Analysis and Classification of Object Poses - Using Visual/Infrared Images and Feature Fusion

Abstract

This master thesis addresses issues in computer vision and pattern classification. More specifically, we are interested in classification of various object poses from images, for examples, poses of human faces or cars. Analysis and classification of visual object poses are important steps towards different applications, e.g., surveillance and traffic safety. In this thesis work, several feature extraction methods are implemented, including HOG (histogram of oriented gradients) and Gabor features. A multi-class object classifier based on multi-class AdaBoost is implemented. Experiments have been conducted on large numbers of face images and car images with different poses. For human faces, the classifier contains 5 classes of poses (frontal, left, right, upward and downward), while for car images the classifier contains 4 classes of poses (frontal, rear, left and right). Two types of images are tested: one is from the visual band and another is from the thermal infrared band. Due to different properties and characteristics in these two types of images, different types of features are extracted. For visual band images, HOG is used as the main feature descriptor. For thermal IR images, Gabor features are used. For classification of object poses, classifiers are tested separately by using visual band images only, and thermal IR images only. Performance is then evaluated for these two types of classifiers. Attempts are also made on classifiers through fusing these two types of features in visual and IR images. Results will be presented and future work will be discussed.