Research On Feature Extraction For Visual Object Tracking

In recent years,visual target tracking technology has become one of the research hotspots in the field of computer vision.Feature extraction is a crucial stage in video target tracking.The complex application scenarios,such as light variation,target shading and blur,have brought great challenges to feature extraction.In order to further promote the effectiveness of the target tracking algorithm,the paper starts with the following tasks from the feature extraction stage:(1)The real-time advantage of the correlation filter tracking framework is gradually lost as the complexity of feature extraction is increased.In order to solve this problem,this article is based on correlation filtering tracking framework,this paper proposes a edge feature,based on the two levels of filtering and confidence discriminant method based on the mean maximum response sequence to update the model to reduce the drift model.Compared with the latest tracking algorithms with better performance,the experiments show that this simple and effective feature further promotes the AUC and Precision indexes of the KCF tracking algorithm.(2)Multi-feature fusion has become the mainstream method for feature extraction in target tracking.In order to improve the tracking accuracy,based on the correlation filtering tracking framework,this paper proposes a multiple features fusion feature extraction method.The feature after two-stage filtering and HOG were used for fusion.The correlation filters are updated with the theory of back propagation algorithm.Compared with the latest multiple tracking algorithms with better performance,the fusion feature further promoted the AUC and Precision indexes of Staple tracking algorithm.(3)The initialization and updating of the convolution filter in the online deep learning feature directly affects the accuracy of target tracking.In order to avoid the artificial interference of initial convolution filter,this paper puts forward an improved online deep feature extraction method,the algorithm of convolution filter based on k-means++ is initialized,and by using the error back propagation algorithm of convolution filter updates.Compared with the latest tracking algorithms with better performance,the results show that the improved online deep feature further promote the AUC and Precision indexes of CNT tracking algorithms.