A fusion-based multiclass AdaBoost for classifying object poses using visual and IR images
Mohamed Hashim Changrampadi · Chalmers Publication Library (Chalmers University of Technology) · 2011
Visual object recognition and classification have wide applications in many areas, e.g., medical, consumer, surveillance.Many investigations have been done in this field, yet there exists great potentials of improvement in order to bridge the gap between computer vision and human recognition system.In this thesis, human faces are considered as the object of interest.The motive of this thesis is to classify object poses in visual and thermal infrared (IR) images, and then apply a fusion technique to improve the classification rate.A thermal IR camera is used to capture thermal images, creating a thermal IR face pose database.Both IR and visual face images are used for training and testing.A new approach, multi-classifier AdaBoost, is proposed for multi-class classification.It is used to classify object poses in visual and IR images.Though individual classification of visual or thermal IR images achieves a classification rate around 95%, the efficiency is further improved with the fusion of these two types of images (> 98%).The classification method is also tested for another database.Viewing that a thermal IR camera is an expensive solution for consumer applications, alternative equipment such like Near-IR webcam and Kinect are investigated and tested in this work.Using images captured by a self-made Near-Infrared (NIR) web-camera or Kinect, our preliminary tests indicate the proposed classifier results in approximately similar performance using such images.Acknowledgement All Praise and thanks to Almighty Allah who has blessed me with innumerable gifts and for providing me a lust for research that has ended me in Chalmers.I utilize this opportunity to thank Chalmers University of Technology and in specific Department of Signals and Systems for sharing the valuable resources and laying down the foundation for my thesis.I foremost express my deep gratitude to my supervisor and examiner Prof. Irene Gu , who has lead me