雷达与视觉信息融合的行人检测方法

Pedestrian detection by radar vision data fusion

首先建立两传感器间的坐标变换模型,将雷达扫描的深度信息映射到图像上。然后对雷达扫描的数据进行聚类分析,并结合行人物理属性,对聚类点簇进行筛选获得有效点簇信息;根据摄像机的成像规律,确定行人成像区域与其所处位置的关系,从而确定行人检测感兴趣区域。在此基础上,提取感兴趣区域的梯度方向直方图特征,运用支持向量机检测行人。实际交通场景测试结果表明本文方法能够对行人实时检测,且准确率可达93%以上。

A pedestrian detection method by radar and vision information fusion is proposed. First, a coordinate transformation model of two sensors is established, by which depth information of radar scanning can be mapped to image. Then, radar scanning data are divided into a number of point clusters. Considering the physical attribute of pedestrian, the radar point clusters are selected again to obtain effective point cluster information. According to camera imaging rule, the relationship between pedestrian imaging region and his/ her position is deduced. As a result, the Region of Interest （ROI） in pedestrian detection is determined. On this basis, the Histogram of Oriented Gradient （HOG） feature of the ROI is extracted,and the hypothesized pedestrian is verified using Support Vector Machine （SVM） classifier. Experiment results indicate that the proposed method is able to detect pedestrian in real time,and the accuracy rate can exceed 93%.