基于判别性局部联合稀疏模型的多任务跟踪

目标表观建模是基于稀疏表示的跟踪方法的研究重点, 针对这一问题, 提出一种基于判别性局部联合稀疏表示的目标表观模型, 并在粒子滤波框架下提出一种基于该模型的多任务跟踪方法(Discriminative local joint sparse appearance model based multitask tracking method, DLJSM).该模型为目标区域内的局部图像分别构建具有判别性的字典, 从而将判别信息引入到局部稀疏模型中, 并对所有局部图像进行联合稀疏编码以增强结构性.在跟踪过程中, 首先对目标表观建立上述模型; 其次根据目标表观变化的连续性对采样粒子进行初始筛选以提高算法的效率; 然后求解剩余候选目标状态的联合稀疏编码, 并定义相似性函数衡量候选状态与目标模型之间的相似性; 最后根据最大后验概率估计目标当前的状态.此外, 为了避免模型频繁更新而引入累积误差, 本文采用每5帧判断一次的方法, 并在更新时保留首帧信息以减少模型漂移.实验测试结果表明DLJSM方法在目标表观发生巨大变化的情况下仍然能够稳定准确地跟踪目标, 与当前最流行的13种跟踪方法的对比结果验证了DLJSM方法的高效性.

Tracking via Multitask Discriminative Local Joint Sparse Appearance Model

Appearance modeling is the research focus in tracking method based on sparse representation. In this paper, a discriminative local joint sparse appearance model based multitask tracking method (DLJSM) is proposed within particle filter framework. The proposed model builds a discriminative dictionary for each image patch within the object-region in order to introduce the discriminative information into the local sparse model, and enhances the structure feature via joint sparse representation. During tracking, the target appearance is modeled firstly. Then the sampling particles are pre-selected according to the target appearance's consecutive changes characteristic to improve efficiency of the algorithm. Next, joint sparse representations of all the candidates are solved jointly. Furthermore, a function is defined to measure the similarities between candidates and the target model. Lastly, the target state is estimated by the maximum posterior probability. Besides, update is judged every five frames to avoid the accumulative error caused by frequent update and the target information in the first frame is reserved to alleviate drifting. Test results show that the proposed DLJSM tracker can maintain a stable and accurate tracking when the target appearance undergoes huge variations. Comparison results on challenging benchmark image sequences show that the DLJSM method out performs 13 other state-of-the-art algorithms.