Classification and Decision-Theoretic Framework for Detecting and Reporting Unseen Falls

Abstract

Detecting falls is critical for an activity recognition system to ensure the well being of an individual. However, falls occur rarely and infrequently, therefore sufficient data for them may not be available during training of the classifiers. Building a fall detection system in the absence of fall data is very challenging and can severely undermine the generalization capabilities of an activity recognition system. In this thesis, we present ideas from both classification and decision theory perspectives to handle scenarios when the training data for falls is not available. In traditional decision theoretic approaches, the utilities (or conversely costs) to report/not-report a fall or a non-fall are treated equally or the costs are deduced from the datasets, both of which are flawed. However, these costs are either difficult to compute or only available from domain experts. Therefore, in a typical fall detection system, we neither have a good model for falls nor an accurate estimate of utilities. In this thesis, we make contributions to handle both of these situations.

In recent years, Hidden Markov Models (HMMs) have been used to model temporal dynamics of human activities. HMMs are generally built for normal activities and a threshold based on the log-likelihood of the training data is used to identify unseen falls. We show that such formulation to identify unseen fall activities is ill-posed for this problem. We present a new approach for the identification of falls using wearable devices in the absence of their training data but with plentiful data for normal Activities of Daily Living (ADL). We propose three 'X-Factor' Hidden Markov Model (XHMMs) approaches, which are similar to the traditional HMMs but have ``inflated'' output covariances (observation models). To estimate the inflated covariances, we propose a novel cross validation method to remove 'outliers' or deviant sequences from the ADL that serves as proxies for the unseen falls and allow learning the XHMMs using only normal activities. We tested the proposed XHMM approaches on three activity recognition datasets and show high detection rates for unseen falls. We also show that supervised classification methods perform poorly when very limited fall data is available during the training phase.

We present a novel decision-theoretic approach to fall detection (dtFall) that aims to tackle the core problem when the model for falls and information about the costs/utilities associated with them is unavailable. We theoretically show that the expected regret will always be positive using dtFall instead of a maximum likelihood classifier. We present a new method to parameterize unseen falls such that training situations with no fall data can be handled. We also identify problems with theoretical thresholding to identify falls using decision theoretic modelling when training data for fall data is absent, and present an empirical thresholding technique to handle imperfect models for falls and non-falls. We also develop a new cost model based on severity of falls to provide an operational range of utilities. We present results on three activity recognition datasets, and show how the results may generalize to the difficult problem of fall detection in the real world. Under the condition when falls occur sporadically and rarely in the test set, the results show that (a) knowing the difference in the cost between a reported fall and a false alarm is useful, (b) as the cost of false alarm gets bigger this becomes more significant, and (c) the difference in the cost of between a reported and non-reported fall is not that useful.