FPGA-Accelerated Implementation and Optimization of Oriented FAST Features for SLAM

This paper proposes an FPGA-accelerated design and optimization framework for the front-end Oriented FAST (oFAST) feature extraction in ORB-SLAM. To address the high computational cost and poor invariance to scale and rotation inherent in traditional FAST algorithms, a multi-scale Gaussian filtering and Difference of Gaussian (DoG) enhancement scheme is integrated into the preprocessing stage. An adaptive binarization mechanism, combined with morphological operations, generates a sparse mask to isolate salient regions, enabling localized and efficient FAST detection. On the hardware side, separable Gaussian filtering and incremental reuse strategies are adopted to reduce resource consumption, while non-maximum suppression is employed to retain only the most responsive keypoints. Experimental evaluations on the Zynq-7020 FPGA platform demonstrate that the proposed design achieves latencies of 1.54 ms and 2.33 ms on the TUM and KITTI datasets, respectively, significantly outperforming existing SLAM accelerators in terms of delay and frame rate.