Sparse Coding of Point Cloud Data

Point clouds, made available through laser range finders, stereo cameras, or time of flight cameras, are frequently used in robot navigation systems. However, no unsupervised machine perception algorithm exists to provide understanding of the data; e.g. that a particular blob of points looks roughly like, say, a car. In this paper, we take steps towards such an algorithm based on sparse coding. The work here generalizes to any binary data. An algorithm for learning the bases and the activations for point cloud data is derived and demonstrated. Given precomputed basis vectors and an input vector, calculating the activations is very fast and could be used in real-time applications.