This framework for evaluating human activity recognition (HAR) systems for body-worn wearable devices is designed to help users make informed decisions about whether the specific system is suitable for a given application. The first-of-its-kind innovation from Georgia Tech helps address the knowledge gap typically encountered at the onset of considering HAR systems by providing recognition performance estimates and HAR configurations for novel tasks before dedicating resources to solve them. This may help domain experts, rather than technology developers, make intelligent choices about whether to invest in a HAR system they may be considering.
Georgia Tech’s framework presents tasks as a small pilot dataset, recorded and annotated prior to a planned deployment. Off-the-shelf analysis tools measure the complexity of each task according to an eight-dimension computation. Mapping a complete analysis task to this numerical representation enables quantification of the complexity and challenges associated with each HAR task. It also enables comparison of unknown tasks to existing ones that already have optimized analysis workflows. This robust categorization aims to provide concrete and actionable guidelines for practitioners regarding specific HAR deployments.
- Novel: Provides a framework designed to address a major challenge associated with HAR deployment decisions, for which there is no prior state of the art
- Insightful: Fills a knowledge gap with information about how well a HAR system is suited to an application before investing in it, helping users make intelligent decisions that may save time and resources
- Practical: Employs off-the-shelf analysis methods and therefore may not require a data expert at the pilot stage
- Robust: Uses multidimensional numerical computation to help accurately categorize HAR task complexity
- Clinical assessments
- Health monitoring
- Sports and fitness monitoring
- Human-computer interaction research
Activity recognition using HAR systems is of great research interest for analyzing human activities, behaviors, and routines. However, at the onset of deploying such systems, there is typically very little knowledge about how well a task can be solved and what would be a recognition estimate that can be obtained on the task. Without a realistic assessment of the complexity of an analysis task, modifications required to develop problem-specific data analysis pipelines can either turn out to be impossible or may require substantial resources with no performance guarantees. Georgia Tech’s framework uniquely addresses this challenge, with no other comparable method currently available.