Distractor-aware visual tracking using hierarchical correlation filters adaptive selection

摘要

In recent years, the ensembled trackers composed of multi-level features from the pre-trained Convolutional Neural Network (CNN) have achieved top performance in visual tracking. However, due to the background clutters and the distractors in the search area, the tracker tends to drift towards an area that is similar to the target. In order to suppress interference of background and similar objects, we propose an effective Distractor-Aware Map (DAM), which can reduce the weights of the interference area in the multi-level features. Thus, the tracker can focus on the target to greatly eliminate the risk of drift. In addition, we build a Hierarchical Correlation Filters Model (HCFM) based on the multi-level convolutional features to track targets in parallel. To further improve the robustness of tracking, a novel Multi-Model Adaptive Selection (MAS) mechanism is presented. This mechanism can evaluate the confidence of the response map in HCFM to adaptively select the most reliable model. Finally, in order to appropriately update the model to adapt to appearance changes of the target, we propose an adaptive updating strategy for the updates of the DAM and HCFM. We perform comprehensive experiments on OTB-2013, OTB-2015 and Temple Color datasets and the experimental results show the superiority of our algorithm over other state-of-the-art approaches.