Research Of Person Re-identification Model Based On Super-resolution Images

Security is the premise of pursuing all development.Intelligent security is put forward to protect our personal and property security,and avoid the occurrence or recurrence of theft,robbery,loss and other unsafe events.With video surveillance technology,the potential adverse events can be predicted in advance by analyzing previous behavior.After the event,trajectory tracking can be used to make up for losses in a timely manner.With the soundness of the video surveillance network,it is not only time-consuming and labor-intensive to watch the exponentially increasing data generated daily by human labor alone,but even miss the best time to deal with them.These problems can be solved by automatic monitoring technologies,such as pedestrian re-identification.Pedestrian re-identification is a technology that matches target people across cameras.That is,the goal of automatic tracking and matching of pedestrians is achieved by analyzing the data collected by the monitoring network.As deep learning continues to mature,new breakthroughs have been made in pedestrian re-identification.However,pedestrian re-identification still needs to cope with many obstacles,such as low resolution,obstruction of other objects,different shooting angles,different shooting lighting,changes in people's posture,etc.These conditions will interfere with the final recognition result.Existing methods focus on dealing with inconsistencies caused by lighting,background confusion,and posture changes,but ignore the problem of poor image quality due to low resolution,which will weaken the pedestrian re-identification model to a certain extent.In order to solve the above problems,this thesis proposes a pedestrian re-identification model based on multi-resolution monitoring data,which is used to effectively match person images with inconsistent resolutions across devices.The main contents of our research are as follows:(1)Processing of public dataset for pedestrian recognition: The existing datasets do not fully meet our requirements,that is,the images are multi-resolution and there is a clear correspondence between multi-resolution images.We use bilinear interpolation algorithm to simulate the low resolution images taken in the actual situation.(2)Aiming at the limitation of the existing pedestrian re-identification research,we propose an end-to-end pedestrian re-identification model based on multi-resolution monitoring data.This model is a hybrid model combining super-resolution reconstruction technology and pedestrian re-identification technology.(3)Using modular design scheme to build a stable pedestrian re-identification model combined with residual dense block,inception block and integrated-attention block.And multi-task training method is used to optimize the super resolution reconstruction and pedestrian re-identification.(4)The proposed model is compared with the traditional low-resolution pedestrian re-identification algorithms and hybrid methods of super resolution reconstruction and pedestrian re-identification.And we also analyze the impact of each key design of the proposed model on the final results.Experiments show that the proposed pedestrian reidentification model based on multi-resolution monitoring data is superior to most of the existing models in recognition performance,and each key design of the model is indispensable.