Machine Learning Projects
My main focus is on using open-source algorithms and off-the-shelf cameras to determine human wrist kinematics. Designing experiments, creating datasets, and refining model parameters has lead me to a high-degree of success in this. The video illustrates a 3D markerset generated through 2D videos and computer vision techniques. The green segment seen represents markers placed on a subjects knuckles and the blue segment represents markers on their wrist. The orange segment is a connection representing the movement of the hand in relation to the wrist.
Laboratory experiments were conducted to build the dataset that was used in training a wrist angle prediction model. A gold-standard motion capture system was used alongside off-the-shelf cameras to build a dataset that contained marker locations with ground-truth wrist angles. I have successfully achieved a high-degree of accuracy when using both a multi-camera and single camera approach to solving this issue. Through the use of deep neural networks, I was able to achieve an angular value with a mean absolute error of approximately 5 degrees when compared to the current gold-standard.
Another project involved using CNNs to detect faces within photos and then predict the race and gender of the subject. This was completed by using YOLOv5 to detect subject faces within several open-source datasets, and crop out the face without the surroundings. The cropped faces were used to train a pre-trained ResNet-34 network to be able to predict the subject race and gender. Gender classification achieved 99% accuracy and race prediction achieved up to 95% accuracy.
Designed experiments to evaluate machine learning models on a multi-modal human action dataset achieving an F1 score of 0.90 in classifying 21 upper body motions related to biomechanical movement.