Research On Pedestrian Detection And Re-Identification Across Disjointed Cameras

In recent years,intelligent security has developed rapidly,under the tide of building a safe city and smart city.Video surveillance system has been widely deployed all over the country and plays an important role in criminal investigation,community security and other fields.Large scale video surveillance network not only enhances the monitoring,but also brings a huge amount of data.The traditional manual method by watching video playback is far from meeting the current demand.Therefore,the intelligent video surveillance system based on computer vision technology has attracted extensive attention of researchers.Person re-identification is the kernel of intelligent video surveillance system,which aims to retrieve whether there are specific pedestrians in cross view camera network.Pedestrian detection is the pre-task of person re-identification,and the detection accuracy is directly related to the retrieval effect of re-identification task.However,due to the influence of camera angle difference,illumination change,background clutter,pedestrian posture change,occlusion interference and other factors,pedestrian detection and person re-identification remain great challenges.This thesis analyzes the relevant recent research extensively,and proposes the corresponding algorithms and models for some difficulties in pedestrian detection and person re-identification.In the pedestrian detection,an efficient contour feature descriptor and the camera imaging geometric model is proposed,and then combine with the normalized central moments to solve the problem of occlusion.In the person re-identification,the metric learning in deep convolution neural network and the person re-identification in top view are studied respectively,and the corresponding optimization models are proposed.The main research contents and innovations include:(1)For the feature extraction in pedestrian detection,a contour feature based on elliptic Fourier descriptor and normalized central moments is proposed.By adjusting the relevant harmonic coefficients of the elliptical Fourier descriptor,the target contour descriptor has rotational variability,so as to distinguish the vertical human contour and the horizontal vehicle contour more effectively.In addition,the normalized central moments in the pedestrian contour features can provide the geometric feature information of the contour,which enhance the description of the contour feature.(2)For the pedestrian detection deviation in the case of pedestrian occlusion,this thesis proposes the camera imaging geometric model to obtain the mapping relationship between the height of the foreground and the vertical coordinate of the foreground center in the video image.The confidence range of the pixel height of the pedestrian foreground contour in the video image is determined with the prior knowledge that the pedestrian height is roughly the same.The pedestrian contour beyond the confidence range is identified as occlusion.On this basis,the distribution characteristics of the contour shape are judged through the normalized central moments in the pedestrian contour features,so as to obtain the accurate positioning of the detection box under occlusion.(3)For the appearance change of the same pedestrian in the video person re-identification,a video sequence invariant loss function and a sequence invariant feature learning model are proposed.From the perspective of metric learning,the video sequence level feature vectors are sliced and re-combined according to sequence differentiation,that is,different sequence features of the same pedestrian are dispersed into each group.Applying the propsed loss to minimize the cosine distance of all feature pairs of the same pedestrian,the sequence invariant feature learning model can effectively resist the interference caused by various appearance changes and environmental changes,and enhance the robustness and retrieval accuracy.(4)For the occlusion in person re-identification,this thesis constructs a large-scale top view video person re-identification dataset according to a video recorded by the top fixed camera,which effectively avoids the person occlusion caused by the installation view of the camera.In the person re-identification backbone network,the attention mechanism is introduced to make the model focus on the most recognizable area in the pedestrian image,so as to optimize the feature extraction of the neural network.In addition,the non local operation algorithm is added to increase the receptive field of the convolution neural network and further improve the performance of the model.