ROPH: A Robust, Optimized, and Parallelized Harris Detector with Flexible FAST-Based Pruning

The Harris corner detector is a foundational algorithm in computer vision, valued for its geometric precision, though its throughput can be a bottleneck for real-time processing. Numerous optimizations proposed often focus solely on accelerating corner detection, creating a trade-off between speed and accuracy, and neglecting the detector’s original capability to find both corners and edges. This paper introduces a novel optimization framework, Robust, Optimized, and Parallelized Harris (ROPH), that addresses these limitations. Our key contribution is a parallelized, two-stage pipeline. The first stage uses a configurable, FAST-based pruning mechanism to intelligently filter pixels, drastically reducing the workload for the subsequent Harris stage. Our method not only achieves a significant average speedup of 6.73× (and a peak of 9.24×) over a standard baseline but also enhances robustness. A key finding is that ROPH’s intelligent pruning provides a feature map that is cleaner than the noise-sensitive baseline and more structurally coherent than the output of traditional Non-Maximum Suppression (NMS). By maintaining perfect precision and restoring the full feature-finding capability of the original algorithm, ROPH provides a versatile and efficient solution for modern computer vision applications on resource-constrained devices.