一种俯视行人的多目标检测和跟踪方法

Zenithal Pedestrian Detection and Tracking

提出了一种针对俯视客流计数的多目标自动检测和跟踪方法。该方法主要包括四个组成部分:运动检测、目标表示与识别、目标跟踪建模、目标跟踪。即首先通过视频帧间差法粗略检测出一帧图像中成片运动或变化的区域,在变化区域里遍历搜索目标;视觉目标表示为外形对称、外观具有某种形状结构、颜色均匀等一系列弱分类特征串联组成的形式,可以快速检测判别俯视行人的头部;再对检测到目标重新建模以便后续的跟踪;最后利用连续自适应的均值移位算法跟踪检测到的多个目标。检测与跟踪并行运行,跟踪可以验证检测到的目标。实验证明,该方法是有效的。

A method of automatic multi-target detection and tracking with respect to zenithal pedestrian counting was put forward. There are mainly four parts. motion detection, object modeling, object detection and tracking. First the motion area is segmented using frame difference. Then the pedestrian head can be quickly detected using a series of simple cascaded characteristic in the area, such as the symmetry or asymmetry of a head, some shape of the appearance and the uniform color of that. After that, a suitable object model is set up to keep track of location and otherproperties. At last these targets can be tracking using the continuous adaptive mean shift tracking algorithm. Object detection and tracking run simultaneously and the tracking can verify the detection. Experiments show that the method is effective.