Enhanced Local Subspace Affinity for feature-based motion segmentation

摘要

We present a new motion segmentation algorithm: the Enhanced Local Subspace Affinity (ELSA). Unlike Local Subspace Affinity, ELSA is robust in a variety of conditions even without manual tuning of its parameters. This result is achieved thanks to two improvements. The first is a new model selection technique for the estimation of the trajectory matrix rank. The second is an estimation of the number of motions based on the analysis of the eigenvalue spectrum of the Symmetric Normalized Laplacian matrix. Results using the Hopkins155 database and synthetic sequences are presented and compared with state of the art techniques.