Pedestrian Re-identification Based On Deep Hybrid Convolutional Neural Networks

With the rise and popularity of intelligent monitoring systems,the human detection method has obvious disadvantages in front of the massive video data,which promotes development of the automatic video analysis field.Pedestrian re-identification is a vital branch of video analysis,and plays a very important role in behavior constraint,crime tracking and forensic evidence.Using machine learning and deep learning approach,discriminant feature extraction and similarity measure method,to solve the problem of pedestrians retrieve matching.Currently,due to the pedestrian data perspective differences in multi-model illumination,background interference and human body occlusion,accurate and robustness pedestrian re-identification still has a great technical challenge.Based on pedestrian re-identification practical application ideas and existing data characteristics,this research studies the above problems in the pedestrian re-identification,and designs the deep hybrid convolutional neural networks.The main work of this dissertation is as follows:Firstly,in view of the problem that the dislocation of the posture features of pedestrians with the same identity in the multi-shot pedestrian re-identification,under the condition of the non-time sequential multiple frames,design an attitude guidance mechanism,guide the depth convolution networks to study the distribution characteristics focused on the pedestrian body area and the joints mutual information by using the global human posture confidence figure.In order to reduce the weight of background interference and implicitly align the spatial distribution of pedestrians with the same identity.Under the condition of temporal multiple frames,embeds non-local convolution blocks into the deep neural networks.Therefore,the network model of the feeling attention can range from local to global,and the model has the ability to establish a mapping relationship for pixels with a certain spatial distance.And finally realizes the alignment of pose features and the extraction of spatial features between temporal frames,mitigates the effect of background and attitude differences on the accuracy of re-identification.Secondly,for the problem of robust feature extraction in multi-shot pedestrian re-identification,under the condition of the non-time sequential multiple frames,designs a pose-guided hybrid convolutional neural networks.The representation information of pedestrian posture can be explicitly introduced into the original data to enhance the learning of human regional features.The 2D residual network structure is improved,the shallow layer feature and the pedestrian intrinsic feature are extracted respectively by using the four-channel convolutional layer and the terminal pooling layer of the networks.At the same time,a weighted precision fluctuation estimation based hybrid representation mechanism is designed to integrate the distribution feature and category identification feature of the pedestrian,so as to enhance the robustness of pedestrian characteristics.Under the condition of temporal multiple frames,a spatio-temporal based three-dimensional non-local hybrid convolutional neural networks is proposed.A spatio-temporal separation convolution block is designed and combined with the above-mentioned non-local convolution block to extract and fuse the pedestrian spatial distribution features and temporal gait features.Meanwhile,the network training parameters are simplified,and the non-linear expression ability of the spatio-temporal features is enhanced.In both temporal and non-temporal network models,fully connected blocks are used to replace the traditional single fully connected layer for pedestrian category division,which is used to enhance the coupling between pedestrian features and pedestrian identity categories and enhance robustness and discriminability of pedestrian feature.In summary,this paper analyzes the characteristics of pedestrian re-identification tasks and data distribution,studies the existing technical challenges,and proposes methods based on deep convolutional neural networks to achieve more accurate performance of pedestrian re-identification retrieval.It provides effective design ideas and opinions for pedestrian re-identification algorithm and practical applications.