Research On Pedestrian Detection Algorithm Based On Residual-Dense Connection And Attention Mechanism

As one of the hot research contents in the field of computer vision,pedestrian detection technology has made major breakthroughs in recent years.However,the pedestrian images obtained from real life are often small in scale,diversified in scale,and have different degrees of occlusion.These factors are bound to affect the detection effect.Therefore,based on the improvement of the existing pedestrian detection model,this thesis proposes two pedestrian detection algorithms,which are as follows:First,in view of the low detection accuracy of small-scale pe destrians and the problems such as missing detection and false detection,a Faster RCNN pedestrian detection algorithm based on residual and dense connection network was proposed.Firstly,by taking advantage of the residual connection and the dense connection network,the residual and dense connection network is used as the backbone network to obtain more key and effective feature information,so as to complete the feature extraction task.Secondly,a new anchor scales generator algorithm is designed to generate prior anchors size with a higher degree of fitting with the experimental data set,so as to improve the positioning effect of small-scale pedestrians.Finally,a multi-scale feature fusion module is formed by combining the last three convolutional blocks of the residual and dense connection network with the feature pyramid network structure to obtain more detailed information of pedestrians.Experimental results show that this method can effectively improve the detection accuracy of small-scale pedestrians.Secondly,aiming at the problem of insufficient feature acquisition and low detection accuracy caused by occlusion in pedestrian detection in dense scenes,a YOLOv4 pedestrian detection algorithm based on attention mechanism was proposed.Firstly,a location attention module is introduced before the feature fusion in the feature enhancement part of YOLOv4.This module could not only expand the receptive field,obtain more feature information of the blocked pedestrians,but also reduce the feature loss during the feature fusion to achieve feature enhancement.Secondly,the spatial pooling pyramid structure of YOLOv4 is replaced by a spatial information enhancement module,which reduces feature loss and cross-layer connection through its internal multi-branch cavity convolution operation to promote the full integration of local details and global semantic features,and enhance the spatial feature characterization ability of the network.Finally,a context attention module is introduced after 5 continuous convolution of the path aggregation network,which can fully extract the multi-scale spatial information in the channel attention vector and effectively integrate multi-scale context features.Experimental results show that the proposed method can improve the detection accuracy of pedestrians with shielding.Third,based on the above two pedestrian detection models,a pedestrian detection system was developed using Pycharm and PyQt5,which could accurately detect pedestrians in the input images.