Research On Pedestrian Re-identification And Target Tracking Based On Multi-modality

In recent years,with the continuous improvement of urban surveillance networks,video surveillance has been widely used in social security.In the process of video surveillance,due to the environment,angle,light and other factors of the camera,there are big differences between the same pedestrian at different moments of the same camera or different cameras,especially during the day and night.Under light conditions,the camera will capture images in two modes: visible light and infrared.How to identify and track target pedestrians under different modalities has become the focus of this article.Aiming at the modal difference between visible light image and infrared image,the single-modal and cross-modal pedestrian re-recognition are studied separately.Singlemodal pedestrian re-recognition uses a standard benchmark model,using Res Net50 as a feature learning network,selecting classification loss and triple loss to train the network,and finally through some training strategies to improve the recognition rate of the model.Cross-modal pedestrian re-identification after analyzing the classic network model and its experimental data,a model based on a generative adversarial network is proposed.The model is divided into two modules: modal transformation and feature learning.The former uses the generative adversarial network.The visible light image is converted into an infrared image to reduce the difference between different modalities;the latter trains the model through cross-entropy loss and triple loss to reduce the difference within the same modal.Comparing the experimental data of training strategies for single-modal pedestrian re-recognition,it is concluded that the two training strategies of REA and BNNeck have the most significant improvement in the recognition rate of the model,which verifies the efficiency of the single-modal benchmark model;based on the experimental analysis of the generational confrontation network model,it is concluded that the recognition rate of this model in the SYSU-MM01 data set is higher than other classic models,which proves that the model based on the generation confrontation network is feasible in solving the problem of cross-modal pedestrian re-recognition.Further applied research is carried out on the single-modal and cross-modal pedestrian re-identification network models proposed above,combined with target tracking,and proposed application models in two modalities.The single-modal application model detects pedestrians through the SSD network,the single-modal network retrieves pedestrians,and the KM matching algorithm tracks pedestrians;the cross-modal application model also uses SSD networks and KM matching algorithms for target detection and tracking,while the re-identification module uses the above Proposed cross-modal network.Put the studied single-modal model and cross-modal model into the video library recorded in this article for experimental analysis.Both modal models have good performance in the video library,verifying the feasibility of the two models.This paper studies pedestrian re-recognition and target tracking in single mode and cross-modality,uses standard pedestrian re-recognition Baseline to solve the problem of single-modal pedestrian re-recognition,and proposes a model based on generative confrontation network to solve cross-modal pedestrian re-recognition Problems,and by combining target tracking to carry out application research,it has high theoretical significance and practical value for the research and application of video surveillance technology.