Visual Tracking Via Correlative Discriminative Model

Visual tracking is a hot topic in the field of computer vision,aiming to accurately find the location and size of the target in subsequent frames only by the location and size of the target in the first frame of a given video.In recent years,visual tracking algorithms based on correlation filters have attracted extensive attention from researchers due to their excellent tracking accuracy and speed.Among them,the visual tracking algorithm based on correlation filters,because it guarantees accuracy and has faster operation speed,many excellent works have emerged.This paper makes the following researches on the visual tracking algorithm based on correlation filters:First,it provides a new solution to the problem that the correlation filter algorithm is susceptible to boundary effects.Previously designed correlation filters roughly used fixedbandwidth window functions to suppress boundary noise,which inevitably lost the information needed for training.Based on the shortcomings of the traditional models,this paper first achieves more effective suppression of boundary effects by adaptively adjusting the bandwidth of the window function,and dynamically changes the bandwidth of the window function to retain more effective information.The window function bandwidth is adjusted according to the change of the filter confidence value of the specified area,so that the filter template fixed area value and the window function bandwidth change in a positive correlation trend.Second,the discriminative correlation filter tracking algorithm based on channel and spatial reliability is improved when the target is occlusion and background interference is easy to lose,the regular penalty of the time term is introduced,and the calculation strategy of channel reliability is adjusted.When training the filter,constraints are made in three aspects: space,time and channel,which greatly enhances the discriminative power of the filter template.The adaptive window function mechanism is fused on the correlation filtering model based on manual features and depth features,respectively,and increase accuracy and success rate on the two OTB datasets by 1 to 3 percentage points.For the channel,space and time constrained algorithm,the success rate and accuracy rate on the OTB100 dataset exceed the baseline by 1.2% and 1%,respectively.