Heng Fan1 and Jinhai Xiang2
1College of Engineering, Huazhong Agricultural University, Wuhan, China
2College of Informatics, Huazhong Agricultural University, Wuhan, China
Dictionary learning for sparse representation has been increasingly applied to object tracking, however the existing methods only utilize one modality of the object to learn a single dictionary. In this paper, we propose a robust tracking method based on multitask joint dictionary learning (MJDL). Through extracting different features of the target, multiple linear sparse representations are obtained. Each sparse representation can be learned by a corresponding dictionary. Instead of separately learning the multiple dictionaries, we adopt a multitask learning approach to learn the multiple linear sparse representations, which provide additional useful information to the classification problem. Because different tasks may favor different sparse representation coefficients, yet the joint sparsity may enforce the robustness in coefficient estimation. During tracking, a classifier is constructed based on a joint linear representation, and the candidate with the smallest joint decision error is selected to be the tracked object. Figure 1 ilustrates the proposed method. In addition, reliable tracking results and augmented training samples are accumulated into two sets to update the dictionaries for classification, which helps our tracker adapt to the fast time-varying object appearance. Figure 2 demonstrates the update strategy. Both qualitative and quantitative evaluations on CVPR2013 visual tracking benchmark demonstrate that our method performs favorably against state-of-the-art trackers.
Figure 1. The proposed tracking method. Firstly we sample the candidates in frame (t+1) within Bayesian framework and then extract multiple features for each candidate. After obtaining the features, a classifier based on multitask joint dictionary learning is utilized to compute the joint decision measure of each candidate. The candidate with the maximum joint decision measure is selected to be the tracking result.
Figure 2. The illustration of the update process.
Figure 3. The precision plots and success plots of OPE for the proposed tracker and the top 10 trackers in benchmark. The performance score for each tracker is shown in the legend. The performance score of precession plot is at error threshold of 20 pixels while the performance score of success plot is the AUC value. Best viewed on color display.
Figure 4. Sampled results of MJDL on the benchmark. For the frame pair of each image sequence, the left shows the first frame with the bounding box of the target while the right one shows the beginning frame that suffers from severe drift. If the severe drift does not happen in that image sequence, the right one shows the last frame with the tracking result. Best viewed on color and high-resolution display.
Email me for the source code if you are interested.
Heng Fan and Jinhai Xiang. Robust Visual Tracking with Multitask Joint Dictionary Learning. IEEE Transactions on Circuits and Systems for Video and Technology (TCSVT), vol. 27, no. 5, pp. 1018-1030, 2017.