Unsupervised Pattern Analysis of Vehicle Tracks

In this effort ALPHATECH will develop algorithms for extraction of purposeful temporal-spatial patterns from moving vehicle data and demonstrate their efficiency on a set of simulated MPA scenarios. We apply unsupervised learning algorithms to find similarities between traffic activities. Traffic activities are constructed using ALPHATECH's multiple hypothesis tracking algorithm and grouping correlated tracks into most probable activities. We use clustering techniques in order to identify motion patterns as groups of similar activities where similarity is defined based on the features inferred from MTI observations. In order to develop a full set of traffic activity features, we will perform Bayes Net estimation of kinematic and motion vehicle/group attributes and traffic centers analysis. Discovered clusters can be used to describe motion patterns in high-level terms for interaction with experts, to build statistical models of the motion patterns and ambient traffic, and to predict future traffic activities. We will perform feasibility and feature importance analysis by simulating known military patterns embedded into ambient traffic. Scenarios will be generated using existing ALPHATECH behavior pattern and motion target generators.