RGB-D目标跟踪综述
A Survey of RGB-Depth Object Tracking
-
摘要: 随着深度学习的不断发展, 近几年已有许多基于深度学习的RGB目标跟踪算法被提出并具有较为显著的性能. 然而, 纯粹依靠可见光进行跟踪的算法, 在光照变化、严重遮挡等高难度场景下很难实现鲁棒跟踪. 因此, 一些研究者认为改变数据模态与质量是跟踪领域的新方案, 他们在RGB数据上融合进深度、红外、事件和语言等模态信息, 使跟踪器在特定场景下发挥更优越的性能. 本文将主要以RGB-D目标跟踪方法为主, 详细列举近几年可见光-深度的多模态跟踪方法, 并对各类不同方法的优缺点进行分析和比较. 其次, 本文也对主流的RGB-D跟踪数据集及其评价指标做介绍. 最后, 本文总结了RGB-D目标跟踪这一技术近来的发展趋势和挑战并对其未来发展提出展望.Abstract: With the development of deep learning, many RGB object tracking algorithms based on deep learning have been proposed with promising performance in recent years. However, algorithms that rely solely on visible light for tracking make it difficult to achieve robust tracking in difficult scenarios such as illumination variation and full occlusion. Therefore, some researchers believe that changing the data modality and quality is a new approach. They integrate the depth, infrared, events and semanteme modal information with RGB data to make the tracker perform better in specific scenarios. This paper will mainly describe the RGB-D tracking method. The RGB-Depth multimodal tracking methods in recent years are listed in detail, and the advantages and disadvantages of each method are analyzed and compared. Secondly, we will introduce the mainstream RGB-D tracking datasets and their evaluation indicators. Finally, we summarize the development trend and challenges of RGB-D target tracking technology and prospects its future development.