1. UKM Learning and Research Repository
2. Master Thesis / Tesis Master
3. Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Title:	Upper body activity recognition with convolutional long short-term memory network and kinematics calculation using inertia measurement unit
Authors:	Lim Xiang Yang (P103213)
Supervisor:	Gan Kok Beng, Assoc. Prof. Dr.
Keywords:	Universiti Kebangsaan Malaysia -- Dissertations Dissertations, Academic -- Malaysia Upper body activity Kinematics calculation Inertia measurement unit Kinematics
Issue Date:	16-Mar-2022
Description:	Human activity recognition (HAR) is the study of identifying specific human movement and action based on images, accelerometer data, and inertial measurement units (IMU). This technique can be applied in the areas of healthcare, rehabilitation, entertainment, and remote monitoring. Human body kinematics is the study of the body's movement and the required for body movement to be conducted. A deep understanding of human kinematics can help in learning the underlying principle for body motion and create specific and designated rehabilitation programs for the patients. Most researchers have applied deep learning technology to recognize low-level activities such as walking, running, and sitting. The pre-trained deep models are mainly intended for image classification, natural language processing, and object detection. There is no common architecture and pre-trained model for the HAR classification using sensor data. In the sensor-based HAR application, most researchers used many IMU sensors to get an accurate HAR classification. The use of many IMU sensors not only limits the deployment phase but also increases the difficulty and discomfort for the users. The imbalanced class distribution is another challenge for the recognition of human activity in real life. In a real-life scenario, the model trained on the imbalanced dataset may predict some of the imbalanced classes with very high accuracy. When a model was trained using an imbalanced dataset, it can degrade model performance. This study proposed a sampling-based method and a loss-based optimization method to handle an imbalanced dataset. A deep convolutional long short-term memory (ConvLSTM) model architecture combining both convolutional neural network and long-term short memory was chosen due to its high performance and better HAR. An upper-body activity recognition model was developed using a deep ConvLSTM network with only five IMU sensors compared to the previous model that used 19 sensors. Based on the experiments, the F1-score showed that the model can classify each activity class with a great improvement after applying both proposed imbalanced dataset optimizations. The results showed that the developed model was able to achieve a good performance of 0.93 in both F1-score and model accuracy using five IMU sensors. Besides, a kinematic evaluation framework was proposed to combine the trained upper body ConvLSTM HAR model with the Madgwick filter to perform sensor fusion. Kinematics metrics, for example, elbow flexion and angular velocity on the upper arm and lower arm were studied using the data output from sensor fusion. In conclusion, the upper body recognition model and kinematics metrics are useful for the assessment of home-based rehabilitation involving activities of daily living.,Master of Science
Pages:	113
Publisher:	UKM, Bangi
Appears in Collections:	Faculty of Engineering and Built Environment / Fakulti Kejuruteraan dan Alam Bina

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.