Pedestrian Re-Identification Method Based On Joint Weak Supervised Learning

Pedestrian re-identification is a computer vision method to judge whether the pedestrian captured by the camera without cross shooting area at diverse times and the target is the same person according to the given pedestrian image.Traditional pedestrian re-identification method have limited feature description ability,which is difficult to meet the needs of complex scenes.The pedestrian re-identification method combined with deep learning mainly focuses on the global or local features of pedestrians,which can adapt to complex scenes such as occlusion and clutter,and the extracted features have stronger robustness.The features extracted by the former cover the whole pedestrian image,and the detection results are easily affected by clutter background,occlusion and other noises;The latter divides the features of pedestrians into blocks and has better detection results,but it ignores the relationship between the features of each part of pedestrians and the whole.When different pedestrians have similar features in the same body part,it is easy to detect errors and reduce the accuracy.In order to make full use of the advantages of the above methods,this thesis proposes a pedestrian re-identification algorithm based on context relation CR-PRA.CR-PRA is based on Res Net-50,contextual relationship between modules CRBM and local relationship modules LRM are added.CRBM combines the advantages of global average pooling GAP and global Max pooling GMP,and designs a new pooling method to extract global features with stronger pedestrian discrimination;LRM not only uses local features to retrieve pedestrians,but also supplements the context of local features.The experimental results show that the m AP value of this method is 9% higher than that of PCB algorithm on public data sets,and it is also better than the mainstream algorithms in recent years.Compared with the mainstream pedestrian re-identification methods,CR-PRA algorithm has a certain improvement in the effect,but there are still some shortcomings in the detection of low resolution occlusion pedestrian.Based on this,this thesis proposes a pedestrian recognition algorithm based on modified context relation MCR-PRA.MCR-PRA combined with weak supervised learning method can improve the situation of insufficient pedestrian feature extraction in low resolution occlusion by sharing network parameters;Finally,the improved k-nearest neighbor algorithm is used to optimize the whole network.The m AP value of MCR-PRA algorithm is 1.8% higher than that of CR-PRA algorithm,which improves the robustness of the network to low resolution occlusion of pedestrians in real scenes and optimizes the whole network.