基于深度外观建模和时空隐变量推理的在线目标跟踪

Online object tracking is an important and hot issue in the fields of computer vision and video processing, which also has many real applications. However, existing tracking algorithms haven’t broken through the bottleneck of the tracking problem, and haven’t achieved satisfying performance on the large-scale benchmark dataset. ..This project will address the online object tracking problem from the perspective of deep learning and temporal-spatial latent variable inferencing. First, we will present “deep learning”-based appearance models, and further extend the theory of deep learning to the tracking filed. On the one hand, we will propose the concept of deep objectness, which aims to gradually learn the objectness of the tracked object by utilizing deep neural networks. On the other hand, we will exploit deep learning to mine effective local feature descriptors, which is combined with pooling algorithms and classifiers to develop tracking methods. In addition, we will investigate the temporal-spatial latent variable model to construct a unified object tracking framework, which aims to integrate the global and local deep appearance models, spatial configuration relationships within one frame, and the temporal consistency and inconsistency among different frames. Finally, based on the theory of probabilistic graphical model, we will design the related model simplification schemes and inference algorithms to effectively balance the tracking accuracy and speed. ..On the basis of above-mentioned research achievements, the results of this project can provide theoretical and practical supports for developing a robust and fast online object tracking algorithm, and can make significant progress on the large-scale evaluation.

在线目标跟踪是计算机视觉和视频处理领域的难点和热点问题，并有广泛的实际应用。然而现有算法尚未突破跟踪问题的研究瓶颈，未能在大尺度数据库上取得令人满意的结果。..本项目拟从深度学习和时空隐变量推理角度出发研究在线目标跟踪问题。首先，拟建立基于深度学习的外观模型，将深度学习理论推广到目标跟踪领域。一方面，提出深度目标性概念，利用深度神经网络逐步学习被跟踪目标的目标性描述；另一方面，利用深度学习挖掘有效的局部特征描述，并将其与池化算法、分类器相结合设计跟踪算法。其次，拟研究时空隐变量模型建立统一的目标跟踪框架，有效结合全局和局部深度外观模型、帧内空间配置关系以及帧间时间一致性和不一致性关系。最后，以概率图理论为基础，有针对性地提出模型简化方案和设计模型推理算法，有效平衡跟踪精度和速度。..拟通过上述研究，为实现鲁棒快速的在线目标跟踪算法提供理论和实践支撑，在大尺度测评中取得突破性进展。

在线目标跟踪是计算机视觉和视频处理领域的难点和热点问题，并有广泛的实际应用。本项目主要从目标外观建模和时空隐变量推理两个方面系统地研究了在线目标跟踪问题，提出了连续异常噪声（TIP2015)，最小软阈值均方（TCSVT2016），时空隐变量遮挡感知（TIP2016)，联合判决性和可靠性相关滤波（CVPR2018），局部空间敏感回归（CVPR2018，VOT2017公开组第一名），“演员-评论家”框架（ECCV2018）等有效的跟踪模型及算法，在跟踪精度和速度上均取得了较为突出的进展，在OTB和VOT标准测评中达到国际领先。在本项目资助下，课题组共发表论文17篇（SCI检索9篇），其中包括IEEE TIP、IEEE TCSVT、CVIU、PR等顶级期刊论文6篇，CVPR、ICCV、ECCV等顶级会议论文7篇。基于上述部分成果，课题组负责人获教育部自然科学二等奖1项（排名第2）。