Occlusion-based Pedestrian Re-identification Based On Deep Learning

Person re-identification is an important task in computer vision,and its purpose is to retrieve specific person in images or videos.In recent years,with the development of deep learning,many breakthroughs have been made in the community of person re-identification.Person re-identification has been widely used in many fields such as security and criminal investigation.At present,traditional person re-identification methods are mostly dedicated to extracting rich image features.However,when faced with occlusion problems in actual scenes,their recognition accuracy will be greatly reduced.According to the degree of the occlusion,this paper divides the occluded person re-identification into two types: local occlusion and partial occlusion.Based on deep learning,this paper conducts research on occluded person re-identification.The main innovations are as follows:(1)Aiming at the problem that the accuracy of most person re-identification baselines is not high enough,this paper proposes a simple and efficient person re-identification baseline.Considering the structural simplicity of the baseline,this baseline uses only one global feature,and the loss function includes a triplet loss and a classification loss.This paper uses six strategies to train and optimize the person re-identification baseline,including warm-up learning rate,random erasing data augmentation,label smoothing,expanded feature map,batch normalization network and attention mechanism.Under the premise of almost no increase in the amount of parameters,the accuracy of the baseline is further improved,and the impact of the above six strategies to the accuracy and complexity of the baseline is analyzed through ablation experiments.Experiments show that the baseline proposed in this paper surpasses most current baselines in the recognition accuracy,without increasing the amount of parameters.This proposed baseline is used as the baseline for the following two models for solving local occlusion and partial occlusion..(2)Aiming at the local occlusion problem in occluded person re-identification,this paper proposes Triplet Erasing-based Data Augmentation for Person Re-identification.The core idea of the method is to introduce more locally occluded samples into the training set to improve the model's robustness to local occlusion.This method includes a local distance branch and a triple erasing branch.The main function of the local distance branch is to align the local features of two person images.The main function of the triple erasing branch is to erase a rectangle area in the triplet.This can generate positive and negative sample pairs that are more difficult to discriminate,which can increase the difficulty of training model and improve the discriminative ability of the model.Because the method itself belongs to an operation of data preprocessing,it can be easily used for other person re-identification networks based on triplet loss.(3)Aiming at the partial occlusion problem in occluded pedestrian re-recognition,this paper proposes Attention-based Feature Alignment for Person Re-identification.Unlike traditional methods that use an additional human pose estimation model or a person segmentation model to extract semantic information of human body,the network proposed in this paper only uses spatial attention mechanism without increasing parameters to make the model pay more attention to the salient regions of the image.Moreover,in the retrieval stage,this paper innovatively uses spatial attention mechanism to align the shared visible human regions of the two images,reducing the interference of large-area obstacles on the model,thereby improving the model's robustness to partial occlusion.