Fast estimation of multiple vector fields: Application to video surveillance

Characterizing object motion in a given scene is a central problem in computer vision and image analysis. Object motion has been recently modeled by using multiple motion fields; this model allows characterizing typical motion patterns and, among other possible applications, may be used to detect abnormal events. However, the estimation of multiple fields from video information (e.g., trajectories) is a challenging task since we do not know which field is active at each instant of time, for each object. This difficulty has been successfully addressed by using iterative approaches in which the estimation of the active field alternates with the field update, using the expectation-maximization (EM) algorithm or variants thereof. However, the EM method for this problem has been shown to be slow and to yield field estimates that depend on the initialization. This paper describes an alternative approach for the estimation of multiple overlapping fields, using a label propagation algorithm. The proposed algorithm, which is not iterative, is fast and has good performance on synthetic and real data.