Traditional methods of Human Activity Recognition (HAR) and gait analysis are being
done using computer vision techniques and in an expensive motion-capture laboratory under a
controlled environment. Although this provides accurate results, these techniques are prone to
privacy and pervasive issues. The alternative is wearable systems that employ standalone
accelerometers, gyroscope sensors, and inertial measurement units (IMU), and inbuilt ones in
smartphones and smartwatches. However, accelerometer-based sensor systems have
limitations in terms of position localization (e.g., accelerometer on the wrist will not be able to
predict ambulatory motion accurately), recognizing complex activities involving multiple
actions at the same time (e.g., eating while sitting), drift due to the double integration to arrive
at the position information, and noisy readings.
The research work reported in the thesis focuses on developing three new wearable
sensing systems and approaches that are drift-free with improved accuracy and a better
recognition capability. The work also includes formulating a robust methodology and
algorithm for gait event identification and on developing a more robust and efficient AI based
drift free wearable systems for human activity recognition and monitoring. The first sensing
approach includes a wearable gait analysis system measuring the ankle angle and MTC for
continuous monitoring of gait parameters using drift free ultrasonic sensors employed in a
simple triangulation technique.
The other sensing approach includes a novel body-worn suspender integrated with a
strain sensor system that captures the body movement's periodicity, resulting in less noisy
readings with non-localized measurements. Along with resolving the localized sensing, noisier
measurements and identifying complex activities, a suspender integrated with a e-textile based
knit jersey conductive fabric is developed, that provides durability to the wearable system. This
system is washable and designed with low energy sensor data acquisition, processing, and
transmission capabilities. The resulting developed system recognizes fourteen human activities
using machine learning and deep learning algorithms comparable to the accuracy levels
obtained in the other suspender based HAR system. In most classifiers, KDA outperformed
LDA methods. The smart suspender system reduces the need for localized sensors to measure
complex activities at various points and lessen the pre-processing time for smoothening the
noise signals. The practicality of all the systems were verified using appropriate prototypes
built and tested