Key Technologies For Person Re-Identification Based On Multi-Strips Of Local Feature Network

Person re-identification,a research hotspot in the field of computer vision and pattern recognition,refers to the task for retrieving specific pedestrians across different camera view in an image or video sequence.What's more,person re-identification is a valuable and challenging research and remains an important application prospects in the fields of intelligent video surveillance and intelligent security.Since global features often fail to capture local details of pedestrians,how to learn a discriminative local feature has become a key component of person re-identification methods nowadays.Different from most other person re-identification methods that use external human segmentation or human posture prediction models to obtain local features,this paper proposes a novel person re-identification method based on multi-strips of local feature network,in which external models are not included to avoid transfer deviations.The proposed method is with a simple structure but a high performance in person re-identification.In this paper,the pre-trained Resnet-50 model on the large image classification dataset ImageNet is used as part of the backbone network.The intermediate features extracted by the backbone network are evenly partitioned and trained separately,by which the category information contained in each partial block is fully utilized.The model is encouraged to learn an effective partial feature representation autonomously,without external data models to divide human local parts.Targeting the specific problem of person re-identification that the cross-class similarity is larger than the intra-class similarity,this paper adopts hard example mining method in the training block to extract difficult samples to facilitate effective learning.What's more,the similarity distance of local feature strips is dynamically adjusted to handle the mismatch between the corresponding local feature strips.By replacing the Euclidean distance to the shortest distance,local feature strips are aligned in the horizontal direction macroscopically.Based on theoretical research and experimental platform,this paper designs and implements an end-to-end person re-identification method,comparative experiments and model analysis are conducted to prove the superiority and effectiveness of the proposed method.The experimental results on the main person re-identification datasets including Market-1501,DukeMTMC-reid and CUHK03 show that the proposed method is superior to most existing person re-identification methods.Multiple sets of comparative experiments are conducted to verify the influence of each model parameter on the performance.The experimental results show that the multi-strips of local feature has a better performance than original backbone model and there is a further improve when the shortest metric learning method is added.For the effect of the number of strips of local feature on the model performance: mAP and Rank-1,the line graph reflects that when the number of strips n increases,mAP and Rank-1 decreases at first and reaches the highest at n=6,then decreases when the number of strips get larger,which means a better performance for n=6.This paper also conduct comparative experiments for different backbone networks.Considering both the performance and the cost,ResNet-50 is choosed as the basic backbone network.The experimental and evaluation results proved the validity and superiority of the person re-identification method based on multi-strips of local features network proposed in this paper.