联合模板先验概率和稀疏表示的目标跟踪  

Target tracking based on a priori probability of template and sparse representation

作　　者：田猛[1,2] 路成 周健[1,2] 施汉琴[1] 陶亮[1] Tian Meng Lu Cheng Zhou Jian Shi Hanqin Tao Liang(Key Laboratory of Intelligent Computingand Signal Processing of Ministry of Education, Anhui University, Hefei 230039, China Institute of Media Computing, Anhui University, Hefei 230601, China)

机构地区：[1]安徽大学计算智能与信号处理教育部重点实验室,合肥230039 [2]安徽大学媒体计算研究所,合肥230601

出　　处：《中国图象图形学报》2016年第11期1455-1463,共9页Journal of Image and Graphics

基　　金：国家自然科学基金项目(61372137;61301295);安徽省自然科学基金项目(1308085QF100;1408085MF113);安徽大学博士科研启动基金项目~~

摘　　要：目的虽然基于稀疏表示的目标跟踪方法表现出了良好的跟踪效果,但仍然无法彻底解决噪声、旋转、遮挡、运动模糊、光照和姿态变化等复杂背景下的目标跟踪问题。针对遮挡、旋转、姿态变化和运动模糊问题,提出一种在粒子滤波框架内,基于稀疏表示和先验概率相结合的目标跟踪方法。方法通过先验概率衡量目标模板的重要性,并将其引入到正则化模型中,作为模板更新的主要依据,从而获得一种新的候选目标稀疏表示模型。结果在多个测试视频序列上,与多种流行算法相比,该算法可以达到更好的跟踪性能。在5个经典测试视频下的平均中心误差为6.77像素,平均跟踪成功率为97%,均优于其他算法。结论实验结果表明,在各种含有遮挡、旋转、姿态变化和运动模糊的视频中,该算法可以稳定可靠地跟踪目标,适用于视频监控复杂场景下的目标跟踪。Objective Although sparse representation-based tracking approaches show good performance, they usually fail to observe the object motion because of noise, rotation, partial occlusion, motion blur, and illumination or pose variation. This study proposes an algorithm based on sparse representation and a priori probability of object template to improve track- ing capability under partial occlusion, rotation, pose change, and motion blur conditions. An L1 tracker is also developed, which runs in real time and possesses better robustness than other L1 trackers. Method The importance of the target tem- plate is measured by a priori probability and is considered in the proposed algorithm when updating the object template. Combined with the regularization model, a novel sparse representation model of the object is presented. Based on the pro- posed target appearance model, an effective template update scheme is designed by adjusting the weighs of the target tem- plates. The tracking particles of the current frame are generated by the last tracking result according to the Gaussian distri-bution. The sparse representation of each particle to the template subspace is obtained by solving the Ll-regularized least square problem, and a target searching strategy is employed to find the particle that well matches the template as the track- ing result. The particle filter is then used to propagate sample distribution in the next tracking frame. Result Compared with existing popular tracking algorithms, the proposed algorithm can achieve better tracking performance in diverse test video datasets. Experimental results demonstrate that the proposed algorithm can handle appearance changes, such as pose varia- tion, rotation, illumination, motion blur, and occlusion. Compared with state-of-the-art methods, the proposed algorithm performs well and obtains the best results in the sequences of FaceOccl, Girl, BlurBody, and Singerl, with average center location errors of 6. 8, 4. 0, 16. 3, and 3.5 pixels, respectively. The average trackin

关 键 词：目标跟踪 稀疏表示 先验概率 粒子滤波 模板更新 正则化模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]