Research On Pedestrian Re-identification Algorithm Based On Deep Learning

In the face of massive urban surveillance video data,intelligent analysis instead of manual viewing has attracted the attention to a large number of researchers in the field of security.Pedestrian re-identification,as the core of surveillance video content analysis,has developed rapidly.Its purpose is to solve the problem of searching for a specific pedestrian in a massive image or video library through the appearance of visual characteristics and motion characteristics of pedestrians after cross-camera.The identities of pedestrians captured by different cameras on the internet are correlated to obtain the movement trajectory of a specific pedestrian in time.In real life,due to the complex and changeable monitoring environment,different camera placement positions,and different light intensity,the same pedestrian captured images under different cameras are inconsistent,which increases the difficulty of pedestrian re-identification.Aiming at the influence of factors in different scenarios on the accuracy of pedestrian re-identification,this thesisi takes a convolutional neural network as the core and proposes two network models that can effectively improve the accuracy of re-identification.The main research contents are as follows:1)Aiming at the effects of noise,partial occlusion between pedestrians,and change of posture on re-identification in pedestrian images captured by cameras,a pedestrian re-identification algorithm based on a spatio-temporal attention mechanism is proposed.First,Res2Net-50 is used to extract multi-scale pedestrian features,and at the same time,the attention mechanism is used to recalibrate and filter the unactivated or incorrectly activated parts from the original feature map extracted by the basic backbone network to extract more distinguishing pedestrian features.Secondly,a 3D aggregation module is used to aggregate the re-screened frame-level pedestrian features to make full use of the spatio-temporal information on the video sequence to obtain a robust pedestrian feature vector representing the entire video sequence.After a large number of ablation experiments on the three data sets of i LIDS-VID,PRID-2011,and MARS,the rank-1 reached 85.5%,92.7%,and 85.6% respectively,which is better than most current methods.2)Aiming at the spatio-temporal dependence on video sequences and the spatial misalignment during the comparison of pedestrian feature maps,a pedestrian re-identification algorithm based on a non-local 3D dense convolutional neural network is proposed.First,a 3D dense convolutional neural network is used to replace the 2D convolution operation,and the space-time dimensional features are extracted at the same time,to fully mine the hidden spatio-temporal information in the input sample sequence.Secondly,to effectively capture the long-distance spatio-temporal dependence on the video sequence,and at the same time solve the problem that the spatial misalignment between the pedestrian feature maps reduces the accuracy of re-identification,a non-local block is added to the 3D dense convolution block.Through non-local operations,the relationship between any two spatial positions in the feature map is calculated to capture the long-distance dependence on the feature map,and the problem of spatial misalignment is solved at the same time.After performing multiple ablation experiments on the i LIDS-VID and MARS data sets,the rank-1 reached 84.3% and 87.8%,respectively,which proved that the method has excellent performance.