Authors:
JunChang Zhang
1
;
ChenYang Xia
2
and
JinJin Wan
3
Affiliations:
1
Northwestern Polytechnical University and Key Laboratory of Photoelectric Control Technology, China
;
2
Northwestern Polytechnical University, China
;
3
Key Laboratory of Photoelectric Control Technology, China
Keyword(s):
Correlation filter, Channel reliability, Tracker, Validator, Siamese convolutional neural network
Abstract:
For most algorithms, the problem of Tracking target performance degradation in the case of fast moving, illumination changes, target deformation, occlusion, out-of-plane rotation, low-resolution images, etc. This paper proposes a tracking verification algorithm based on channel reliability. The tracker part of the algorithm is tracked by the method of correlation filter based on channel reliability. By calculating the reliability weight of each feature channel of the input correlation filter, and multiplying the weight by the response of the corresponding channel to obtain the final response, so that the target positioning will be more accurate. The validator part uses the Siamese dual input network in the deep learning convolutional neural network. Every few frames, the verifier will verify the results of the tracker part of the algorithm. If the reliability is verified, the tracking result will not be modified. Otherwise, the validator will re-detect the target location and verify
the reliability through the Siamese dual-input network. The tracker will regard this location as the new location of our target continues to be tracked, making target tracking more durable and robust. The experimental evaluation of the OTB13 video sequence proves that the proposed algorithm has good adaptability to target fast motion, illumination change, target deformation, occlusion, and out-of-plane rotation, and has good robustness.
(More)