Self-Tuned Unsupervised Learning of Motion Trajectories

Abstract

This paper presents a novel technique for clustering of video motion clips using coefficient-space representation of object trajectories. Trajectories are treated as time series and modelled using orthogonal basis function representation. Various function approximations have been compared including Chebyshev polynomials, Piecewise Aggregate Approximation, Discrete Fourier Transform (DFT) and Modified DFT (DFT-MOD). A novel framework (HSACT-SOM) is proposed for unsupervised learning of motion patterns without having prior knowledge of number and types of patterns hidden in datasets. Experiments,
using simulated and complex real life trajectory datasets, demonstrate the superiority of our proposed HSACT-SOM based motion learning technique compared with other recent approaches. The comparison is performed in terms of the quality of clustering, based on cluster validity indices, and the number of clusters
discovered in variety of datasets.