Research On Visual Tracking Algorithm Based On Correlation Filters

Visual tracking has always been a core technology in the areas of autonomous driving,intelligent monitoring,and human-computer interaction.With the advent of the era of artificial intelligence and the rapid development of computer hardware and equipment.As an important branch of computer vision,visual tracking has become a key research object for researchers at home and abroad.But it is always difficult to design a tracking algorithm that can not only effectively meet the challenges of various environmental changes,but also can achieve the real-time and high-speed tracking effect.This thesis is guided by the above objectives,and aims to develop efficient,real-time and robust tracking algorithms based on correlation filters.This thesis mainly researches from the three aspects of building algorithm models,target feature expressions and response graph fusion strategies.The main work and innovations of this thesis are as follows:From the perspective of constructing the algorithm model,a cooperative model of global block and local block was established based on KCF.The collaborative model relies on the cooperative interaction between the global filter and the two local filters.The local filter provides an initial estimate of the final position of the target,and then the global filter determines the final result.On the basis of the collaborative model,a method to guide model update based on effective local blocks is proposed,and the evaluation criteria of effective local blocks are given.This method can better deal with the model drift caused by partial occlusion.At the same time,the model estimates the scale of the target by analyzing the distance change between the local blocks of the two frames before and after,and judges whether the current target scale and the filter updated based on the stored average of the scale factors of all frames since the filter is reinitialized.The method can solve the problem of tracking failure caused by the change in target scale.The overall performance of the collaboration model is better than several other improved algorithms based on KCF,and the tracking speed reaches 32 frames per second.Based on the target feature expression and response maps fusion strategy,four different features were proposed to represent the target object,and a response map fusion strategy was proposed to improve the tracking performance in different scenarios.This fusion strategy uses the peak sidelobe ratio to weight the response of each feature to filter the noise of each response.Then the processed four characteristic response maps are fused to obtain six different response maps.The final weighted fusion of the six response maps results in a final improved,more confident,and less noisy response map.Integrating the proposed scheme into the BACF constitutes the MFOL algorithm of this thesis.The tracking accuracy of this algorithm is better than many current advanced tracking algorithms based on Correlation Filters.At the same time,it can better meet the challenges of various environmental changes,has better robustness,and the tracking speed reaches 21 frames per second.