基于多粒度感知SVM的复杂场景人车分类方法

针对复杂场景中的人车分类问题, 提出一种基于多粒度感知SVM (support vector machine)的复杂场景人车分类方法。该方法首先对视频场景进行运动区域分析, 结合角点检测方法提取运动区域视觉感知信息, 在时空域中采用Kalman滤波将感知信息进行关联推理, 去除噪声干扰。 再以运动区域质心点为中心, 构造目标的多粒度感知特征, 最后构造2级SVM分类器, 将目标多粒度感知特征向量集输入SVM分类器进行训练及分类, 得到人车分类结果输出。实验结果表明, 该方法取得了良好的分类效果, 人、车全天候平均分类正确率分别达到93.6%以上, 能有效避免光照、色彩、目标大小等变化导致的误分类问题, 适用于智能交通视频的人车分类应用。

Human and Vehicle Classification Method for Complex Scene Based on Multi-granularity Perception SVM

For solving the problem of human and vehicles classification in complex scene, a novel method based on multi-granularity perception SVM (support vector machine) is proposed. Firstly, motion regions of the video are detected and analyzed, and visual perception information is extracted by corner detection in motion regions. In order to reduce the noise interference, perception information is inferenced by Kalman filter in time-space domain. Furthermore, multi-granularity perception features of objects are constructed with the mass center of motion regions. Finally, a two-level SVM classifier is constructed, and classification results are obtained by training and classifying on the SVM classifier with multi-granularity perception features vector set. The results of experimentation show that the proposed method is good. The average correct rate of all-day classification between human and vehicles is up to 93.6% separately, and it is valid to avoid the influence of illumination, colors and object’s size variation. It is suitable for intelligent traffic system.