ScholarWorks@성균관대학교

초록

This paper presents a Zynq SoC-based autonomous driving system with traditional image processing and Deep Learning Processing Units (DPUs). A 1/5-scale kid's electric vehicle was modified and tested on an autonomous driving track. The traditional method with Hough transform achieved ~13.64 FPS but was vulnerable to environmental noise. In contrast, the DPU-based system integrated three optimized models: UFLD (6.29 FPS), YOLOv3-tiny (3.54 FPS original, 85.47 FPS optimized), and YOLACT (2.90 FPS original, 20.27 FPS optimized). Optimizations included quantization, input resizing, and backbone replacement. The results support real-time AI perception on embedded systems through DPU optimization. This work may contribute to future research on AI accelerator-based autonomous driving systems.

키워드