based on YOLOv5 has achieved good results in practical applications,but there are still some shortcomings,such as the detection in complex scenes is not accurate enough,and the processing speed for target de-tection is not ideal.In order to solve the above problems,the following improvement method is proposed:replace the C3 module with the RepGhost module to reduce the computational complexity of the model and improve the detection speed.In addition,the pedes-trian detection accuracy is further optimized by combining with the Optimal Transport Assignment(OTA)method,which optimizes the allocation of detection frames by minimizing the cost function.As a result of these improvements,the amount of parameters in the model is reduced by 35%,the amount of computational GFLOPs is reduced by 20%,and the processing speed is increased by 7%on the RTX3090,while the model size is reduced by 33%.Accuracy is improved by 2.7%on mAP50 and 1%on mAP50-95.
deep learningYOLOv5target detectionmodel improvementlightweighting
万方数据
