Deep trajectory representation-based clustering for motion pattern extraction in videos
Jonathan Boyle, Tahir Habib Nawaz, James M. Ferryman · 2017
We present a deep trajectory feature representation approach to aid trajectory clustering and motion pattern extraction in videos. The proposed feature representation includes the use of a neural network-based approach that uses the output of the smallest hidden layer of a trained autoencoder to encapsulate trajectory information. The trajectory features are then fed into a mean-shift clustering framework with an adaptive bandwidth parameter computation to yield dominant trajectory clusters. The corresponding motion patterns are extracted based on a distance minimization from the clusters' centroids. We show the effectiveness of the proposed approach on challenging public datasets involving traffic as well non-traffic scenarios.