Visual learning of a spatio-temporal model of cyclic human locomotion for recognition and tracking

Miha Peternel (2004) Visual learning of a spatio-temporal model of cyclic human locomotion for recognition and tracking. MSc thesis.

Preview

PDF Download (6Mb)

Abstract

Visual tracking is a field of computer vision research that investigates techniques of object tracking in time sequences of images. These techniques can be applied to people tracking, gait recognition and activity recognition. We envision an autonomous system which is able to learn appearance dynamics of humans from past video recordings and apply the information learned to assist recognition and tracking in new recordings. Human body is a relatively complex articulate object with a multitude of possible poses and motions, making for a large database of learning images, therefore we need a model that compactly abstracts the motion in a video sequence, but doesn't discard individual characteristics. The main motive is the development of an algorithm which estimates the parameters of a model based on the video recordings of human locomotion in partially controlled environment and uses the learned model on new video recordings in arbitrary environment for recognition or tracking. This thesis presents a novel representation of cyclic human locomotion based on a set of spatio-temporal curves of random tracked points on the surface of a person. We start with an algorithm for tracking of multiple random points on a person in a monocular video sequence, followed by a method to determine the cycle interval, align repetitions and extract a set of continuous, phase aligned spatio-temporal curves. We analyze a PCA representation of the cyclic curves, pointing out properties of the representation which can be used for spatio-temporal alignment in tracking and recognition tasks. We model the curve distribution density by a mixture of Gaussians using Expectation-Maximization algorithm. For recognition, we use maximum aposteriori likelihood estimate combined with linear data adaptation. We tested the algorithms on CMU MoBo database with favorable results for the recognition of people identity and locomotion mode from monocular video sequences captured from the side view. We conclude with discussion of results based on laboratory and outdoor testing and propose extensions of the method.

Actions (login required)