Accelerating robust 3D pose estimation utilizing a graphics processing unit

The spin-image pose estimation algorithm is an accurate method for estimating pose of three-dimensional objects while being both robust to clutter and sensor noise. Unfortunately, the algorithm has a high computational complexity, thus preventing its use in applications that require a robotic system to interact with a dynamic environment. Upon inspection, the spin-image algorithm can be broken down into five portions where a single portion called spin-image matching commands 96% of the computation time in estimating pose. Because, the matching of individual spin-images can be performed independently regardless of order, this portion of the algorithm is ideal for the massively parallel architecture of the graphics processing unit (GPU). This paper introduces a GPU implementation of the spin-image matching portion of the spin-image algorithm which makes no modifications to the spin-image algorithm, thus not compromising its robustness and accuracy. This implementation results in a speed-up in spin-image matching of 515x and a total algorithmic speed-up of 24.6x out of a theoretical maximum of 26.0x over a MATLAB implementation. This GPU implementation extends the use of the spin-image algorithm towards practical real-time robotic applications.