FIRDAUS BIN MOHD NAZRI
I major in Artificial Intelligence and minor in Mathematics in my Bachelor's Degree in Computer Sciences. I am deeply passionate in learning new techniques and employing more efficient algorithms in the projects that I am involved in. I believe that Machine Learning is becoming a norm in our daily life and thus is vital that industry leaders to employ them in the business based on their respective needs to be ahead of their competitors.
Matrix No:
137067
Student Email:
Supervisor:
Dr. Ahmad Sufril Azlan Mohamed
Supervisor Email:
SC016
Markerless Motion Capture Using Deep Learning
The purpose of this project is to develop a system that is able to track the body movement of a person from a video source while augmenting the labelled skeleton joints onto the body of the person. This project has endless applications in the real world especially in the physical-demanding working environment as well as in the sports industry.
This project is implemented using deep learning techniques, specifically by using a pretrained model from Google's mediapipe BlazePose which is able to automatically detect a person's joints. The model is trained from an image dataset containing around 85,000 images including 30,000 of the images obtained from consented images of people using a mobile AR application captured with smartphone cameras in various 'in-the-wild' conditions.
The outcome from this project is a working system that is able to correctly identify and label the skeleton joints on a person's body as well as perform various calculation such as movement velocity and the angle of joints. This information is invaluable for researchers and clinicians to identify potential risks of muscle strains and stress. In addition, the system will also ease them in processing 2-D video images without the need of expensive Motion Capture System.
An Inertia Measuring Unit bodysuit (Rokoko) was used as the benchmark and results show that the system exhibits a 12.24%, 7.89% and 8.65% deviation at torso, arms and legs respectively. The results justify that the system is comparable to other 3-D Markerless-Based system and has approximately less than 15% mean difference.