Research On Single Target Tracking Algorithm In Complex Scenes

The core idea of target tracking is that the tracking model estimates the target position in subsequent images based on the target bounding box and position information given by the first frame image of the video sequence.In the past few decades,the research on target tracking algorithms has received extensive attention and rapid development.However,how to achieve accurate and robust tracking in complex scenes is still a difficult point.Although the introduction of correlation filtering theory improves the accuracy and processing speed of target tracking,the tracking results are easily affected to fail by the noise in images.The tracking model based on the discriminative correlation filter(DCF)introduces background information for training to significantly improve the robustness of the tracker,while it is still susceptible to boundary effect and temperal degradation.In order to improve the accuracy and robustness of the target tracking algorithm in complex environment,this paper proposes two improvements of the tracking algorithm based on the discriminative correlation filter.The main results are as follows:(1)Aiming at the problem that traditional tracking algorithms based on correlation filter are prone to tracking failures in complex scenes,this paper proposes an aberrance learning via time-driven correlation filter(ALTCF).Specifically,the model introduces a temporal regularization term that can suppress aberrance data,and uses the features in the temporal domain to ensure that the target tracking model maintains the temporal similarity.At the same time,the model alleviates the negative impact of aberrance data due to the noise,and suppresses the over-fitting phenomenon of model training.Finally,the model combines the similarity of filter response and temporal features to search for target positions in the images,which improves the robustness of correlation filter target tracking and alleviate the temporal degradation of the filter.(2)As the tracking algorithm based on discriminant correlation filter will be affected by spatial boundary effect and temporal degradation of filter in complex scenes,this paper proposes a target tracking model based on the learning temporal consistency correlation filter(LTCF).In order to improve the stability of the tracking model,the model adds the second-order information,that is,the gradient information of the filter for training and updating the correlation filter.Therefore,the LTCF tracking model combines the stability of gradient feature with the temporal consistency,and uses the correlation between features to improve the accuracy of model target tracking.In order to explain the performance of the tracking algorithms proposed in this paper,this paper conducts comparative experiments based on different data sets.The tracking models involved in the comparative experiments include 30 advanced algorithms.By comparing and analyzing the tracking results,the tracking algorithm proposed in this paper shows good tracking effects in complex scenes and can ensure the accuracy and robustness of target tracking.