Mining regular behaviors based on spatiotemporal trajectory multi‐dimensional features

In the military and civilian surveillance domain, it is of great significance to mine regular behaviours of targets for situation awareness and command decision support. Most of the existing trajectory clustering algorithms only consider the similarity of spatial position of the trajectory, without sufficient multi‐dimensional information such as time, course and velocity. Some approaches based on information fusion take these multi‐dimensional information into account, but the features with different dimensions fused by weight coefficients are not robust and universal for different scenarios. In this paper, a regular behaviour mining method based on spatiotemporal trajectory multi‐dimensional features and density clustering is proposed. Firstly, multi‐dimensional Hausdorff similarity is defined to measure spatiotemporal trajectory from different feature dimensionalities. Different from methods based on information fusion, the proposed method defines trajectory density in feature similarity of different dimensions and adaptively determines parameters according to feature distribution in different dimensions. Experimental results in simulated and radar measured trajectory data show that the proposed method can be accurate and robust in clustering evaluation indexes such as Purity, Precision, Recall and Rand Index from different scenarios, which has a good application prospect in intelligent surveillance tasks.