Yu-Hsun Lin MSc Dissertation 2014/15

Gesture Recognition using Kinect and Machine Learning

Supervised by N.Lawrence

Abstract

This project aims to solve the gesture recognition problem based on ChaLearn Gesture Challenge (ChaLearn 2011), an one-shot learning challenge which means we have only one training sample for each given gesture. The original motivation to conduct this research is to shorten the gap between the performance of human and the computers. We proposed the use of two latent variable models, Probabilistic Principal Component Analysis (PPCA) and Bayesian Gaussian Latent Variable Model (BGPLVM) to model the high-dimensional video data, with two feature extraction techniques, Histograms of Oriented Gridents (HOG) and Histograms of Optical Flow (HOF) to emphasis the features. By using the class conditional density model we compare the value of test likelihood for new data in terms of classification. Our best result (33.67%) is performed by BGPLVM with linear kernel which is similar to that of using PPCA. Although our result could not outperform the results of previous challenge participants, it could be ranked top-15 of all published results in the first 20 video batches.