Person Re-identification In Complex Scenes Based On Local Feature And Spatial Transformation

Research of person re-identification based on deep learning has made some progress.When applying the person re-identification to a monitoring system in a real scene,it often encounters changes in camera viewing angles,occlusions,and undesirable lighting conditions.These non-ideal scenes are important challenges facing the current person re-identification and the key to future real project applications.This thesis proposes innovative algorithms for person re-identification for two most common disturbances in non-ideal scenes:For non-ideal scenes with occlusion,this thesis proposes a local person re-identification algorithm based on multi-scale occlusion.According to the different characteristics of deep and shallow networks,a deep occlusion branch and a shallow occlusion branch are designed.The two parallel branches are weighted to provide a richer scale features.The occlusion mask designed in the two branches makes the network learn the global and local features of person by randomly covering part of the image.The introduced attention mechanism enables the network to not only learn the discriminative information after occlusion,but also the changes of occlusion.The algorithm has stronger robustness and learning ability,and to some extent alleviates the challenge of non-ideal scene recognition caused by occlusion.This thesis proposes cross-domain person re-identification based on spatial transformation for another non-ideal scene caused by dramatic lighting changes.When the light is poor,the infrared image provided by cross-domain person re-identification can effectively compensate for the impact of the lack of color.The algorithm proposed in this thesis spatially transforms images of different modalities to achieve feature alignment,thereby accurately obtaining the shared features of the two modalities.Spatial transformation enables the network model to extract local features that are still expressed stably in low-light environments,so that the network maintains stable performance in dark scenes.