压缩感知W-HOG特征的运动手势跟踪  

Motion gesture tracking based on compressed sensing W-HOG features

作　　者：史东承[1] 倪康[1] 

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130012

出　　处：《智能系统学报》2016年第1期124-128,共5页CAAI Transactions on Intelligent Systems

摘　　要：运用单一特征或较简单的特征进行手势跟踪容易产生跟踪偏差。为提高手势跟踪的精确性,将压缩感知运动目标跟踪运用到手势跟踪中,并提取运动区域的HOG特征替换原来广义的类Harr特征进行压缩感知目标跟踪。为减少不同图像块HOG特征等权值串联对手势跟踪所产生的累积跟踪误差,需计算HOG特征权值,并将特征权值与HOG特征进行有效融合,形成W-HOG压缩特征。实验数据统计结果表明本文所改进的算法在精确度方面较CT算法提高近16%,较HOG-CT算法提高6%左右。且在复杂背景下能够较精确地检测出运动手势,提高了手势跟踪在光照变化、背景存在类肤色物体等情况下的鲁棒性,减少手势跟踪漂移的发生。The use of a single or simple feature for gesture tracking always induces tracking errors. To improve the accuracy of hand tracking,this study uses compressed-sensing motion tracking to track a hand and extracts HOG features in the movement area instead of original generalized class Harr features to track the target. At the same time,to reduce the accumulation of errors of gesture tracking generated by weight series,such as HOG features of different image blocks,the study calculates the HOG feature weight and effectively integrates the weight with HOG features to form W-HOG compression characteristics. The statistical experimental results show that the improved algorithm provided increased accuracy by approximately 16% compared with CT algorithm and approximately 6%compared with HOG-CT algorithm. Moreover,the algorithm can accurately detect the moving gesture in a complex background,improve the tracking robustness of gesture tracking in circumstances of illumination changes and background objects with a color similar to the skin,and reduce the occurrence of gesture tracking drifting.

关 键 词：压缩感知 Harr特征 HOG特征 手势跟踪 跟踪漂移 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]