Event monitoring via local motion abnormality detection in non-linear subspace

We present a computational approach to abnormal visual event detection, which is based on exploring and modeling local motion patterns in a non-linear subspace. We use motion vectors extracted over a region of interest (ROI) as features and a non-linear, graph-based manifold learning algorithm coupled with a supervised novelty classifier to label segments of a video sequence. Given a small sample of annotated normal motion vectors, the non-linear detector ranks segments in a sequence as a function of abnormality. We evaluate the proposed method and compare its performance against the use of other low-level features such pixel appearance and change detection masks. Our choice of feature space compares favorably to the alternatives in terms of classification performance, sensitivity to noise as well as computational complexity.