Optimization Of Multi-Object Identity Recognition Algorithm Based On Video Tracking In Complex Scenes

Gait recognition has become an important research direction in the field of multi-object identification because it does not require the subject to cooperate and support remote recognition.The existing gait recognition algorithms have been basically able to complete the identification of multiple targets in the daily monitoring scene,but the accuracy and efficiency of recognition need to be improved in the complex scene where the analogue interference and the significant appearance changes of the target occur.Therefore,this dissertation studies and optimizes the multi-objective identification algorithm based on video tracking.In this thesis,firstly,the multi-target contour segmentation algorithm based on video tracking is designed to obtain the gait sequence of multiple targets based on the scene with the interference of similar objects.Then,the identity recognition algorithm based on the multi-target gait contour is explored to extract the gait features and carry out identity recognition.Then,the identification model is pruned to improve the recognition efficiency.Finally,based on the above algorithm,the multi-object identity recognition system in complex scenes is designed and implemented.The main innovative work of this thesis is as follows:(1)A multi-object segmentation algorithm based on video tracking is proposed to solve the problems that the segmentation accuracy and speed of the existing video object segmentation algorithms are not ideal in complex environment and the segmentation effect decreases significantly when segmenting long video.This algorithm extends the STM algorithm.Firstly,object constraint module is added and a multi-angle attention mechanism is proposed to solve the problem of wrong fragmentation in segmentation results.Then the tracking module is added and the relative position coding method is improved to solve the problem of missegmenting similar objects.Finally,the feature interaction function is added to the memory reading module and the long-term memory algorithm is introduced to improve the segmentation effect of long video.Experiments were carried out on DAVIS2017 and You Tube-VOS data sets.The results show that the video tracking based multi-objective segmentation algorithm designed in this thesis can get better segmentation results under complex environment,and provides a feasible solution to multi-objective segmentation.(2)Aiming at the problems that the current gait recognition algorithm will reduce the recognition accuracy and the efficiency of multi-person recognition due to complex factors such as occlusion,analogue interference and illumination,an identity recognition algorithm based on multi-object gait contour was proposed.Firstly,in order to extract gait features with significant discriminant,pyramid attention mechanism is introduced into feature extraction network.Then this thesis proposes a multitarget recognition algorithm based on machine learning algorithm to balance the recognition accuracy and speed.Finally,the model pruning method is adopted to reduce the multi-dimensional channel redundant information of the convolutional layer and improve the recognition efficiency of the model.Experiments were carried out on CASIA-B data set,and the experimental results show that the multiobjective gait contour based identity recognition algorithm designed in this thesis can achieve high recognition accuracy and efficiency.(3)Based on the multi-object segmentation algorithm based on video tracking and the identity recognition algorithm based on multi-object gait contour,a multi-object identity recognition system based on video tracking in complex scenes is designed and implemented.The system mainly includes multi-object data input module,multi-object segmentation module and multi-object gait recognition module.In addition,in order to ensure the real-time data,it supports the online recording of multiobject video.After system demonstration and test,it is proved that the system has high accuracy and efficiency in the identification of multiple targets in complex scenes.