Research On Person Re-identification Based On Fusion Neural Network

With increased awareness of public safety,video surveillance is seen in many areas as a primary tool for enhancing security and solving crimes,or as a crime deterrent,and it will play an important role in security enforcement.The task of person re-identification comes into play when a pedestrian disappears from the view of a surveillance camera and appears in the view of another camera.The technology has become one of the most popular research topics in recent years due to its importance in the search for missing children,the tracking of criminals and the recovery of pedestrians.However,the huge amount of surveillance video data hinders the development of person re-identification tasks,and the traditional manual processing of data images does not meet the current stage of security and criminal surveillance systems for identification and pedestrian needs mining.Therefore,this thesis will further promote the research work of person re-identification technology in safe life application.Firstly,a feature fusion-based person re-recognition is proposed to address the problems such as the difficulty of effective extraction of feature descriptors by designing a feature extraction framework with fusion-based ideas,combining traditional manual-based feature extraction with Convolutional Neural Networks(CNN)based feature extraction,using a neural network to learn in an end-to-end manner,and using internal state(memory)to accumulate key information of each sequence to construct the final features that represent the input.Secondly,image pre-processing knowledge is used to deal with the problems of pedestrian pose variation,height variation and spatial misalignment faced in the person re-identification task.Finally,the person re-identification problem is transformed into a distance metric problem as a measure of similarity between images,which in turn indirectly solves problems such as the existence of angular anomalies in the pedestrian image dataset.In order to further improve the feature representation method,pedestrian re-recognition embedded in a ternary convolutional neural network is proposed.Firstly,the ternary model is embedded into the pedestrian re-recognition extraction framework,and the ternary model is used to expand the breadth of extraction of the dataset,and the original dataset is used to synthesise new data or edit the existing database to make the training network better match without losing information,increasing the driving force of the network to extract features.Secondly,new pose training networks are generated using pedestrian images to obtain a more comprehensive set of pedestrian features,while generating obscured body parts to provide a new feature set for the model.In addition,the loss function is effectively designed to help the whole network framework converge quickly as well as to speed up the classification of pedestrians.Finally,to address the inherent deficiencies of the ternary neural network model,the central loss function is combined with the ternary loss function,and the gradient descent algorithm is used to update and optimise the network weights and parameters to overcome the sample imbalance and refine the similarity matching of the network model.Following the above analysis,experimental demonstrations were carried out through relevant datasets and comparison experiments with several different algorithms,which effectively demonstrated the feasibility,advancement and scalability of the fusion-type idea.