Research On Target Tracking Technology Based On Correlation Filter

As an important branch in the field of state estimation,target tracking has been widely study by many researchers because of its wide application background.In recent years,target tracking algorithm based on correlation filtering have attracted more and more attention due to their high tracking speed and considerable tracking accuracy.Although correlation filter has certain advantages in tracking speed and accuracy,due to its boundary effect caused by cyclic sampling and non-detected tracking mode,the robustness of correlation filter is insufficient.Tracking drift and tracking failure are easy to occur.To solve these problems,this paper improved the tracking algorithm form three aspects: backgroundaware,feature fusion and reliability detection.So that the tracking algorithm has higher tracking accuracy and robustness.Specific contributions are mainly as follows:Firstly,on account of boundary effect of training sample discontinuity caused by cyclic sampling,this paper in-depth analyzed the influence of foreground and background information on tracking accuracy.The color probability graph is build based on statistical color histogram,so that the tracker could better distinguish foreground and background information.After that,the background block around the target box is sampled and integrated into the optimization function as a background regularization term to improve the algorithm;s aware of the background information.Finally the qualitative、quantitative and ablation experiments were conducted on the proposed tracking algorithm,which proved that our algorithm,which proved that our algorithm could effectively improve the tracking accuracy.Secondly,using multiple features to track targets is the mainstream method in the filed of tracking algorithm.This paper analyzes the shortcomings of the existing tracking schemes,proposes a set of feature preprocessing algorithms to carry out the value normalization and dimension normalization of different features,so as to prevent the problem of tracker deviation caused by different values and dimensions.Then,the concept of group learning is introduced into the field of the target tracking.By setting a group of correlation filter trackers and using a set of robust detection mechanism,the optimal feature combination is selected for tracking.For the reliablity problem of target tracking results,this papper proposes a set of reliability dection machanism and adaptive updating mechanism to reduce the impact of bad samples on tracking by adjusting the learning rate of the model.Finaly we conducted qualitative,quantiitative and ablation experiments on the proposed algorithm,demonstrating the improvement of tracking accuracy and robustness of the proposed algorithm.Finally,with the development of sensor technology,target tracking algorithm based on mult-source data fusion has become the main research direction in the future tracking field.Based on the above two parts,a target tracking algorithm based on mult-source data fusion is proposed in this section.The algorithm collects the visable and infrared information of the target and fuses the data at the feature level and decision level to achieve accurate target tracking.After that,the algorithm is tested on the self-made mult-source target tracking data set,which proves that our algorithm can track the traget accurately in the multi-source data scenario.For the single visual data,through testing on the public data set UAV123,it is proved that the proposed algorithm has certain improvement in accuracy and robustness compared with other classical target tracking algorithms.