基于稀疏表示的多尺度目标跟踪算法  

Multi-scale object tracking algorithm based on sparse representation

作　　者：徐文强[1] 高以成[1] 周煜坤[1] 

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094

出　　处：《计算机应用》2013年第A02期179-182,共4页journal of Computer Applications

摘　　要：由于目标的突发性移动、复杂的目标结构、遮挡、摄像机移动等原因,使得目标跟踪变得十分困难。结合粒子滤波和稀疏表示思想,提出一种鲁棒的多尺度目标跟踪算法。该算法首先经粒子滤波采样,并由随机森林分类器拒绝部分非目标粒子,再由稀疏重构算法计算出每个粒子属于目标的后验概率,从而刻画出当前目标状态。对比分析实验数据,可以看出该算法在准确性和正确性上都比其他跟踪算法优越,并且对目标姿态以及光照等变化具有较好的鲁棒性。As an important issue of intelligent surveillance, object tracking has been a hot topic in computer vision as well. However, problems such as object＇s abrupt motion, complex target structure, occlusion and camera＇s movement would definitely bring difficulties to academic study and engineering application. Hence, combining the theory of particle filtering and sparse representation, a new efficient algorithm for multi-scale object tracking was proposed. In order to approximate target＇s probability distribution, this algorithm acquired particles by the sampling theory of particle filtering at first, dispatched part of them next and computed each particle＇s post probability belonging to the target by signal reconstruction algorithm at last. Through analyzing and comparing the tacking results with other algorithms＇, it turns out that this algorithm is actually superior in both accuracy and correctness and also more insensitive to pose and illumination change.

关 键 词：目标跟踪 粒子滤波 稀疏表示 随机森林 多尺度 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]