Abstract: An efficient car detection method is proposed to detect cars with arbitrary aspect ratios or angles, which is based on vehicle shape features, object proposals and cascade Boosting classifiers. The train dataset is clustered according to different angles or aspect ratios of cars firstly. Then the Accumulated Channel Features（ACF） of car areas are extracted, which is learned by AdaBoost to detect cars. Moreover, the object proposal, which is calculated by EdgeBox, is used instead of sliding windows for computation issues. A dataset with three difficulty levels, which consists of more than 3?500 images, is used to test the proposed method. The experimental results indicate that compared with the ACF-based, the DPM-based, the NPD-based and the HOG-Haar-based methods, the proposed method is the best with an average detection accuracy of 85% on all test sets with different difficulty levels.

Key words: vehicle detection, shape clustering, object proposals, cascade classifier

摘要： 为了快速定位监控场景中不同姿态的车辆位置，结合车辆外形特征、目标候选区域和级联Boosting分类器进行多角度车辆检测。对不同角度和纵横比的车辆进行聚类，然后对每种姿态的车辆提取候选区域的累积通道特征（ACF），使用AdaBoost学习分类器用于车辆检测，此外，检测时用边缘框计算可能存在物体的区域代替滑动窗法进行提速。以含有较难、中等、较易3种难度水平超过3?500个样本为测试集进行了快速车辆检测测试，并与ACF、DPM、NPD和HOG-Haar等4种方法进行了对比，实验结果表明基于候选区域的车辆检测方法性能最优，在3种测试集上平均达到了85%以上的检测率。

关键词: 车辆检测, 形状聚类, 目标候选区域, 级联分类器