Detection of activities and events without explicit categorization

Abstract

We address the problem of unsupervised detection of events (e.g., changes or meaningful states of human activities) without any similarity test against specific models or probability density estimation (e.g., specific category learning). Rather than estimating probability densities, very difficult to calculate in general settings, we formulate the event detection as binary classification with density ratio estimation [9] in a hierarchical probabilistic framework. The proposed method takes pairs of video stream data (i.e., past and current) as input with differing time-scales, generates density ratio models in a way of online learning, and judges if there is any 'meaningful difference' between them based on the multiple density ratio estimations. Through experimental studies on real-world scenes of specific domains using challenging datasets from sports scene (i.e., tennis match) with complex background, we demonstrate the potential advantage of our approach over the state-of-the-art in terms of precision and efficiency.

Density-Difference Estimation

“…Density differences are useful for various purposes such as class-balance estimation under class-prior change (Saerens et al, 2002;Du Plessis & Sugiyama, 2012), change-point detection in time series (Kawahara & Sugiyama, 2012;Liu et al, 2012), feature extraction (Torkkola, 2003), video-based event detection (Matsugu et al, 2011), flow cytometric data analysis (Duong et al, 2009), ultrasound image segmentation (Liu et al, 2010), non-rigid image registration (Atif et al, 2003), and image-based target recognition (Gray & Principe, 2010).…”

Detecting Interesting Events Using Unsupervised Density Ratio Estimation

“…This approach exploits the insight from machine learning that it is much easier to learn a ratio of two probability densities in a high dimensional space than to learn each separately. This is why density ratio estimation is used in many fields, such as outlier detection [2] and change points detection [16], etc. In this model, we view R and I as training data for a two-class classification problem in which we assign a label y = +1 if the frame is classified as originating from R and y = −1 if it is classified as originating from I.…”