High-performance trajectory planning: A GPU-acceleration performance study

Abstract

Automated driving technologies and advanced driver assistance systems (AD/ADAS) have been a popular research topic since the automotive industry started pursuing software-defined vehicles. An instrumental part of AD/ADAS is the trajectory planning algorithm, which decides the trajectory for the given traffic environment. In recent years, trajectory planning algorithms have improved in both run time and trajectories. While the algorithmic improvements have been apparent, there has been a lack of research on the suitability of parallelization and graphical processing unit (GPU) acceleration. Targeting the GPU is also highly relevant due to the increase of GPUs in a vehicle’s computer architecture. This paper implements a spline-based trajectory planning algorithm in C++ for a single-core central processing unit (CPU), multicore CPU, and GPU-accelerated platform. Implementations were tested on a relevant automotive computing platform for accurate comparisons in a realistic scenario. Ultimately, this thesis concludes that the GPU-accelerated implementation is better in every aspect measured and, in some cases, achieves a speedup of 2 to 3 orders of magnitude. Due to the much higher throughput, more solutions could be generated in a real-time scenario, leading to safer trajectories overall.