Pedestrian Head Detection And Counting Based On Improved YOLOv5s Algorithm

At present,target detection techniques are widely used,and good detection results are obtained on general data sets.However,pedestrian head detection is characterized by low resolution,difficulty in feature extraction and insufficient semantic information due to its small size and small proportion in image,at the same time,there are pedestrian congestion,occlusion and other problems,resulting in unsatisfactory detection results,lower accuracy and recall rate.At the same time,the missed detection rate is high.In view of the above-mentioned problems,In order to improve the accuracy of pedestrian head detection,recall rate,and reduce the rate of missed detection,an improved algorithm based on YOLOv5 s is proposed to detect and count human head on SCUT＿HEAD data set,the specific improvement includes two aspects:1.Using K-Means++ algorithm to generate clustering which is more suitable for head detection of SCUT＿HEAD data set;Mosaic-9 is used to enhance the data,enrich the small target sample data set,add a detection layer,improve the structure by fusing modules,improve the capture of semantic information,add SE + CBAM merge attention mechanism,the guidance model pays more attention to the spatial and channel features of head information,improves the performance of small target detection,improves the activation function in SPPF module,and optimizes the loss function by using EIo U,realizes the improvement of YOLOv5s-LJ Algorithm,to Improve the precision and recall of pedestrian head detection.2.On the basis of Yolov5s-LJ Algorithm,Transformer Prediction Heads(TPH)are used to replace the shallow Prediction Heads to improve the performance of the algorithm.Using the local instead of the whole method,by detecting the number of head count.The experimental results show that the T-bh-4 + YOLOv5S-LJ algorithm is effective in pedestrian head counting.The missing detection rate is 6.9%,which is 10.1% lower than the original model At the same time,the validity of YOLOv5S-LJ and T-bh-4 + improved structure +CBAM algorithm in pedestrian head detection is also verified.MAP of YOLOv5s-LJ and T-bh-4 +improved structure + CBAM is 90.2% and 90.5% respectively,compared with the original algorithm,it is improved by 3.6% and 3.9% respectively.