Abstract: This paper proposes a novel pedestrian detection method. An angular-diffused Shape Context(SC) descriptor is proposed to obtain the histogram of the sampled edge map with considerations of different edge orientations. Modified Hausdorff Distance(MHD) is employed as the similarity likelihood between the codebook of the model and the test image. A voting map of the central hypotheses of pedestrians is generated, followed by the foreground and background segmentation which utilizes the binary masks of the templates. In the people candidate regions, a color- segmentation-based verification and a SVM-based shape classifier are subsequently performed to reduce two types of false positive results. Experiments on a pedestrian image database and PASCAL database illustrate the improved performance of the angular-diffused SC in representing the shape of up-right human body as well as the reduced false positive rate introduced by the two-step verification.

Key words: Shape Context(SC), pedestrian detection, codebook, SVM, angular-diffused

摘要： 提出一种行人检测算法。该算法使用角度泛化的形状上下文描述子提取边缘采样点集的直方图分布，以改进的Hausdorff距离作为模型与待测图像码本之间的匹配度量，利用该相似度为人体中心位置投票，并结合模板的二值掩码分割人体的前景与背景。采用颜色聚类与支持向量机形状分类器的两步验证法去除2类假阳结果。在自建行人图像库与PASCAL库上的实验结果表明，角度泛化的形状上下文提高了对直立人体形状的局部描述性能，两步验证明显降低了误检率。

关键词: 形状上下文, 行人检测, 码本, 支持向量机, 角度泛化

CLC Number: 

• TP391.41