College of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China.
Big Data Institute, East University of Heilongjiang, Harbin 150066, China.
Sensors (Basel). 2020 Sep 19;20(18):5374. doi: 10.3390/s20185374.
Most trackers focus solely on robustness and accuracy. Visual tracking, however, is a long-term problem with a high time limitation. A tracker that is robust, accurate, with long-term sustainability and real-time processing, is of high research value and practical significance. In this paper, we comprehensively consider these requirements in order to propose a new, state-of-the-art tracker with an excellent performance. EfficientNet-B0 is adopted for the first time via neural architecture search technology as the backbone network for the tracking task. This improves the network feature extraction ability and significantly reduces the number of parameters required for the tracker backbone network. In addition, maximal Distance Intersection-over-Union is set as the target estimation method, enhancing network stability and increasing the offline training convergence rate. Channel and spatial dual attention mechanisms are employed in the target classification module to improve the discrimination of the trackers. Furthermore, the conjugate gradient optimization strategy increases the speed of the online learning target classification module. A two-stage search method combined with a screening module is proposed to enable the tracker to cope with sudden target movement and reappearance following a brief disappearance. Our proposed method has an obvious speed advantage compared with pure global searching and achieves an optimal performance on OTB2015, VOT2016, VOT2018-LT, UAV-123 and LaSOT while running at over 50 FPS.
大多数跟踪器仅专注于鲁棒性和准确性。然而，视觉跟踪是一个长期存在的问题，时间限制很高。一个具有鲁棒性、准确性、长期可持续性和实时处理能力的跟踪器具有很高的研究价值和现实意义。在本文中，我们综合考虑了这些要求，提出了一种具有优异性能的新型、最先进的跟踪器。首次通过神经架构搜索技术采用 EfficientNet-B0 作为跟踪任务的骨干网络，这提高了网络的特征提取能力，并显著减少了跟踪器骨干网络所需的参数数量。此外，最大交并比被设置为目标估计方法，增强了网络稳定性，提高了离线训练的收敛速度。在目标分类模块中采用通道和空间双重注意力机制，提高了跟踪器的辨别能力。此外，共轭梯度优化策略提高了在线学习目标分类模块的速度。提出了一种两阶段搜索方法和筛选模块相结合的方法，使跟踪器能够应对目标突然运动和短暂消失后的再次出现。与纯全局搜索相比，我们提出的方法具有明显的速度优势，在 OTB2015、VOT2016、VOT2018-LT、UAV-123 和 LaSOT 上的性能达到最优，同时运行速度超过 50 FPS。
