Identifying Important Action Primitives for High Level Activity Recognition

Abstract

Abstract. Smart homes have a user centered design that makes human activity as the most important type of context to adapt the environment according to people's needs. Sensor systems that include a variety of ambient, vision based, and wearable sensors are used to collect and transmit data to reasoning algorithms to recognize human activities at different levels of abstraction. Despite various types of action primitives are extracted from sensor data and used with state of the art classification algorithms there is little understanding of how these action primitives affect high level activity recognition. In this paper we utilize action primitives that can be extracted from data collected by sensors worn on human body and embedded in different objects and environments to identify how various types of action primitives influence the performance of high level activity recognition systems. Our experiments showed that wearable sensors in combination with object sensors clearly play a crucial role in recognizing high level activities and it is indispensable to use wearable sensors in smart homes to improve the performance of activity recognition systems.

Dimensionality reduction of data sequences for human activity recognition

“…In [27], the potential of recent machine learning methods for discovering universal features is investigated for context-aware applications of activity recognition. In [28], the authors utilize action primitives that can be extracted from data collected by sensors worn on human body and embedded in different objects and in the environment to identify how various types of action primitives influence the performance of high level activity recognition systems. In [29], a technique is presented for using on-body accelerometers to assist in automated classification of problem behavior during such direct observation.…”

Signal processing technologies for activity-aware smart textiles

“…Such 'daily routines' are usually composed of a complex interleaving of lowerlevel activities ). Thus, recognizing high-level activities builds upon the recognition of lower-level action primitives in a hierarchical manner (Hong et al , 2009;Manzoor et al , 2010).…”

Neutrosophic Concept Lattice based Approach for Computing Human Activities from Contexts

“…Study of effectiveness of context data in describing activities on same dataset was also found in [26]. Purpose of evaluation was to select the relevant set of contexts.…”

“…The accuracy of non-deterministic approaches is slightly lesser, as our approach penalizes the recognition in such activities. The recognition algorithms used in [26] work like a black box and recognition process is not explainable. In our case the recognition process is semantically clear and self-explanatory.…”