A Novel Method for Trajectory Analysis in Surveillance

In this paper, we propose a nonparametric grammar based framework for analyzing trajectories, aiming to discover the motion pattern of objects and assist human understanding. The framework works in three steps. 1) Raw trajectories are smoothed to eliminate noise, and then, points and segments are sampled as primitive units. 2) The primitive units are clustered based on DPM and HDP-HMM, in order to learn the pre-terminal symbols in the grammar. 3) Trajectories (sequences of primitive units) are modeled with ISCFG, and parse trees are achieved by using Viterbi algorithm for further research. Compare with previous works, our approach includes temporal, spatial and structural information in a single model. All the parameters can be learned from training set and can be adapted online. The parse tree of trajectories can be exploited for further applications, such as path prediction and anomaly detection.