基于矢量量化的动态场景中运动车辆的检测  

Moving vehicle detection in dynamical scene based on vector quantization

作　　者：高磊[1] 牛润新[2] 刘伟[2] 戚利民 梁华为[2] 

机构地区：[1]中国科学技术大学工程科学学院,合肥230022 [2]中国科学院合肥物质科学研究院,合肥230022 [3]合肥晟泰克汽车电子有限公司,合肥230022

出　　处：《电子测量技术》2014年第10期66-70,共5页Electronic Measurement Technology

摘　　要：汽车为了提高动态道路场景中的运动车辆的检测的准确性,提出在Harris角点检测与金字塔Lucas-Kanade光流的基础上,以角点光流矢量的方向、幅值和角点的坐标值作为输入特征,引入矢量量化的方法对图像的光流场进行聚类,并结合了欧式距离和相似系数作为相似性测度,再通过计算每个类别中角点分布的方差剔除误匹配的光流,并依靠方差区分背景类和车辆类,从而提取出运动的车辆。实验结果表明,此算法能有效的剔除误匹配的光流,准确的从动态场景中提取出运动的车辆,并同时适合单个和多个车辆的检测。In order to improve the accuracy of moving vehicle detection in dynamical scene, a new method which was on the basis of the Harris corners and pyramidal Lucas-Kanade optical flow was presented. Firstly, the direction and magnitude of the optical flow were taken as inputs, and Vector Quantization was adopted for clustering them to several classifications, and then computed variance of feature point distribution of each classification. Finally, the mismatched optical flow class could be eliminated by the value of variance, and by which the vehicle class could also be distinguished from background. Eventually, the moving vehicles could be extracted from background. Experiments show that this algorithm can eliminate the mismatched optical flows effectively and detect the moving vehicles accurately and stably, and it can apply to single and multiple vehicle detection.

关 键 词：运动车辆检测 动态场景 矢量量化 角点检测 Lucas-Kanade光流 

分 类 号：U491.116[交通运输工程—交通运输规划与管理]