Prediction of Inhabitant Activities in Smart Environments

Bryan David Minor

Human activity prediction is a challenging problem which poses a number of machine learning challenges. Predicting occurrences of activities is valuable in understanding human behavior and creating services that not only react to human activities, but proactively anticipate and prepare for future activities in advance. We hypothesize that human activities can be predicted from smart environment sensor data. In order to validate this hypothesis we introduce machine learning-based methods to create an Activity Predictor (AP). AP learns a model of inhabitant behavior from sensor observations of their daily lives. It then automatically generates numeric predictions of the time until future occurrences of activities of interest. These predictions can be used to facilitate a variety of smart environment services including activity prompting and energy management systems. We explore the utility of drawing contextual features from sensor event data in order to inform the predictor. These features draw on the information provided in the temporal distribution and the values of the sensor events. We further develop a method of using information about activity relationships to provide joint activity prediction. These connections are established through the use of prediction lag features which supplement other contextual information in a recurrent model. We also propose a method of using a combination of independent and recurrent predictors in a two-pass fashion to further enhance activity learning. We utilize data collected from several CASAS smart environment datasets to validate the methods in support of our hypothesis. We also explore the use of various evaluation metrics and underscore the importance of using a combination of metrics to fully understand activity prediction performance. Our findings suggest that the proposed automated solutions are able to predict human activities with reasonable accuracy. To further evaluate our methods we have deployed them as part of two pilot tests of our CAFE prompting app. Predictions were found to be less than 30 minutes in most cases and the prompts were found to have a noticeable effect on participants' behavior.

Prediction of Inhabitant Activities in Smart Environments

Files and links (1)

Abstract

Metrics

Details