Unsupervised Activity Perception by Hierarchical Bayesian Models

Abstract

We propose a novel unsupervised learning framework for activity perception.

Learning Rare Behaviours

“…sharing substantial visual words of typical behaviours) are usually overwhelmed (explained away) by the more obvious and common behaviours, and therefore cannot be detected. Moreover, all outlier detection based approaches [1][2][3][4][5] have no mechanism to classify different types of rare behaviours.…”

“…They can deal with simultaneous interaction of multiple objects or activities. Despite these advantages, all unsupervised topic models [3][4][5] for rare behaviour detection must compute how well new examples can be explained by the learned typical behaviour model. This exposes their key limitation: unsupervised topic models are only sensitive to rare behaviours which are visually very distinct from the majority.…”

“…In particular, we explore prior domain knowledge about spatio-temporal conflict of multiple existing topics to synthesise new topics of unknown behaviours. Although these behaviours do not cover every possible case in the real world, they are usually of particular interest but can easily be missed by standard approaches based on outlier detection [1,[3][4][5]. The synthesising procedure is as follows:…”

“…These models however suffer from lack of robustness to noise and clutter in crowded scenes. Recently a number of studies on modelling complex scenes have instead exploited statistical topic models such as Latent Dirichlet Allocation (LDA) [8] to cluster low level features into topics -or activities in video -for robust modelling of behaviour [3][4][5]. Topic models aim to automatically discover meaningful activities via co-occurrence of visual words.…”

“…Existing studies usually treat this problem as either an outlier detection problem, flagging any behaviour which is badly explained by a model representing typical behaviours [1][2][3][4][5]; or a supervised classification problem when examples of rare behaviours are known and available for training an explicit model of rare versus typical behaviours [6,7]. For both, the real challenge in model building is that visual evidence for rare behaviours is subtle and under-represented, which makes them neither distinctive enough to be detected among many concurrently occurring typical behaviours, nor repeated enough to be modelled precisely.…”