This paper presents a novel feature descriptor called TreeBASIS that provides improvements in descriptor size, computation time,\nmatching speed, and accuracy. This new descriptor uses a binary vocabulary tree that is computed using basis dictionary images\nand a test set of feature region images. To facilitate real-time implementation, a feature region image is binary quantized and the\nresulting quantized vector is passed into the BASIS vocabulary tree. A Hamming distance is then computed between the feature\nregion image and the effectively descriptive basis dictionary image at a node to determine the branch taken and the path the feature\nregion image takes is saved as a descriptor. The TreeBASIS feature descriptor is an excellent candidate for hardware implementation\nbecause of its reduced descriptor size and the fact that descriptors can be created and features matched without the use of floating\npoint operations. The TreeBASIS descriptor is more computationally and space efficient than other descriptors such as BASIS,\nSIFT, and SURF. Moreover, it can be computed entirely in hardware without the support of a CPU for additional software-based\ncomputations. Experimental results and a hardware implementation show that the TreeBASIS descriptor compares well with other\ndescriptors for frame-to-frame homography computation while requiring fewer hardware resources.
Loading....