Correlation-filter-based trackers have shown favorable accuracy and efficiency in visual tracking. However, most of these trackers are prone to drift in cases of heavy occlusions and temporal tracking failures because they only maintain the short-term memory of target appearance via a highly adaptive update mode. In this paper, we propose a reliable visual tracking method based on a dual-memory selection (DMS) model to alleviate tracking drift. Considering that long-term memory is robust to heavy occlusions while short-term memory performs well in rapid appearance changes, the proposed DMS model combines these two memory patterns of the target appearance and adaptively selects a reliable memory pattern to handle the current tracking challenges via a memory selector. For each memory pattern, a memory tracker is established based on discriminative correlation filters. The short-term tracker aggressively updates the target model to capture recent appearance changes via a linear interpolation update model, while the long-term tracker conservatively updates the target model to maintain historical appearance characteristics with a memory-improved update model and a dynamic learning rate. Furthermore, a novel memory evaluation criterion (MEC) is developed to evaluate the reliability of each tracker for memory selection. From credibility and discriminability measurements considering the temporal context, the memory tracker with the highest reliability score is selected to determine the target location in each frame. Extensive experiments on public benchmark datasets demonstrate that the proposed tracking method performs favorably compared to multiple recent state-of-the-art methods.

基于相关滤波器的跟踪器在视觉跟踪中显示出良好的准确性和效率。但是，大多数这些跟踪器在重度遮挡和时间跟踪失败的情况下易于漂移，因为它们仅通过高度自适应的更新模式来维持目标外观的短期记忆。在本文中，我们提出了一种基于双记忆选择（DMS）模型的可靠视觉跟踪方法，以减轻跟踪漂移。考虑到长期记忆对重度遮挡具有鲁棒性，而短期记忆在快速的外观变化中表现良好，因此建议的DMS模型结合了目标外观的这两种记忆模式，并自适应地选择了一种可靠的记忆模式来处理当前的跟踪挑战内存选择器。对于每种记忆模式，基于判别相关过滤器建立内存跟踪器。短期跟踪器通过线性插值更新模型积极地更新目标模型以捕获最近的外观变化，而长期跟踪器通过内存改进的更新模型和动态学习率来保守地更新目标模型以保持历史外观特征。 。此外，开发了一种新颖的内存评估标准（MEC）来评估每个跟踪器进行内存选择的可靠性。从考虑时间上下文的可信度和可分辨性测量中，选择具有最高可靠性得分的内存跟踪器以确定每个帧中的目标位置。