Research on Lightweight Deep Correlation Filter Tracking Algorithm Based on Fuzzy Decision

Abstract

Deep correlation filter tracking method based on the fusion of correlation filter and deep convolutional neural network is one of the research hot topics in the field of visual object tracking. But how to choose an effective decision-making mechanism for implementing the online updating of feature network to fully adapt to the changes of target and environment in the tracking process is one of the key problems in the research of deep correlation filter tracking. It is obvious that the decision-making mechanism that only considers single factor can hardly meet the complex situation of the changes of target and environment. To address such an issue, this paper proposes a “Lightweight Deep Correlation Filter Tracking Algorithm Based on Fuzzy Decision”. In the process of tracking, the cosine similarity based on Siamese network and the SSIM similarity both for the predicting tracking targets in two consecutive frames are calculated in real time. And then these two kinds of the similarity are fused together into the final similarity of the predicting tracking targets by full use of the fuzzy decision, which is taken as the criterion to determine whether the feature network needs updating and whether the tracking fails. When the feature network needs to be updated, the model is updated online while the tracking continues. In the case of tracking failure, the target is searched again, and the tracking is resumed. We tested the model on the OTB data set, and the experiments show that the tracking model designed in this paper can improve the tracking accuracy under the conditions of real-time tracking.