Utilizing spatial locality of ColourFAST features for GPU-Accelerated object recognition
ColourFAST is an alternative technique to FAST developed by Ensor and Hall used to extract feature point descriptors from an image based on colour change values. The extracted descriptor is compact and, therefore, efficient to compute and match. The purpose of this thesis is to extend the Colour-FAST feature descriptor from a 4-dimensional vector to a 6-dimensional vector to improve feature point matching accuracy. This is achieved by incorporating spatial locality to gain a sense of the shape of an object alongside its colour change information. The main focus is designing, developing and testing feature point matching algorithms specifically architected for the GPU pipeline with an emphasis on accuracy while maintaining high throughput.