Research On People Re-identification Algorithm Based On Deep Learning

Person Re-Identification,also known as Re ID,is a specialized form of computer vision technology.It possesses the capability to ascertain the presence of a specific pedestrian in any given image or video,commonly regarded as a subcategory of the broader image retrieval issue.Deep learning has shown exceptional advancements across various domains of computer vision,which includes pedestrian re-identification that finds extensive applications in sectors like security,criminal investigations,and smart cities.This paper deeply studies the current pedestrian re-identification algorithm.Traditional methods are difficult to obtain high accuracy due to complex problems such as background,lighting,pose and occlusion of the picture.The main work of this paper is as follows:1.A two-branch multi-level loss network based on ResNet50 is proposed.This method draws on the multi-branch structure of the multi-granularity network.First,the network structure is based on the ResNet50 model of pre-trained ImageNet,which consists of two branch structures and uses multiple loss functions.Second,a new branch structure,the scSE-Res branch,is proposed,which is based on the scSE-res module,which consists of a scSE module and a residual module.Third,only one loss function is used in the general network.Here we design a multi-level multi-loss function to obtain more layers and more types of features.Experiments are carried out using the proposed network to verify the effectiveness of the network.2.An attention mechanism network based on full-scale features is proposed.Due to the camera angle,the environment of pedestrians,and the transformation of pedestrian poses,the intra-class differences among pedestrians with the same ID are large,and the inter-class differences between different IDs are small.Aiming at the problem of insufficient feature extraction capabilities of full-scale feature modules,an improved model based on full-scale feature module networks is proposed.At the same time,the CBAM attention mechanism is used to integrate the first convolutional layer and the last convolutional layer of the backbone network,and The loss function is optimized using the center loss.On the public datasets of Market1501 and DukeMTMC-reID,comparative experiments were conducted,and the improvement effect of the model was obvious.3.A pedestrian re-identification algorithm based on non-local features is proposed.The convolution operation of ResNet50 is a local operation and can only obtain local information.In order to obtain more dimensional information in the pedestrian reidentification task,the model uses non-local attention The mechanism builds spatiotemporal attention to obtain global information,and the calculation of a certain location feature can associate all locations,thus avoiding the limitations of convolution operations in the basic network.In order to make the smallest similarity of the same kind greater than the largest similarity of the same kind,the loss function is optimized,because the original loss function lacks flexibility,and the unified loss is a unified expression of various losses,which is more flexible.In multiple comparative experiments,the improvement effect of the model is obvious.