Abstract

Human motion analysis attempts to understand the movements of human body using techniques found in various disciplines. The movements of human body can be interpreted on a physical level through pose estimation, i.e., static reconstruction of three dimensional (3D) articulated configurations, or on a higher more semantic level through action recognition, i.e., understanding the body’s movements over time. It has a wide array of applications in the areas like gaming, sports analysis, security and surveillance, and healthcare. In the gaming industry, learning human action style and creating character animation from a repertoire of actions is very popular. Gait analysis is a crucial step in many biomedical applications as well as security, surveillance and biometric applications. A plethora of sensors are available to capture the human motion data in various modalities easily and in a very cost effective manner. The sheer amount of data produced by researchers, using such sensors, every day demand for human motion analysis methods that are computationally efficient. This thesis attempts to develop effective techniques, based on computer vision and computer graphics, to solve some of the important problems in application areas of human motion analysis. Specifically, three key application areas, namely, sports activity analysis, surveillance and security, and healthcare are considered. New methods for applications like human style sequence learning, gait analysis, gesture recognition,
and time series matching are proposed.

In the first half of this thesis, the problem addressed is learning and synthesis of structured sport activities with kickboxing as a case study. Monocular video data is used as input and human style sequences are investigated in order to identify higher level sequence style. Main idea is to learn the style embedded in action sequencing and transitions in between, and then to synthesize new sequences for virtual characters.
Widely popular computer vision techniques are employed to obtain the sequence of actions performed in the video and to train a model with this sequence to drive a virtual character. Furthermore, style information embedded in transitions are also obtained from video and is used to create style preserving realistic transitions.

In the second half of this thesis, high dimensional sensor data in different modalities is analyzed. Since human motion data is a high dimensional time series data, a novel shape aware multidimensional time series matching algorithm is developed and tested for a variety of scenarios like noise, missing data, different data modalities, small amount of training as well as testing data and different application domains. Firstly, processes for two new biometric systems, namely, gait and visual password are proposed towards surveillance and security. Secondly, processes for recognition of gestures and activities are proposed towards healthcare.

Extensive experimentations are performed to demonstrate the effectiveness and validity of the various techniques developed in this thesis. Performance of the proposed methods is compared with that of the state of the art methods used in the human motion analysis under best possible conditions. For this purpose, the results obtained are validated using some popular benchmark databases as well as a few in-house created databases. It is shown that, the proposed methods outperform the state of the art in most of the cases.