GPU accelerated feature algorithms for mobile devices
Mobile devices offer many new avenues for computer vision and in particular mobile augmented reality applications that have not been feasible with desktop computers. The motivation for this research is to improve mobile augmented reality applications so that natural features, instead of fiducial markers or pure location knowledge, can be used as anchor points for virtual mobile augmented reality models within the constraints imposed by current mobile technologies. This research focuses on the feasibility of GPU-based image analysis on current smart phone platforms. In particular it develops new GPU accelerated natural feature algorithms for object detection and tracking techniques on mobiles. The thesis introduces ColourFAST features which contain a compact feature vector of colour change values and an orientation for each feature point. The feature algorithms presented in this thesis process information in “real time”, with the objective on high data throughputs, whilst still maintaining suitable accuracy and correctness. It compares these new algorithms with well-known existing techniques as well as against their modified GPU-based equivalents. The research also develops a new GPU-based feature discovery algorithm for finding more feature points on an object, forming a cluster, which can be collectively used to track the object and improve tracking accuracy. It looks at clustering algorithms for tracking multiple objects and implements an elementary GPU-based object recognition algorithm using the generated ColourFAST feature data.