Research Of A Spatial-Aware Tracking Algorithm Based On Adaptive Feature Combination

Visual tracking is an important task in computer vision.It can be defined as estimating the trajectory and scale change of a object in the image plane.Visual object tracking is very challenging,because the background of the object is very complex and changes frequently,even the object itself is changing constantly.Object tracking is an on-line task,which requires high tracking accuracy and speed.At present,most of t he traditional algorithms can't meet the accuracy requirements,and deep learning methods can't meet the real-time requirements.In this paper,a spatial-aware correlation filter based on adaptive feature combination is proposed,which obtains high tracking accuracy while meeting the premise of real-time.The main contributions of this paper are divided into the following two parts:(1)By analyzing the characteristics of the Siamese network-based tracking algorithm and correlation filtering tracking algorithm,a spatial-aware correlation filter is proposed.The tracking algorithm based on correlation filter uses cyclic sampling to generate training sample which contaminates positive samples with background information(boundary effect),which limits the training and detection area of correlation filter,moreover,using cosine window to suppress boundary effect will further reduce the effective detection area,which makes correlation filter unable to cope with fast moving scenes.To solve this problem,a strategy of using Siamese networks to expand the search area of correlation filter tra cker is proposed,and the tracking results of Siamese networks are taken as a clue.On the one hand,the Siamese network always uses the first frame as a template to cope with model drift;on the other hand,the search area of the Siamese network is not limited,which can provide a reliable spatial clue when fast motion occurs.Finally,a more reliable space is selected through the proposed spatial selection mechanism.The proposed spatia l-aware correlation filter achieves 0.865 distance accuracy,0.656 overlap accuracy and 32 frame rate on OTB-2015.The tracking accuracy exceeds that of all trackers using manual features.(2)The proposed spatial-aware correlation filter integrates multiple correlation filters using different feature combinations and a Siamese network tracker using a multi-cue tracking framework.This multi-cue tracking framework employs exhaustive feature combinations and trains the corresponding correlation filter for each feature combination,although this approach can achieve good tracking accuracy.However,a large number of filters need to be trained will increase the time complexity of the algorithm.To solve this problem,an adaptive feature combination mechanism is proposed to improve the multi-cue tracking framework.It can select some appropriate correlation filters adaptively in complex dynamic tracking scenarios,which can reduce the number of correlation filters to be trained to improve tracking accuracy.The proposed spatial-aware correlation filter based on adaptive feature combination achieves 0.859 distance accuracy,0.654 overlap accuracy and 47 frame rate on OTB-2015.