CAVIAR Short-term Activity Recognition


IST developed a set of 29 features describing both the target global trajectory and internal motion. The first subset of features are derived from target positions over time, characterizing the target instantaneous or time-averaged global velocity and speed as well as information related to second order moments (variance, covariance matrices, etc). The second set of features characterizes the target internal motion and are derived from instantaneous optic flow measurements.

IST derived and implemented a Bayesian classifier for short-term (2 seconds of video) activity recognition. The likelihood function is modeled as a mixture of Gaussians and estimated with the EM algorithm. Temporal integration is achieved assuming independence over a set of consecutive frames (naive Bayes approach).Classification accuracy compared to the ground truth is more than 98%.

Five activity types were represented: {Active, Inactive, Walking, Running, Fighting} and recognized in the Bayesian hierarchical classifier, where 3 of the 29 features were selected for each classifier stage. This diagram shows the classifier tree that had the best results, with the labels showing which activities were discriminated at each classifier:

A paper that describes the process is:

  1. P. Ribeiro, J. Santos-Victor. Human Activities Recognition from Video: modeling, feature selection and classification architecture. Proc. Workshop on Human Activity Recognition and Modelling (HAREM 2005 - in conjunction with BMVC 2005). pp 61-70, Oxford, Sept 2005.

Back to CAVIAR home page.