Locality-constrained Multi-Instance Learning for Abnormal Trajectory Detection

摘要

Abnormal event detection based on trajectory has been extensively investigated in recent years; however, problems remain when processing an incomplete trajectory that usually has abnormality in some parts of the whole trajectory and the rest are normal. In this paper, we propose a locality-constrained multi-instance learning framework for abnormal trajectory detection. We explore local adaptability for robust trajectory classification, and partition each trajectory into tracklets by control points of cubic B-spline curves. Then, the tracklets are modeled by Hierarchical Dirichlet Process-Hidden Markov Model (HDP-HMM). Finally, the whole trajectory is considered within the multi-instance learning framework as bags, when abnormal ones are positive bags consist of tracklets, normal trajectories are negative bags with tracklets. With experimental results on the CAVIAR dataset, it shows that the proposed method achieves better performance than several recent approaches.