high performance computing on graphics processing units: hgpu.org

In this thesis we design, implement and study a high-speed object detection framework. Our baseline detector uses integral channel features as object representation and AdaBoost as supervised learning algorithm. We suggest the implementation of two approximation techniques for speeding up the baseline detector and show their effectiveness by performing experiments on both detection quality and speed. The first improvement to our baseline classifier focuses on speeding up the classification of subwindows by formulating the problem as sequential decision process. The second improvement provides better multiscale handling to detect objects of all sizes without rescaling the input image. This speed-up builds upon the scale invariance property of image statistics in natural images that offers a powerful relationship for approximating feature responses of adjacent scales. While these techniques are not new itself, to our best knowledge we are the first to combine these into a framework for high-speed object detection. Our detection framework is built from the ground up using a fast GPU implementation. Based on these approximation techniques and the GPU implementation for extracting channel features we report detection speeds of 55 fps on a laptop. In a series of experiments we study the contribution of each component to the overall detection time and the possible change in detection quality due to the approximations. We train and test the detector on our car dataset that was constructed for this work. More specifically we focus on rear-view car detection. However the methods discussed are not limited to this object class.