Clustering and visualizing actions of humans and animals using motion features

We propose a technique to cluster actions of humans and animals. We use domain specific motion features and employ spectral clustering on them to cluster activities. For humans, we use existing optical flow features. For animals, we cluster behaviors of a grasshopper. We track it in 3D and construct features using 3D object movement which discriminate between different classes of actions. We employ spectral clustering on the extracted features for each domain. Due to the large amount of data we use the Nystrom extension which samples from the data and computes the eigenvalues and eigenvectors of affinities between them and extends it to the eigenvectors of the full affinity matrix. We use the K-means algorithm to do the final clustering. We experimented with our method on the KTH data set and videos of one grasshopper. We create a summary visualization of our results using an extension of an existing framework.