TY - THES A1 - Klaumann Gubert, Gabriel Vitor T1 - Evaluation of the impact of the posit quantization on the inference performance of a deep learning neural network for 3D object detection and classification with RADAR and camera N2 - The evolution of autonomous driving is influenced by the convergence of technologies shaping the capabilities of perception systems. This study reflects the interaction between deep learning, specialized hardware accelerators, and advanced numerical representations, specifically Posits, in the context of CenterFusion - a middle-fusion method for fusing radar and camera data designed for autonomous vehicles. In order to optimize the CenterFusion for embedded systems, a parameter quantization and hardware inference technique using Posit numbers was developed, named Posits4Torch and Posits4TorcHA. Posit is a type of floating-point numerical representation and arithmetic whose fields have variable sizes, resulting in non-uniform precision. The main objective of this study is to evaluate the reduction in memory usage, processing speed gain, and degradation of CenterFusion performance, when subjected to Posit quantization, in terms of the average errors and the NuScenes Detection Score. The results demonstrate that the inference errors for both the Deep PeNSieve-emulated model on an AMD EPYC 7413 24-Core CPU and the Posits4TorcHA-based model on an AMD Kria KV260 Vision Starter Kit FPGA were identical. The FPGA-based model's inference time of 3,49 seconds was nearly 1000 times lower than the 3194,69 seconds of the CPU model but 300 times higher than the PyTorch GPU model, which ran on a NVIDIA GeForce RTX 3090. Moreover, FPGA implementation metrics for the 2D 8x8 Posit MAC Unit Array on the AMD Kria KV260 showed a power consumption of 2,939 W, 3,033 W, and 3,090 W for Posit precisions of 6, 7, and 8 bits, respectively, highlighting the approach's resource efficiency. When quantizing CenterFusion's regression heads, an 8-bit Posit precision for Posit-as-Storage (PaS) resulted in a degradation in the Mean Average Precision (mAP) and NuScenes Detection Score (NDS) of, respectively, 0,7% and 0,5%. For Posit-as-Arithmetic (PaA) on the FPGA, certain metrics like Mean Average Velocity Error (mAVE) and Mean Average Translation Error (mATE) were more affected, with 8-bit precision leading to degradations in mAP and NDS by 9,8% and 7,7%, respectively, which can be improved by training the model with a higher Posit precision and subsequently quantizing to a lower Posit precision using Posits4Torch. This approach presents a promising perspective for optimizing neural network performance in perception and sensor fusion systems for autonomous vehicles. By improving efficiency using lower bit-widths without compromising too much accuracy, this research contributes to the development of more energy-efficient high-speed AI solutions for autonomous vehicle's embedded systems, demonstrating ongoing progress in the field of autonomous driving technology. Key-words: posit, autonomous driving, hardware acceleration. Y1 - 2024 UR - http://nbn-resolving.de/urn/resolver.pl?urn:nbn:de:bvb:573-50204 CY - Ingolstadt ER -