A Rule-Based Approach to Activity Recognition

This paper presents a rule-based framework for activity classification and illustrates how domain-specific expert knowledge and observation of data in its feature space can be used for rule construction. To demonstrate its practical value, the framework is applied on datasets collected during an orientation-independent activity recognition experiment. Through an implementation based on the Java Expert System Shell (JESS), two types of rules are compared: rules that are specifically constructed for each individual device orientation and those constructed without assuming any prior knowledge on device orientations. Overall accuracy improvements of 7.97% and 9.25% are observed on training and test datasets when orientation-specific rules are used.