Video event segmentation and visualisation in non-linear subspace

We introduce the use of dimensionality reduction for video event detection without explicitly using motion estimation or object tracking. Raw data from video sequences are used to construct a low-dimensional mapping representing the input frames. We compare principal component analysis, multi-dimensional scaling, isomap, maximum variance unfolding and Laplacian eigenmaps and implement an approach based on local, non-linear dimensionality reduction. We propose an approach with a graph based on the similarity of frames and enriched with the temporal information from the sequence processed by Laplacian eigenmaps. This makes it possible to visualise the manifold of motion in the scene and to detect unusual events in a low-dimensional space. We demonstrate the approach on standard traffic surveillance test sequences.