基于Hausdorff距离的视觉监控轨迹分类算法

针对智能视觉监控系统中的运动目标轨迹分类问题,提出了一种基于多维Hausdorff距离的轨迹聚类算法。该算法使用流矢量序列描述目标运动轨迹,由多维Hausdorff距离进行轨迹相似性测量,通过谱聚类实现轨迹分类。该算法在轨迹描述中同时包含位置和方向信息,解决了Hausdorff距离不能区分轨迹运动方向的问题。为降低计算复杂度,本文还提出一种保距变换对轨迹相似性测量进行优化。与相关算法的对比实验表明,提出的轨迹分类算法可达到更高的聚类准确率;提出的保距变换可以显著降低算法的计算复杂度。

Trajectory lcassification based on Hausdorff distance for visual surveillance system

A trajectory clustering algorithm based on multi-dimensional Hausdorff distance is proposed for classification of trajectories of moving objects in intelligent visual surveillance system. First, the trajectory of a moving object is described using a sequence of flow vectors. Then the similarity between trajectories is measured by their respective multi-dimensional Hausdorff distances. Finally, the trajectories are clustered by the spectral clustering algorithm. The proposed algorithm is different from other schemes using Hausdorff distance that it includes both the position and direction information in the flow vectors; hence it can distinguish the trajectories in different directions. A distance preserving transformation is also proposed to reduce the computational complexity of the similarity measure. Experimental results show that, comparing with other algorithm, the clustering accuracy of the proposed algorithm is better, and the proposed distance preserving transformation can greatly reduce the computational cost.