Unusual event detection and prediction based on sectional contextual edit distance

Abstract

We redefine the unusual event detection problem from a different point of view. Several fundamental event features are investigated and adopted. These features are redescribed in a uniform model. Thus, using this model, supervised/unsupervised unusual event detection algorithms can be designed to fit various situations. Trajectory is treated as the most important feature. To more accurately measure the similarity of different moving object trajectories, a novel distance measurement, the sectional contextual edit distance (SCED), is developed. In the SCED, cost functions are designed according to contextual information and trajectories are segmented into subsections automatically, based on the relevant contexts. Velocity and orientation are also taken into account in cost functions to build an integrated distance similarity measurement. Experimental results demonstrate better performance using the newly proposed similarity measurement while being compared with the existing methods, and some cases of the unusual event detection problem are also demonstrated.