Research Of Person Search Based On Deep Learning

Person search aims at jointly optimizing pedestrian detection and person re-identification to realize identity consistency association of person across cameras based on visual information.It expands single camera video surveillance into multi-camera collaborative video analysis to optimize the performance of the overall monitoring system.At the same time,it can better meet the actual application.Through a large number of related research at home and abroad and the main problems of person search at this stage,this paper proposes high precision end-to-end person search algorithm and improved algorithm based on the idea of the attention mechanism and evenly splitting the image from the perspective of multi-task learning.Person search can be broken down into two separate tasks,pedestrian detection and person re-identification.But there are many deficiencies with operability and efficiency in practical applications.We use the shared convolutional neural network to extract features in a single multi-task network jointly dealing with these two tasks.However,the end-to-end neural network is too large and needs to process a large amount of input information.Relying on human attention mechanism,it can select the key information for processing to improve the efficiency.Specifically,we use the SE module to explicitly model the interdependence between feature channels which mean to automatically learn the importance of each feature channel through the neural network.Then we enhance the useful features and weaken the less useful features to this importance.The experiments on the dataset CUHK-SYSU verify that the end-to-end person search processing model based on the attention mechanism can better implement the person search function.Due to the diversification of shooting angles,sharpness,lighting conditions,person poses and the presence of partial occlusion in the actual surveillance scene,the appearance of person will strongly change.So person search across cameras is more difficult.At present,researchers focus on constructing excellent feature representations and learning reasonable feature matching models.Local body features can provide fine-grained information for matching tasks with the precondition that each part is accurately located.Instead of using extra labels,we directly locate local locations,emphasizing the content consistency of each local location.Specifically,in order to learn the distinctive features for person,we evenly split the detected person image and use a convolution descriptor composed of several local features as the feature of person for matching.The experiments show that the accuracy of the improved end-to-end person search processing model has been further increased.