特征压缩在线距离度量学习跟踪  

On-line distance metric learning tracking using compressed feature

作　　者：陈东成[1,2] 朱明[1] 贺柏根[1] 杨文波[1,2] 

机构地区：[1]中国科学院长春光学精密机械与物理研究所、中国科学院航空光学成像与测量重点实验室,吉林长春130033 [2]中国科学院大学,北京100039

出　　处：《光电子．激光》2014年第8期1592-1597,共6页Journal of Optoelectronics·Laser

基　　金：国家自然科学基金(61203242);中科院航空光学成像与测量重点实验室开放基金(Y2HC1SR121)资助项目

摘　　要：为提高在线学习目标跟踪的实时性和准确率,结合压缩感知理论,提出一种将距离度量学习(DML)运用到目标跟踪的算法。首先,根据所选定的目标位置分别提取目标和背景样本集,运用随机投影理论对样本的Harr-like特征进行压缩;然后,用压缩后的低维特征向量集训练度量矩阵;最后,在新的一帧中抽取目标和背景的样本,用训练得到的度量矩阵计算已知目标和样本间的Mahalanobis距离,距离最小的样本的位置就是所要跟踪的目标的位置。对不同视频序列的测试结果表明,用压缩特征表示目标,使特征计算的计算量压缩到原来的1/4,减少了特征计算的时间;用训练后的度量矩阵计算目标位置,即跟踪器能够根据目标的不断变化自适应调整参数,提高了跟踪的准确率。To improve the precision and real-time quality of on-line learning object tracking, combined with the compress sensing theory, an algorithm using distance metric learning is proposed. First, target samples and background samples around the selected target are sampled. The Harr-like feature vectors are compressed using the random projection theory. Then,the distance metric is trained using the com- pressed feature vectors. Finally, the Mahalanobis distance between the samples in the new coming frame and the known target is calculated. The location of the sample closest to the known target is the location wanted. Experiments on variant videos show that the caculating load of the compressed features is 3/4 less than that using the uncompressed ones. Calculating the location of target using the trained distance metric makes the tracking precision higher.

关 键 词：压缩感知 目标跟踪 距离度量 MAHALANOBIS距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]