Research On Key Feature Extraction And Enhancement Method For Video-based Person Re-identification

Person re-identification is a key technology for tracking pedestrian across a wide range of surveillance cameras,and plays an important role in the systems construction of intelligent security and intelligent business.At present,due to attribute parameter,shooting angle and shooting environment of different monitoring equipments,pedestrian target video has large differences in lighting,image resolution,attitude and background,which makes it difficult for person re-identification task to meet the actual application requirements in terms of recognition accuracy.There are many practical problems that need to be solved in person re-identification algorithms,such as there are some redundant information in the surveillance pedestrian video,insufficient extraction of local key features.Therefore,focus on the above problems,the paper starts from deep learning methods and proposes a key feature extraction and enhancement method for video-based person re-identification.The research content of the paper can be mainly generalized as the following two points.(1)In order to solve the problem of low recognition accuracy caused by video sequence redundancy in real scenes.A key information multi-frame fusion video-based person re-identification algorithm(Multi-frame Fusion Part-based Convolutional Baseline,MFPCB)is proposed in the paper.First,perceptual hash algorithm and three-frame difference method are used to cross-screen the frames in the single-camera pedestrian video which to be detected.After that,redundant frames with high similarity and low quality are screened and the key features in the video are extracted.Then,the improved Res Net-50 network is used as basic feature extraction network to further perform highlevel semantic feature extraction on the selected optimal key frame.Finally,the multiframe fusion network MFPCB is used to aggregate key information from different views into a single compact discriminant feature description.At the same time,the cosine distance is used to measure the similarity between videos.A series of experiments were conducted on the public datasets MARS and i LIDS-VID,and the experimental results show that the proposed MFPCB video-based person re-identification algorithm has a high matching rate of Rank-1 and m AP,which can solve the problem of video information redundancy effectively.(2)In order to solve the problem that deep convolutional neural network cannot effectively extract and enhance local key features in video,a video-based person reidentification algorithm based on progressive spatio-temporal attention is proposed in the paper.Without pre-segmentation or posture alignment of pedestrian body parts,the spatiotemporal attention module uses multi-scale pooling can adaptively assign different weights to local regions in pedestrian video frames.Therefore,the global feature and local key feature of the human body can be effectively extracted.In practical application scenarios,the multi-scale pooling operation can make the local key features of pedestrians unaffected by illumination,perspective changes and complex backgrounds as much as possible.The multi-task learning idea is used in the entire network,combining hard sample sampling triplet loss function and cross entropy loss function to train and optimize the network.The experimenal results show that Rank-1 matching rate of progressive spatiotemporal attention algorithm on MARS and i LIDS-VID is 84.1% and 81.3% respectively,which is about 3% and 5% higher than the image-based baseline algorithm for person reidentification.It can achieve a recognition rate that is comparable to or even higher than that of existing typical video-based person re-identification algorithms,and has excellent performance in the extraction and enhancement of local key features.The paper explores key frame extraction,local key feature extraction and enhancement of video-based person re-identification,the recognition accuracy of person re-identification has been improved,and has good social significance and commercial prospects in intelligent security.However,In the future,it is necessary to further study end-to-end training model,lightweight construction and compression operation of the model,so as to realize real-time recognition on inexpensive devices and improve the practicality of person re-identification system.