基于稀疏表示的在线视觉跟踪

Object tracking is one popular and challenging topic in the field of computer vision and pattern recognition. The goal of object tracking is to locate the target accurately in the existence of occlusion, rotation, illumination change, abrupt motion, shape deformation, cluttered background, and other inferences. Recently, sparse representation theory becomes more and more active in the research of object tracking, and it is still a theory in process. The project intends to use sparse representation theory for the research of robust tracking method that can handle the difficulties of object rotation, pose variation, illumination change, occlusion and cluttered background. The objectives of the project are presented as follows: We explore the joint optimization of the discriminative dictionary and classifier in the object tracking problem. By learning a superior dictionary and a SVM classifier simultaneously, generative and discriminative models are incorporated organically. We also aim to combine structural information with sparse representation by constructing a structural local sparse appearance model of the target. Meanwhile, in order to make better use of prior information, we explore a variety of metrics measuring the reconstruction error in sparse representation and propose a learning based weighted distance metric. To speed up the tracking process, we plan to substitute the traditional particle filter by solving a collaborative model which incorporates Lucas-Kanade method and sparse representation to obtain affine parameters.

目标跟踪是计算机视觉、模式识别领域一个非常热门而且极具挑战性的课题，其目的是能够克服遮挡、旋转、光照、突然运动、变形、复杂背景等因素的干扰，准确的确定目标的位置等参数。近来，稀疏表示理论在目标跟踪中的研究非常活跃，而且其理论在不断完善的过程中。本项目拟针对跟踪问题，围绕稀疏表示理论，研究能够处理目标旋转，姿态，光照，遮挡，以及复杂背景等鲁棒的跟踪算法。本课题的具体目标是探讨目标跟踪中的判决性稀疏字典学习与分类器的联合优化，同时学习出较好的字典和SVM分类器，从而实现生成模型和判决模型的有机融合；探讨基于局部模型的目标结构特征模型，从而将结构化信息与稀疏表示相结合；探讨各种可能的稀疏表示重构误差描述方法，建立一种基于学习的权重的重构误差描述方法，从而能更好的利用一些先验信息；最后探讨Lucas-kanade方法与稀疏表示目标函数联合优化，求解仿射变换参数，替代撒粒子的方法，加快跟踪的速度。

随着交通智能监控、安防系统、机器人等硬件系统的广泛应用，产生了海量视频数据，视频目标跟踪作为理解视频数据的关键任务，在人工智能领域发展过程中成为急需克服的挑战难点。为了解决目标跟踪任务中难以应付的遮挡、旋转、光照变化、形变等困难，本项目主要从稀疏表示方法与其他联合优化的思路，提出了基于稀疏哈希编码模型、判决性稀疏字典模型、结构化局部稀疏模型、自适应权重学习模型、以及距离度量等视频目标跟踪算法。在本项目资助下，课题组共发表论文47篇，其中在顶级期刊TIP/IJCV共发表论文16篇，在顶级会议CVPR/ICCV共发表论文11篇。基于上述部分成果，项目负责人致力于深入研究复杂场景下显著性目标检测与跟踪，获得教育部自然科学二等奖。