Learning rich feature representation and aggregation for accurate visual tracking

Visual tracking is a key component of computer vision and has a wide range of practical applications. Recently, the tracking-by-segmentation framework has been widely applied in visual tracking due to its astonishing performance on accuracy. It attempts to learn from the framework of video object segmentation to realize accurate tracking. Although segmentation-based trackers are effective for target scale estimation, the segmentation network makes the trackers have high requirements for the extracted target features due to the need for pixel-level segmentation. Therefore, in this article, we propose a novel feature representation and aggregation network and introduce it into the tracking-by-segmentation framework to extract and integrate rich features for accurate and robust segmentation tracking. To be specific, firstly, the proposed approach models three complementary feature representations, including contextual semantic, local position, and structural patch feature representations, through cross-attention, cross-correlation and dilated involution mechanisms respectively. Secondly, these features are fused by a simple feature aggregation network. Thirdly, the fusion features are fed into the segmentation network to obtain accurate target state estimation. In addition, to adapt the segmentation network to the appearance changes and partial occlusion, we introduce a template update strategy and a bounding box refinement module for robust segmentation and tracking. The extensive experimental results on twelve challenging tracking benchmarks show that the proposed tracker outperforms most of the state-of-the-art trackers and achieves very promising tracking performance on the OTB100 and VOT2018 benchmarks.