Siamese Network with Feature Fusion for Visual Object Tracking

Abstract

Due to the limitations of learning richness of the model, it is challenging for the trackers to complete tracking precisely under the circumstances of large appearance variance of the objects. Therefore, the improved version of Siamese-RPN are proposed to solve the problem and get a higher accuracy rate. In this paper, multi-level features extraction from layer-wise aggregation are applied to address the limitations of the robustness of Siamese-RPN in some special cases such as large occlusion and illumination variation. By utilizing the feature information in earlier layer and semantic information in latter layer of Siamese-RPN, the performance of CNN can be improved and a complete feature map of the objects can be established. With the improved version of Siamese-RPN, the tracker could have performance gain in its accuracy rate. Experiments show that when we use the original SiamRPN model to test the dataset, the success score is 0.6053. However, we find that the success score of a weighted sum of the features extracted from 0.4 times layer2 and layer5 is 0.6156. The dataset in this paper is based on OTB100K. Experiments show the high efficiency and accuracy rate of our method.