Parallelization of a Vine Trunk Detection Algorithm For a Real Time Robot Localization System

Developing ground robots for crop monitoring and harvesting in steep slope vineyards is a complex challenge due to two main reasons: harsh condition of the terrain and unstable localization accuracy obtained with Global Navigation Satellite System (GNSS). In this context, a reliable localization system requires an accurate detector for high density of natural/artificial features. In previous works, we presented a novel visual detector for Vineyards Trunks and Masts (ViTruDe) with high levels of detection accuracy. However, its implementation on the most common processing units - central processing units (CPU), using a standard programming language (C/C++), is unable to reach the processing efficiency requirements for real time operation. In this work, we explored parallelization capabilities of processing units, such as graphics processing units (GPU), in order to accelerate the processing time of ViTruDe. This work gives a general perspective on how to parallelize a generic problem in a GPU based solution, while exploring its efficiency when applied to the problem at hands. The ViTruDe detector for GPU was developed considering the constraints of a cost-effective robot to carry-out crop monitoring tasks in steep slope vineyard environments. We compared the proposed ViTruDe implementation on GPU using Compute Unified Compute Unified Device Architecture(CUDA) and CPU, and the achieved solution is over eighty times faster than its CPU counterpart. The training and test data are made public for future research work. This approach is a contribution for an accurate and reliable localization system that is GNSS-free.