Research On Person Re-identification Algorithm Based On Aligned Multi-granularity Body Structure

In social life,video surveillance is one of the important methods to protect citizens'life safety and property safety.With more and more surveillance cameras,traditional methods of video surveillance which using professional surveillance personnel face huge challenges on cost and efficiency.When the security incidents occur,traditional methods which depends on security personnel have low timeliness and poor performance.In the security industry,many companies want to use computer vision technologies to solve above problems.Person re-identification technology is one of the important technologies applied in this field.With the rapid development of deep learning in the field of Person re-identification,more and more novel deep learning models are invented to solve this task.In the task of person re-identification,researchers face many problems such as camera settings,illumination changes,human body posture changes,pedestrian body occlusion,image similarity,background differences,etc.As far as now,the most advanced algorithms cannot solve all the above problems well.But Many research results in this field indicates that the combination of global and local features of pedestrian images can be one of the key points to improve the discriminative performance of person re-identification algorithms.This paper proposes a machine learning model based on deep convolutional network called MHSANet,which can extract the global and local features of pedestrian and reduce the influence of background in image,and through aligning multi-granularity feature and utilizing multi-loss function to train itslef,this network performance is improved.The main work and innovations of this paper is as follows:(1)A human body component detection network based on RPN is proposed.The network module is mainly used to detect various parts of the human torso in the pedestrian image.The human body component detection network proposed in this paper can locate the image region where human body components may exist.Compared to the algorithms which divide a pedestrian image into regions by lines or grids.This algorithm proposed in this paper can reduce the influence of background in image,and the extracted features in this network can be directly fused without realigning,so this network can avoid the error caused by component alignment.In addition,compared with the algorithm which directly locate the human body region by the key points of the human body,the network proposed by the present invention integrates the human body components.The network can reduce the final body component feature error caused by the key point positioning.(2)Combining the human body component detection network and image feature pyramid network,a multi-level multi-granularity human body local feature extraction algorithm based on ROI Align and feature map fusion network is proposed,and multi-granularity re-identification network is used to accomplish pedestrian re-identification task.The multi-granularity identification network has three branches according to the granularity level of the human body features,making full use of the multi-level multi-granularity pedestrian features.Compared with the single-task learning network,this paper combines the metric learning method,uses the multi-task learning strategy,and uses the double loss function to jointly optimize the network,and can complete the pedestrian re-identification task while making each kind of granular pedestrians.Features are discriminative and specific,which in turn improves the ultimate performance of the network.The overall network working principle is more interpretable,which is also beneficial to the deep network parameter adjustment process.(3)In summary,this paper proposes a pedestrian re-identification algorithm with multi-granularity human body structure alignment.The algorithm combines human body part detection network,multi-granularity identification network and multi-task learning into a deep network model,which enables the network to be optimized at the same time,reducing training time and training operation complexity.The experimental results on Market1501 show that the Rank-1 re-identification rate of the algorithm reaches 79.3%,which proves the feasibility and effectiveness of the deep learning network proposed in this paper.