Triplet Network with Multi-level Feature Fusion for Object Tracking
Yang Cao, Bo Wan, Quan Wang, Fei Cheng · 2020
In recent years, Siamese network-based trackers have received increasing interest because of the balanced accuracy and speed. However, these tracking methods only extract the high-level features as target representations and merely utilize the first frame as the exemplar branch, which are less discriminative to distinguish similar distractors and are vulnerable to background clutter. To address these issues, we propose a novel Triplet network with a multi-level feature fusion structure (TripMFF) to leverage the merits of Triplet network and fuse multi-layer features for robust object tracking. Firstly, the Triplet network is adopted as our backbone architecture to take full advantage of the correlation among frames in the video sequence and make the exemplar branches have more useful information. Secondly, in order to capture more abundant features, a multilevel feature fusion structure is put forward to combine the low-level fine-grained and high-level abstract information, which improves the discriminative capability and stability of the tracker. Experimental results on tracking benchmarks prove that our proposed method achieves competitive performance and real-time tracking speed compared with other state-of-the-art trackers Contribution-We propose a tracker TripMFF which adopts the Triplet network as backbone architecture and employs a multi-level feature fusion structure for robust object tracking.