The University of Edinburgh -
Division of Informatics
Forrest Hill & 80 South Bridge

MSc Thesis #9839

Title:Person Recognition from Gait
Date: 1998
Abstract:The aim of this project was to reconstruct a skeleton-based gait analysis system and to investigate neural network based methods for person recognition. Subjects were filmed with twenty-three reflective markers attached to certain anatomical landmarks. A three-dimensional motion caputre system was used to acquire the video sequences of the walks of the individuals. The joint trajectories of the individuals were extracted from the video sequences by using the software package EVaDemo. Principal Component Analysis was then applied to the data to align the walks of all the individuals with one of the axis. The data was then transformed using the Fast Fourier transform. The Fourier components of all the joint trajectories are considered as the feature vector of each subject. To reduce the dimensionality of the data, the correlation coefficients of all the marker trajectories were calculated and some markers were subsequently ignored. A clustering technique was also performed on the first few Fourier components to determine whether the Fourier components of the subjects were distinguishable. Finally, a two-layer feed-forward neural network was trained (using error back-propagation) and tested on the fundamental and next four harmonics of the selected markers. The results obtained showed that the neural network could only recognise two subjects among all the subjects considered.

[Search These Pages] [DAI Home Page] [Comment]