With the recent invention of depth sensors, human gesture recognition has gained significant interest in the fields of computer vision and human computer interaction. Robust gesture recognition is a difficult problem because of the spatiotemporal variations in gesture formation, subject size, subject location, image fidelity, and subject occlusion. Gesture boundary detection, or the automatic detection of the onset and offset of a gesture in a sequence of gestures, is critical toward achieving robust gesture recognition. Existing gesture recognition methods perform the task of gesture segmentation either using resting frames in a gesture sequence or by using additional information such as audio, depth images, or RGB images. This ancillary information introduces high latency in gesture segmentation and recognition, thus making it inappropriate for real time applications. This thesis proposes a novel method to recognize time-varying human gestures from continuous video streams. The proposed method passes skeleton joint information into a Hidden Markov Model augmented with active difference signatures to achieve state-of-the-art gesture segmentation and recognition.

Active body parts are used to calculate the likelihood of previously unseen data to facilitate gesture segmentation. Active difference signatures are used to describe temporal motion as well as static differences from a canonical resting position. Geometric features, such as joint angles, and joint topological distances are used along with active difference signatures as salient feature descriptors. These feature descriptors serve as unique signatures which identify hidden states in a Hidden Markov Model. The Hidden Markov Model is able to identify gestures in a robust fashion which is tolerant to spatiotemporal and human-to-human variation in gesture articulation.

The proposed method is evaluated on both isolated and continuous datasets. An accuracy of 80.7% is achieved on the isolated MSR3D dataset and a mean Jaccard index of 0.58 is achieved on the continuous ChaLearn dataset. Results improve upon existing gesture recognition methods, which achieve a Jaccard index of 0.43 on the ChaLearn dataset. Comprehensive experiments investigate the feature selection, parameter optimization, and algorithmic methods to help understand the contributions of the proposed method.

Library of Congress Subject Headings

Computer vision; Image processing--Digital techniques; Patter recognition systems--Mathematics; Hidden Markov models

Publication Date


Document Type


Student Type


Degree Name

Computer Engineering (MS)

Department, Program, or Center

Computer Engineering (KGCOE)


Raymond Ptucha

Advisor/Committee Member

Andreas Savakis

Advisor/Committee Member

Nathan D. Cahill


Physical copy available from RIT's Wallace Library at TA1634 .K86 2014


RIT – Main Campus

Plan Codes