Direct optimization of inference model for human activity and posture class recognition

Abstract

The demand for ambulatory monitoring is rising due to the widespread adoption of home-based healthcare models for the increasing elderly population. Since monitoring patients makes it possible to provide remote medical services based on real-world contexts, context-aware technology is regarded as a crucial element in core module development for ubiquitous healthcare solutions. This paper describes the design and implementation of a class label tagging system for human activity recognition based on sensor data from a single tri-axial accelerometer attached to the waist of a human subject. Our human activity and posture classifier (APC) was designed to model more effectively an event that takes place over a period of time. Consequently, the APC problem becomes the process of tagging class labels onto sequential data from a single tri-axial accelerometer. This sequential tagging naturally led to our hypothesis that a linear combination of the classification model and the prior, approximated by observing a state transition event of human activity and posture (AP), may decrease the error rate of the system. Therefore, this paper aims at testing the aforementioned experimental hypothesis while explicating our method of direct optimization, also known as direct loss minimization. Our experimental hypothesis is based on the view that the task of labeling a series of class tags on a data object of the AP event should be supported by a statistical model capable of robustly capturing the probabilistic context influencing the generation of a set of features at a given discrete time. A radial basis function network algorithm was used for parameter estimation of a class likelihood inference model augmented by training a state transition model to capture the sequential nature of a posture or activity generation event. The linear combination requires our design step to have a further optimization, in addition to learning the parameters of each inference model, of the model parameters that represent the relative contribution to the decision process. Previous studies have elaborated on an effective approach to direct optimization of model parameters when an appropriate evaluation metric exists for a given problem domain. In our case, f-measure was used as an objective function of the direct optimization. We used local beam search to optimize directly the weight vector for the likelihood and the prior. Our method is tested on the set of nine-dimensional feature vectors contained in 69,376 lines of sequential AP data. The experiment shows that embedding prior probability into the recognition process decreases the error rate in a set of categories. Several experiments proved the efficiency of our proposed method and the usefulness of the system.