Discovering activity interactions in a single pass over a video stream
Omar U. Florez, Curtis Dyreson · 2012
We propose a novel, unsupervised learning framework to find interactions of activities in streaming video data. In traditional streaming problems there is a priori knowledge of the atomic elements in the stream. But in a video stream, the elements only emerge as activities of moving objects, and consequently must be learned or discovered in the stream. This paper describes a two-stage framework that discovers activity interactions in a single pass over the stream. In the framework, activities are modeled as distributions over low-level visual features and interactions as frequent co-occurrence relationships over those activities. Off-line, similar low-level visual features are grouped into activities. On-line, co-occurring relationships caused by frequent activity interactions in the same scene are discovered in the video stream. By decoupling the stages, we can incrementally learn interactions regardless of the variable distribution of co-occurring relationships over time. The usefulness of the proposed framework is empirically tested in experiments that compare its performance with two state-of-the-art Hierarchical Bayesian Models for segmenting activities and discovering interactions in surveillance videos containing complex traffic scenes governed by multiple semaphores.