结合场景运动模式的有向加权AdaBoost目标检测

针对现有检测算法对场景先验信息和群体运动规律考虑甚少这一局限性,提出一种结合场景运动模式的有向加权AdaBoost目标检测算法.该算法首先建立了一种基于速率加权方向直方图矩阵的场景运动模式模型,并在此基础上通过稀疏光流投票方法获取场景的运动模式信息.同时,针对该模型提出一种有向加权AdaBoost检测算法,通过建立多个有向AdaBoost分类过程,并利用局部区域的运动模式对分类过程加权,最终实现运动目标检测.通过交叉验证分类实验和视频检测实验验证,该算法在相同假阳性率条件下的查准率较传统AdaBoost检测器的高出约10%,充分验证了算法的有效性和优越性.

Directed AdaBoost weighting classifier based on the scene motion pattern for object detection

The regular behavior reflected by the interaction of spatial layout and interior elements of the scene with the moving object is named the scene motion pattern. In this paper, a novel approach of the object detection combining scene motion pattern with the directed AdaBoost weighting method is proposed to remedy the limitation of the existing detection algorithms that give little consideration of the object motion regularity and the priori information about the scene. For this reason, a model of the scene motion pattern described by a matrix of speed-weighting directed histogram is created, and the information about the scene motion pattern is acquired by the voting of the sparse optical flows on that basis. Meanwhile, a directed AdaBoost weighting detection algorithm is developed correspondingly. A set of directed AdaBoost classifiers which are then weighted according to the motion pattern of the region are established in the algorithm. According to the specially designed cross-validation classification experiments and video tests, the precision rates of the algorithm are about 10％ higher than that of standard AdaBoost detector under the same condition of the false positive rate, which proves the effectiveness and the advancements of the proposed approach.