Improvement And Application Of Video Pedestrian Re-recognition Based On Deep Learning

The application of searching for a specific pedestrian from the video based on the image is more extensive,and it is mostly used in the security field.At present,most of the research on pedestrian re-recognition focuses on image retrieval,while pedestrian re-recognition for video is relatively scarce,and at the same time,it is in great demand.The paper combines pedestrian re-recognition and pedestrian detection,uses pedestrian detection to capture all pedestrian images from the monitoring perspective,and uses the corresponding model to retrieve whether these images contain specific pedestrians,so as to realize the search and location of the video through pictures.Specific target pedestrians.Furthermore,after completing the combination of pedestrian re-recognition and pedestrian detection,the method selection of pedestrian detection and pedestrian re-recognition was studied and discussed,and the pedestrian re-recognition network was improved and optimized.For the pedestrian re-identification preprocessing part,namely pedestrian detection,the traditional machine learning pedestrian detection algorithm represented by HOG+SVM and the deep learning-based pedestrian detection algorithm represented by YOLO v3 are used respectively,and the two representative algorithms are respectively used Experimented.The test results show that the YOLO v3 pedestrian detection algorithm based on deep learning has higher recognition accuracy and more accurate pedestrian candidate frames,so it is more suitable for the current research environment.In terms of pedestrian re-identification,a basic pedestrian re-identification network model is established,and on the basis of this model,a variety of data preprocessing and data enhancement methods are applied.Research has shown that adding random erasure,padding,random cropping,random flipping,and random noise to the basic network of pedestrian re-identification improves the recognition effect of pedestrian re-identification.According to the evaluation index m AP of target detection,the model increased by 4.5% on the market1504 data set,and its rank?1 index increased by 1.2%.Aiming at the over-fitting phenomenon of the pedestrian re-recognition network in the later training period,the strategy of adding Dropout to the network to randomly disconnect neurons is adopted.The experimental results show that when the dropout probability is 0.3,the over-fitting of pedestrian re-recognition can be relieved to a certain extent.At the same time,after replacing Batch Norm regularization with Group Norm regularization at the end of the network,the pedestrian re-recognition model can converge ahead of time at the expense of very few m APs,while the network's Rank?1 index hardly changes.Finally,the paper expands two applications based on this function for the above model:(1)Locate and find specific pedestrian targets that need to be retrieved;(2)Record the path of a specific pedestrian and draw a heat map.Through experiments,the feasibility of the above two applications has been verified.