Research On Pedestrian Detection And Tracking Algorithm Based On Convolutional Neural Network

With the rapid growth of computer vision,pedestrian detection and multiple object tracking technology have piecemeal become a hot research direction for the past few years,and play a crucial role in the domains of security monitoring,intelligent transportation,automatic driving and intelligent robots.Because of the restriction of device computing power and memory,which restricts some network models with large parameters and high complexity real application,particularly some convolutional neural network models that rely on deepening the depth and width of the network to improve feature extraction and fusion abilities.It is painful to arrange to embedded platforms with restricted computing power.So as to settle this problem,this paper proposes a lightweight detection-based pedestrian multiple target tracking algorithm,furthermore,the specific research content is as follows:Directing against the issues of low detection rate,target occlusion,and small size pedestrian missed detection in the detection process,this paper proposes three modifications built on the YOLOv4 algorithm: making use of the lightweight network Mobile Netv3-small without the last convolutional layers and pooling layer as the backbone network,decreases the width and depth of the network,enhances the inference speed of the detection algorithm.The designed SA attention module is used to replace the SE attention module in the lightweight backbone network,the ability of the detection network to extract local features of pedestrian targets is enhanced,and the detection accuracy of occluded targets is improved.The multi-scale output is reconstructed,on the basis of using the adaptive multi-scale feature fusion structure,the detection effect of multi-scale pedestrian targets is improved,and the computational burden is less.Finally,the designed EIo U loss is used as the regression box loss function to train the detection model,and the comparison experiment with the commonly used detection algorithm under the self-built pedestrian dataset verifies that the proposed detection model has least parameters and fastest inference speed while ensuring the detection accuracy.Aiming at the problems of large changes in the motion speed of pedestrian targets,weak extraction ability of pedestrian target appearance features and slow tracking frame rate throughout the pedestrian multiple target tracking,this paper stems from Deep SORT tracking algorithm,by leading acceleration components into the Kalman filter and using the VOVNetv1 person re-identification model,which enhances the position prediction and matching correlation ability of linear moving pedestrian targets,the whole tracking outcome of the algorithm is improved.Ultimately,the performance of the modified Deep SORT tracking algorithm is estimated on the MOT16 dataset,the results are compared with other tracking algorithms,it shows that the proposed algorithm has a competent tracking impression,and it also has a high tracking rate on hardware platforms with low computing power.According to the raised detection model and tracking algorithm,the pedestrian collision counting rule is devised,and the crowd flow statistics system in the public scene is constructed,the accuracy rate can reach more than 95% after testing.Figure [42] Table [20] Reference [81]...