基于单目视觉的车辆识别通常分为候选车辆生成( CG)和候选车辆验证( CV)两个步骤。传统的CG步骤往往采用遍历的方法，获得的候选车辆窗口数量庞大，增加了后续CV阶段的计算耗时，难以满足实际应用的实时性要求。本文提出一种基于几何和深度信息的CG方法，在不丢失有效车辆区域的前提下极大减少了候选车辆的数量。该方法首先将图像以超像素形式进行分块，同时利用预先训练的Adaboost分类器获取超像素图像的几何信息和粗糙深度信息。然后利用车辆在世界坐标系下的垂直度、位置和尺寸等先验知识，采用了一种分层聚类策略，合并图像中属于车辆的超像素块并生成候选车辆。与传统算法的比较结果表明，本方法以检测率的微小降低为代价，实现了候选车辆窗口数量的大幅度减少。%Monocular vision based vehicle identification are often divided into two steps:candidate genera-tion ( CG) and candidate validation ( CV) . Traditional CG procedure adopting ergodic approach often generates a large amount of candidate windows, which dramatically increase the calculation time in CV phase and hence is hard to meet the real-time requirements of practical application. In this paper a novel vehicle candidate generation meth-od is proposed based on geometry and depth information, which can greatly reduce the number of candidate windows generated. With the method, firstly images are divided into super pixel regions, and the geometry information and coarse depth information of images are obtained with pre-trained Adaboost classifier. Then by using the prior knowl-edge of vehicles ( verticality, location and size) in global coordinate system, a hierarchical clustering strategy is a-dopted to merge the vehicle super pixel blocks in images and generate vehicle candidates. The results of comparison with traditional algorithms show that the method proposed achieves a great reduction in the number of candidate win-dows with a cost of minor drop in detection rate.
展开▼
