Motion Clustering using Spatiotemporal Approximations

In this paper a new technique is proposed for the clustering and classification of spatio-temporal object trajectories extracted from video motion clips. The trajectories are represented as motion time series and modelled using Chebyshev polynomial approximations. Trajectory clustering is then performed to discover patterns of similar object motion. The coefficients of the basis functions are used as an input feature vector to a Self-Organising Map which can learn similarities between object trajectories in an unsupervised manner. It is shown that applying machine learning techniques in the Chebyshev parameter subspace leads to significant performance gains over previous approaches that encode trajectories as point-based flow (PBF) vectors. Experiments using the PETS’04 tracking dataset demonstrate the effectiveness of clustering in the parameter subspace and improvements in overall classification accuracy in comparison with PBF vector encoding. We also show how this technique can be further extended to the detection of anomalous motion paths. Applications to motion data mining and event detection in video surveillance systems are envisaged.