Research On Pedestrian Re-Identification Technology Based On Feature Depth Analysis

With the continuous advancement of smart cities,people pay more and more attention to public safety.With the continuous increase of surveillance camera equipment,the surveillance image data also continues to increase.It is often difficult to find the target pedestrian quickly and accurately by relying on manual methods to find and identify pedestrians,and it costs a lot of money.This article studies pedestrian re-identification technology.Try to solve the above problems.Pedestrian Re-Identification(Re-ID)is to identify the same person from videos of different cameras without any overlapping areas.This paper starts with constructing image feature blocks and uses the data fitting capabilities of deep neural networks to design a multi-feature fusion network model.The work completed in this article is mainly in the following areas:(1)According to the feature representation method,a human body pose estimation algorithm is used to divide different feature parts of a single human body.The pose estimation algorithm is used to extract the joint point information of the pedestrian image,and the human body image is divided into rigid and non-rigid sub-images based on the 18 bone point positions of the pedestrian image.These two sub-images respectively contain specific body sub-regions,which can better extract the local features of the pedestrian body in the image,and at the same time,can handle the misalignment between pedestrian images caused by changes in posture and perspective.(2)A two-branch neural network structure with feature fusion is proposed.Based on the ResNet50 network,neural network structures with pedestrian global images and local feature sub-images as input data were constructed.For the global feature neural network structure model,a soft attention module is added to reduce the interference of noise information in the region.According to the proportion of different local components in the pedestrian image in the entire pedestrian image,a weight coefficient is set to allow the global features and local features to be weighted and fused to form pedestrian features with more individual discrimination.Finally,in the model training phase,the triplet loss function is applied as an objective function to the overall network model.(3)Experiments were performed on the CUHK03 dataset and the DukeMTMC-ReID dataset using the improved convolution network model.The research results show that the improved dual-branch convolution neural network has better performance.