Deformable attention object tracking network based on cross-correlation

Abstract

With the increasing popularity and widespread application of embedded devices, lightweight target tracking networks have emerged as a crucial research direction in the field of computer vision. In this paper, we propose a deformable attention object tracking network based on cross-correlation (DAM-LOTNet), which utilizes the lightweight MobileNetV3 network for efficient feature extraction. Aiming at the problem that lightweight networks cannot better capture target features and morphological changes, deformable attention modules are proposed to generate attention maps along the channel and spatial dimensions. Furthermore, a dual cross-correlation operation is proposed, aiming at fusing channel information and pixel-level information. Thus, the model obtains a more comprehensive feature representation and a more accurate similarity calculation capability while maintaining a lightweight framework. Finally, a feature enhancement network constructed by depthwise separable convolution was added to the classification and regression branches. The designed DAM-LOTNet has 0.552M parameters and 0.121G FLOPs. On the UAV123 dataset, DAM-LOTNet achieves state-of-the-art performance compared to other lightweight object tracking algorithms. Extensive experiments on visual tracking benchmarks (including VOT2018, VOT2019, OTB100, etc.) have demonstrated that DAM-LOTNet has extremely low model FLOPs and parameters compared to common deep learning-based trackers, and is essentially on par with these trackers in terms of performance.