Research in assistive healthcare, in particular home rehabilitation, has spawn huge potential owing to the recent advancement of internet-of-things technology and the wearable hardware, Inertial Measurement Unit (IMU) in wearable sensors and smartphones become a affordable for community usage. However, using low cost IMU sensors or smartphones face certain challenges, such as accurate orientation estimation for lower-limb motion tracking, which is usually less of a problem in specialized motion tracking sensor devices. To address these issues, the candidate has made three main contributions: a new and better orientation estimation algorithm which combines quaternion-based Kalman filter with corrector estimates using gradient descent (KFGD), an auto-detector of post-filtered lower-limb orientation signal oscillation and the machine-learning based state identification of rehabilitation exercise. Firstly, obtaining accurate orientation readings with noise-prone IMU and post-processing drift is a key challenge in motion tracking research. It is the result of accumulated errors over the integration of the gyroscope signal to calculate the angular displacement, in other words, the orientation of the limb, in the motion tracking application. Thus, the candidate proposes two sensor fusion algorithms: the complementary filter feedback (CFF) and the quaternion-based Kalman filter with corrector estimates using gradient descent (KFGD). The complementary filter feedback (CFF) focuses on the components’ performance of high-pass filter (from angular velocity) and low-pass filter (from fusion of gravity and earth magnetic field). These components contribute to the estimated orientation while the proposed feedback loop can correct the drift. KFGD is later introduced to further improve the limitation of the low-pass filter and the fixed fusion threshold of the CFF. Gradient descent method and quaternion-based Kalman Filter are chosen for their progressive features. The performance was evaluated on the case study of early stage rehabilitation exercises, namely, leg extension and sit-to-stand. The result shows that CFF is capable of fast motion tracking and confirms that the feedback loop is capable of correcting errors caused by integration of gyroscope data. KFGD outperforms the state-of-the-art Madgwick algorithm and is recommended for obtaining accurate orientation readings using motion sensors. Secondly, upon observing the characteristics of the post-filtered orientation signals of the lower-limb, a noticeable artifact in the output signal that it would oscillate from positive to negative and vice versa. To address the oscillations in the signals of both motion capture and inertial measurement sensors, the candidate applied machine learning algorithms and compared them with the rule-based approach. Machine learning methods, such as Logistic Regression, Support Vector Machine and Multilayer perceptron, were adopted in order to automatically detect the oscillation. The results showed that machine learning methods are able to learn the oscillation patterns in wearable sensor data and identify the tendency of fluctuation thereby allowing the errors to be filtered out more efficiently than rule-based method. Lastly, in order to realize meaningful home rehabilitation, there is a need for informative feedback or intervention in parallel with the exercise monitoring. The study aims to use the collected data and the understanding of wearable signal to simulate the high-level observations by the physiotherapist towards the patients and provide informative feedback during exercising at home. Therefore, the candidate proposes the study on machine-learning based state identification of rehabilitation exercise by using wearable sensors on the lower limbs. The informative feedback and quality assessment could be obtained by selectively segmenting the exercise into four states: rest, raise, hold and drop. The segmentation potentially increases the frequency of detection resulting in almost real-time feedback. In addition, identifying the abnormal sequences against the correct pattern in the respective state results in more specific and informative feedback. In this work, the candidate analyses the impact and derives valuable insights of the extracted sensor signals in relation to the predicted. As a result, the predictive model yields up to 95.89% (SVM) and 94.04% (SVM) accuracy for binary and multi-label pattern recognition respectively. The experiment and recommended framework show the efficiency and potential of using signal data as features in motion-based exercise pattern recognition. The work presented in this thesis demonstrates the realization of home rehabilitation from the hardware-level to the simulation of user intervention. The methodologies exploit the a ordable hardware to correctly track the limb motion while the motion signal prediction model and analysis boost the potential of intervention strategy for the user’s home exercise feedback.