Research On Dense Pedestrian Detection Method Based On Improved YOLOv5

With the rapid development of computer vision technology,pedestrian detection,as one of the important research contents in this field,not only has high academic value,but also attracts more attention due to its contribution to the emergence of applications such as safety monitoring,assisted driving,and intelligent robots.However,in the dense pedestrian scene,due to factors such as large differences in pedestrian scales and severe mutual occlusion,the existing pedestrian detection methods have the problems of low accuracy and high missed rates.Therefore,based on YOLOv5,this thesis studies the dense pedestrian detection method,which involves three aspects: network improvement,loss function optimization and post-processing algorithm design.The main research work is as follows:Aiming at the problem of large differences in pedestrian scales in dense scenes,which leads to a high rate of missed detection of small-scale pedestrians,a dense multi-scale pedestrian detection method DP-YOLOv5 is proposed in this thesis.In order to improve the ability of the backbone network to extract pedestrian features at different scales and enhance its attention to small-scale pedestrians,this method first uses the idea of reparameterization and replaces the 3×3 convolutional module with an improved Rep VGG module,which improves detection accuracy with a small increase in test time overhead.Secondly,the high Efficiency Attention ECA was added to the C3 module to improve the weight distribution of feature extraction for small-scale pedestrians.Finally,in order to enhance the utilization rate of the model for different scale features,a weighted cross-layer Path aggregation network is proposed.By linearly weighting different scale features,the fused features can have both shallow position information and deep semantic information.Aiming at the problem of internal occlusion of pedestrians in dense scenes,which leads to poor positioning accuracy of occluded pedestrians and the bounding boxes are easily suppressed in the post-processing stage,this thesis optimizes and designs the loss function and post-processing algorithm respectively.Firstly,the regression loss of the bounding box loss function was improved and an exclusion loss term was added.The optimized loss function can not only improve the accuracy of predicted bounding boxes in size regression and localization regression,but also reduce the interference caused by nearby ground truth boxes and predicted boxes.Secondly,considering the high density of pedestrians in dense scenes,highly overlapping prediction boxs may be violently deleted by the non-maximum suppression algorithm.In this thesis,a two-threshold post-processing algorithm is designed.The distance intersection over union is used as the evaluation criterion of target similarity,and the boundary box confidence is processed by two thresholds.The experimental results verify the effectiveness of the improved loss function and the post-processing algorithm in the dense occlusion scene.Finally,combined with DP-YOLOv5,the final dense pedestrian detection method in this thesis is obtained.Finally,a series of experimental verifications are carried out on the City Persons and Crowd Human datasets.The experimental results show that,compared with the original YOLOv5,the method proposed in this thesis can effectively solve the above problems.