Multi-agent detection and labelling of activity patterns

Abstract

We address the problem of automatic detection and labelling of far-field activity patterns from target trajectories using vector field abstractions that are estimated by multiple communicating agents in a Kalman filter (KF) framework. We propose a novel 2D application of the diffusion KF. The proposed approach yields multiple vector field abstractions at the same spatial regions by allowing the estimator agents to create a library of states. We compute internal consistency measures to assess the estimated vector fields, and establish thresholds that signal low performance estimates. Experimental results on synthetic and real data sets show that the proposed approach correctly detects activity patterns from training trajectories and, using new test trajectories one at a time, either matches them to previously detected activity patterns or detects new activity patterns that are added to the agents library. Moreover, the internal consistency measures correctly flag the low-performer vector field abstractions.