Research On Single Object Tracking Algorithm Of Siamese Network Framework

The object tracking task predicts the location of a specific target by analyzing its characteristics in a video sequence.In recent years,the object tracking algorithm based on Siamese network has gradually become a mainstream method.The pre-training and usage methods of the feature extraction network in current target tracking algorithms based on Siamese network architecture limit tracking accuracy,and there is a lack of suitable hyperparameter optimization algorithms in the tracking inference process.Based on the above problems,this thesis designs and implements a video surveillance recognition and tracking system based on the object tracking algorithm of the Siamese network.The main research work and contributions can be summarized as follows:Firstly,an improved target tracking algorithm based on Siamese networks is proposed.A comparative learning network framework is designed to pretrain the feature extraction network,the integration of multi-scale features improves tracking performance,and a quadratic tracking model is introduced to enhance the accuracy of position prediction.Compared with the original Ocean algorithm,the algorithm proposed in this thesis improves Success and Precision by 3.7% and 2.8% respectively;compared with the mainstream Siamese network architecture target tracking algorithm Auto Match,the algorithm proposed in this thesis improves the Success value by 0.9%while maintaining the same Precision value.Secondly,a Siamese network target tracking hyperparameter optimization algorithm based on the improved particle swarm optimization algorithm is proposed.The particle swarm optimization algorithm is applied to the parameter optimization of Siamese network target tracking,the upper limit of particle velocity is dynamically changed based on performance feedback method,the dynamic inertia weight method is designed to associate the inertia weight value with the particle fitness value,the experimental results under different mapping relationships between inertia weight and particle fitness value are studied,and the simple position mutation and binary position mutation methods are introduced to help the particles jump out of the local optimum.Experimental results show that,compared with the target tracking algorithm optimized by the classical particle swarm optimization and the target tracking algorithm optimized by Bayesian optimization,the target tracking algorithm optimized by the improved particle swarm algorithm improves the average overlap degree by 1.04% and 1.09%respectively.Finally,a video surveillance identification and tracking system is designed and implemented.Integrating the above algorithms,the system designs and implements subsystems such as video upload,video recognition,target tracking,and user information management,and the success rate of the algorithm and the functionality of the system are tested.