Unusual event detection and prediction based on sectional contextual edit distance

摘要

We redefine the unusual event detection problem from andifferent point of view. Several fundamental event features are in-nvestigated and adopted. These features are redescribed in a uni-nform model. Thus, using this model, supervised/unsupervised un-nusual event detection algorithms can be designed to fit variousnsituations. Trajectory is treated as the most important feature. Tonmore accurately measure the similarity of different moving objectntrajectories, a novel distance measurement, the sectional contextualnedit distance (SCED), is developed. In the SCED, cost functions arendesigned according to contextual information and trajectories arensegmented into subsections automatically, based on the relevantncontexts. Velocity and orientation are also taken into account in costnfunctions to build an integrated distance similarity measurement. Ex-nperimental results demonstrate better performance using the newlynproposed similarity measurement while being compared with the ex-nisting methods, and some cases of the unusual event detectionnproblem are also demonstrated.