Lower Limb Gait Activity Recognition

Abstract

Lower limb activity recognition, recently, has shown large demand in health care and rehabilitation fields. Wearable activity recognition system has the ability to monitor patients' activities, especially those need to be followed up such as people suffering from traumatic brain injuries. Furthermore, home-based self rehabilitation can be provided for disabled people using assistive systems where activity recognition plays an essential role. However, sensor data acquisition and preprocessing, data segmentation, feature extraction and selection, training and classification are considered as the research challenges that face any lower limb recognition system. Although the sensor data acquisition step affects the prediction performance, only few studies dealt with the sensors positioning challenge. In this research work, the author targets to determine the lower limb segments with the highest contribution to the activity recognition process. The 3-D kinematics and orientation profile of the lower limb segments are acquired using a sensor network of four Inertial Measurement Units (IMU) spread over the lower limb segments of one leg. Time (statistical) and time-frequency (wavelet components) features are extracted from the segments motion profile (kinematics and orientation). Those features are used in the proposed algorithm for sensor localization, which depends on determining the sensors of the most discriminative features selected by a filter type feature selector. Most of the discriminative features are found to be extracted from the sensors fixed on the thigh segment followed by the foot sensors. The results are validated by using random forest classifier for lower limb activity recognition. The validation process supported the results that the thigh kinematics and orientation data are sufficient to recognize the lower limb activities. The overall recognition accuracy using thigh sensor only is 95.7%, while 97% if both thigh and foot sensors are used, in other words the accuracy decreased by 2.3% and 1%, respectively, compared to the accuracy using the four IMU sensors. To overcome the reduction in the accuracy rate several amendments to the extracted features and classification techniques are suggested. V Keywords