Human activity data analytics

• Main Author: Wee, Jia Yi
• Other Authors: Tan Ah Hwee
• Format: Final Year Project
• Language: English
• Published: 2015
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/62601
• Institution: Nanyang Technological University

Human activity recognition related technologies are enjoying a fuelled growth in investment as they gains popularity around the world, especially so for its application in home-based patient care, elderly living and many more. There exists a need to support this trending innovation with a robust and extensible motion recognition framework to serve as a widely applicable foundation for this research arena. Majority of the existing work face the problem of having a recognition system with limited generalisation capability (e.g. handling users from different age groups or handling situations when sensor devices are unfixed) as well as the inability to address high level motion transitions (e.g. Standing Up or Falling Down). In this project, we propose for the establishment of a highly robust human activity recognition framework which can accurately detect both low level motion (e.g. Sitting, Standing) and high level motion transitions (e.g. Standing Up, Falling Down). This new framework will be referred to as the Robust Activity Recognition with Motion Transition (RARMT) framework in the rest of this report. Furthermore, an extensive set of experiments will be conducted to test for the performance of the RARMT framework, providing assurance on the reliability of the system. The three major contributions in this project are listed as follows. Firstly, we established the low level motion recognition processing components of RARMT framework and used several benchmark datasets for performance testing. The benchmark results achieved are positive and there is even an accuracy improvement over what was achieved by the benchmark models. Secondly, we collected a new dataset covering two distinct age groups (i.e. youth and elderly) while also capturing human behavior in their most natural environmental settings. Extensive experimentations were performed and promising results were achieved, proving the RARMT’s robustness in handling diverse datasets and noise. Lastly, we designed a novel way of detecting higher-level motion transitions. With that, we extended the RARMT framework to allow for motion transition recognition, enabling us to recognize a wider range of human activities (i.e. not limited to basic low level motions). Thus, providing the system with the extensibility to support many other activity recognition applications like those that can handles health-critical situations (e.g. falling down motion). With the RARMT framework, we believe that it can provide a robust foundation for many more exciting technological innovations that hinges on reliable human motion recognition in the future.