Research On Pedestrian Target Detection Method Based On Full Convolutio

In recent years,pedestrian detection is one of the hot research fields of object detection,which mainly uses computer vision technology to detect whether there are pedestrians in images or video sequences and mark the location of each pedestrian.Pedestrian detection has important research value and broad application prospects in intelligent security monitoring,autonomous driving,human-computer interaction,etc.Most of the existing pedestrian detection methods use feature pyramids for multi-level prediction,which can detect multi-scale targets,but it is easy to generate many prediction frames far from the center of the object,resulting in low detection accuracy.In the case of dense pedestrians and serious occlusion,the relevant network model cannot obtain enough target features,resulting in false detections and fewer detections.In view of the above problems,this paper mainly carries out the following researches:(1)A pedestrian detection method based on dynamic selection of optimal features is proposed for the feature pyramid that tends to generate a large number of low-quality prediction frames.First,by adding a centrality branch to the detection head network to reduce the weight of low-quality prediction boxes,thereby reducing the number of low-quality prediction boxes.Secondly,the module of coupling Dynamic Selection of Optional Feature(DSOF),by calculating the total loss function of different feature layers,performs feature training on the layer with the minimum value to achieve dynamic selection of the best feature.Furthermore,the labeling method of positive and negative samples is optimized by dividing the sample area,and the accuracy of the detector’s distinction between the front and back backgrounds is improved.Finally,the sufficient experimental results show that the method can suppress the low-quality prediction frame and select the best feature,and improve the accuracy of distinguishing the background before and after the image,and effectively improve the model detection performance.(2)Aiming at the problem of insufficient feature extraction in pedestrian detection occlusion scenes,a pedestrian detection feature enhancement method based on full convolution is proposed based on the pedestrian detection method based on dynamic selection of the best features.First,the optimized regression feature(ORF)module is jointly optimized to obtain more location information of the regression feature by optimizing the location regression branch feature in the network structure.At the same time,the Enhanced Boundary Feature(EBF)method is introduced to directly use the boundary features collected from each boundary to enhance the original point features,so that the feature map has a high response to the extreme points of the target boundary and is affected by background noise.smaller.Finally,the experimental results verify that the method can effectively enhance the feature extraction of pedestrian targets and improve the pedestrian detection ability in occluded scenes.