Abstract: The occlusion problem in the visual object tracking task is one of the most challenging scene attributes. The study of effective anti-occlusion model learning schemes is of great significance for building long-term robust tracking models that adapt to complex scenes. First, it analyzes essentially why occlusion affects the tracking performance, and then advanced tracking algorithms with better anti-occlusion performance are taken as the research object, it systematically analyzes the effective anti-occlusion mechanism in model learning, and compares its effectiveness in improving long-term and short-term occlusion problems. The analysis includes training sample quality improvement strategies to provide sufficient discriminative information for the model that contains hard negative sample mining, effective sample management, and occlusion of hard positive samples generation; Passive and stable learning methods based on time consistency learning, adaptive appearance learning, and multiple active learning strategies suitable for tracking tasks builds robust models that can resist scene interference, target deformation and other factors, such as domain attributes, target perception, interference perception, feature fusion, etc.; The update strategy balances the adaptability and stability of the model tracking object with change state online such as manual confidence evaluation, adaptive decision-making, and time series memory storage, adaptive estimation template. Then, by comparing the performance of the representative tracking algorithm in occlusion and background clutter, out of view, in and out of the plane rotation, and deformation, the effectiveness of each strategy against occlusion is analyzed in detail, and it is pointed out that compared to the update strategy, the data processing and learning strategy design is more helpful to improve the anti-occlusion performance; Meanwhile, the applicability of each strategy to long-term occlusion, background clutter, out-of-view, and other attributes is analyzed, as well as the strategies applicable to multiple types of complex scenes. Finally, the effective anti-occlusion strategies are summarized, and research direction of backbone network replacement and migration scene understanding, motion law, and other prior information to the tracking task is proposed.

Key words: object tracking, occlusion, high-quality training sample set, time consistency, confidence

摘要： 视觉目标跟踪任务中的遮挡问题是最具挑战的场景属性之一，研究有效的抗遮挡模型学习方案，对构建适应复杂场景的长期鲁棒跟踪模型具有重要意义。剖析了遮挡影响跟踪性能的本质原因，以抗遮挡性能较好的先进跟踪算法为研究对象，系统分析了模型学习中有效抗遮挡机制，并对其改善长短期遮挡问题的有效性进行比较分析，包括以硬负样本挖掘、有效样本管理、类遮挡硬正样本生成的训练样本提质策略，提供模型充足判别信息；以时间一致性学习、自适应外观学习的被动稳定学习方式和基于多域属性、目标感知、干扰感知、特征融合等适用跟踪任务的主动学习策略，构建可抵抗场景干扰、目标形变等因素可适用跟踪的鲁棒模型；以手工置信度评估、自适应决策、时序记忆库、自适应估计模板的更新策略，平衡模型在线跟踪状态变化目标的适应性与稳定性。通过对代表跟踪算法在遮挡及背景杂乱、出视野、平面内外旋转、形变场景下的性能比较，详尽分析了各策略抗遮挡有效性，指出相比更新策略，数据处理、学习策略设计更有利于提高抗遮挡性能；同时分析了各策略对长期遮挡、背景杂乱、出视野等属性的适用性及适用多类复杂场景的策略。总结了有效抗遮挡策略，提出骨干网替换及迁移场景理解、运动规律等先验信息至跟踪任务的研究方向。

关键词: 目标跟踪, 遮挡, 高质训练样本集, 时间一致性, 置信度