Human activity recognition and tracking

• Main Author: Quak, June Ren Feng
• Other Authors: Tan Ah Hwee
• Format: Final Year Project
• Language: English
• Published: 2014
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/58966
• Institution: Nanyang Technological University

As the population in Singapore grows older, the health-related problems trend will also be seen more commonly. As there will be an increasing number of elderly people with health-related problems likely to need assistant in daily living, the number of caretakers will soon be outnumbered. Therefore, it is necessary to build smart systems to aid the elderly people in their daily living environment. In order to provide unobtrusive monitoring, every day devices such as smartphones could be used. The more advanced way of data collection for body movement is the use of smartphone sensors; and the common way to build a model for motion recognition includes the use of Hidden Markov Model (HMM). HMM was used in many motion recognition applications. However, many of these applications only included continuous motions (Running, Walking, Standing, and Sitting) recognition. This report focuses on building of a model that is able to recognize continuous and transitional motions. In this project, a model was developed with HMM to classify temporal sequences of sensory data. Human motions are categorized into two categories, the lower-level motions and higher-level motions. Lower-level motions include Running, Walking, Standing, Sitting and Lying; higher-level motions include Sit down, Stand up, Lie down, and Fall. A rule-based system was built upon the HMM to achieve transitional (higher-level) motions recognition. Experiments were conducted in real-time to demonstrate the performance of the motion recognizer. The HMM which can recognize lower-level motions had an accuracy of 99.98% while measuring against the cross-validation methodology and 93.60% during real-time testing. The incorporated rule-based system for recognizing higher-level motions had the accuracy of 76.25%.