Wearable sensors such as inertial measurement units (IMUs) have been widely used to measure the quality of physical activities during daily living in healthy and people with movement disorders through activity classification. These sensors have the potential to provide valuable information of the movement during the activities of daily living (ADL), such as walking, sitting down, and standing up, which could help clinicians to monitor rehabilitation and therapeutic interventions. However, high accuracy in the detection and segmentation of these activities is necessary for proper evaluation of the quality of the performance for a given activity. In this research, we devised a wearable inertial sensor system to measure physical activities and to calculate spatio-temporal gait parameters. We presented advanced signal processing and machine learning algorithms for accurate measurement of gait parameters from the sensor values. We implemented a fusion factor based method to deal with the accumulated drift and integration noise in inertial sensor data in an adaptive manner. We further implemented a quaternion sensor fusion algorithm for joint angle measurement and achieved less noisy values for static and dynamic joint angles with a fourth order Runge-Kutta method. For classification of daily life activities,we rigorously analyzed and handcrafted sixty-six statistical, and frequency domain features from the accelerometer and gyroscope time-series. This feature set was then used to train several state-ofthe- art and novel classifiers. We designed and trained Decision trees, Support Vector Machines, k-nearest neighbour, Ensemble algorithms during our experiments. Our investigation revealed that support vector machine classifier with quadratic kernel and a bagged ensemble classifier met the required value of accuracy of above 90%. Since hand-crafting of features require substantial domain knowledge, we devised a novel deep convolutional neural network to extract and select the features from the raw sensor signals automatically. For this, we proposed and implemented a CNN with dropout regularization and batch normalization, achieving a 96.7% accuracy proving the superiority of automatically learned features over hand-crafted ones.