A novel multi-intensity image labeling algorithm for real-time computer vision and robotics applications

Abstract

This paper presents a novel multi-intensity image labeling algorithm tailored to provide high throughput for real-time computer vision and robotics applications. The algorithm overcomes and alleviates many drawbacks of the earlier labeling techniques. With the aid of the newly introduced pre- and post-equivalence solving stages, it assigns each pixel the correct label value with the minimal possible accesses, hence approaching the maximal processing rate. With its efficient algorithmic design, it does not require buffering the entire image and uses the minimal possible memory space required to perform the multi-intensity labeling task. It was designed in a way to reduce the required conditional statements (e.g. comparisons) that stall the labeling pipeline and prevent parallelizability. These merits make it very suitable for hardware design (e.g. GPUs). Testing results prove that our algorithm competes with the world's fastest labeling algorithm, and is 2×~3× times faster in its worst case scenario.