Independent walking is a significant challenge for stroke patients, and improving mobility is of crucial importance for them. Supervised walking training is the standard rehabilitative program for this population;​ however, it can often be performed for only a few hours a week due to limited resources. Beyond these short periods, patients have to perform training independently, which can affect the training quality. The biofeedback-based therapeutic training has been shown to expedite the re-learning process for walking training. However, this approach requires dedicated laboratory equipment, which is rarely available at hospitals. As such, there is a lack of a practical motion capture system for walking training post-stroke.

To address this issue, this thesis aimed to develop a novel wearable technology using inertial measurement units (IMUs) to measure walking patterns during therapeutic training sessions. Then, the measured walking patterns can be compared to normal walking patterns, and the differences can be translated into auditory/visual biofeedback for the patient and therapist.

First, we proposed two simple, yet effective, sensor-to-segment calibration procedures, including (1) quiet standing and (at least) ten hip flexion/extension and (2) quiet standing and straight walking for (at least) eight steps. Using these calibration procedures, we transformed the measured quantities such as joint angles from the IMU sensor frame to the segment anatomical frame with high accuracy and repeatability to obtain clinically meaningful parameters.

Second, we performed a comprehensive survey of sensor fusion algorithms (SFAs) for body segment orientation tracking using IMUs. Using SFAs, we can combine the recordings of the accelerometer, gyroscope, and magnetometer embedded in an IMU to obtain an accurate and robust estimate of body segment orientation. This survey identified efficient SFAs in the literature and techniques for obtaining robust performance under various motion patterns and intensities.

Third, we developed a framework for adaptive gain regulation of SFAs. We showed that the performance of SFAs depended highly on their chosen gains, and poor initialization of the gains would degrade their performance. Our experimental study showed that an optimized gain regulation scheme based on switching gains between two/three levels obtained sufficient accuracy. Fourth, we proposed a novel linear Kalman filter and a novel robust extended Kalman filter for orientation tracking with IMUs. We included error sources in the raw IMU readouts in the state vector of our proposed Kalman filters so that the raw IMU readouts could be corrected before orientation estimation. Our experimental study showed that our proposed Kalman filters obtained more accurate and robust estimation in long-duration dynamic tasks. Also, in a benchmarking study, we compared the accuracy and robustness of our proposed SFAs to those of more than 30 SFAs in the literature and identified the most efficient choices for different applications.

Fifth, using the estimated foot orientation in the sagittal plane obtained with our proposed SFA, we proposed a novel real-time algorithm for gait event detection. Foot orientation provides physiologically meaningful features corresponding to our observational recognition of the foot’s initial and terminal contacts with the ground. Our experimental study showed that using our proposed biomechanically meaningful rules and constraints resulted in (1) sensitivity and precision of 100% and (2) a temporal accuracy higher than or comparable with the literature.

Finally, using a single chest-mounted IMU, we developed a novel method for the detection and classification of a wide range of physical activities, including standing, sitting, lying, level walking, and walking upstairs and downstairs. The trunk inclination angle and variation of the gravitational component of the accelerometer readout were used for the detection and classification of postural transitions and walking modalities. Our experimental study showed that the proposed method had higher accuracy, sensitivity, and specificity in detecting postural transitions and walking modalities than other methods in the literature.

Our research outcomes based on the steps above enable us to develop wearable sensor technology for gait training. In the future, a biofeedback control system should be designed to report the measured gait kinematics and the difference between the pathological movement patterns and the targeted normal ones to the patient and therapist. Also, our proposed daily activity recognition technology can reveal the efficacy of the training by assessing the users’ activity in their natural living environmen