| Visual impairment may be caused by various factors varying from trauma, birth-defects, and diseases. Until today there are no viable medical treatments for this condition; hence bio-medical approaches are being employed to overcome that. The Cortivision team has been working on an intra-cortical implant that can bypass the retina and optic nerve and directly stimulate the visual cortex. In this work we aimed to implement a modular, reusable, and parameterizable object recognition system that tends to "simplify" video data prior to stimulation; hence opening new horizons for partial vision restoration, navigational and even recognition abilities.;We identified the Scale Invariant Feature Transform (SIFT) algorithm as being a robust candidate for our application's needs. A multithreaded software prototype of the SIFT and Lucas-Kanade tracker was implemented to ensure proper overall operation. The feature extractor, difference of Gaussians (DoG) part of the SIFT, being the most computationally expensive, was migrated to an FPGA implementation due to the real-time restrictions that is not achievable on a host machine. The VHDL implementation is highly parameterizable for different application needs and tradeoffs. We introduced a novel architecture employing the sub-kernel trick to reduce resource usage compared to preexisting architectures while still being comparably accurate to a software floating point implementation. In order to alleviate transmission bottlenecks, the system also includes a new parallel Huffman encoder design that is capable of performing lossless compression of both images and scale space image pyramids taking into account spatial and scale data correlations during the predictor phase. The encoder was able to achieve compression ratios of 27.3% on the Caltech-256 data-set. Furthermore, a new camera and fiducial markers setup based on image processing was proposed in order to target the phosphene map estimation problem which affects the quality of the final stimulation that is perceived by the patient. |
