ScholarWorks@성균관대학교

초록

Surround view monitor (SVM) systems are at the heart of advanced driver assistance systems (ADAS), and SVM technology has recently been actively researched in autonomous driving applications beyond parking. Preprocessing to build SVMs is difficult to implement in CPU or embedded environments because it requires processing large amounts of video information, and SVMs using traditional FPGAs are lookup table (LUT)-based implementations that are real-time but difficult to compensate for camera damage or position changes. The architecture proposed in this work performs real-time, pixel-by-pixel inverse mapping using dynamically modifiable parameters. The inverse mapping also handles fish-eye distortion correction, bird's-eye view (BEV) transformation, and rotate to reduce memory usage and enable high-speed processing. The system operates at 200 MHz and processes 4-channel video with 960x640 resolution at over 30 frames per second with minimal image loss. We also implemented SVM with post-processing in PS. We show that the proposed structure is a very flexible real-time calibration and transformation architecture for SVMs and is a suitable solution for next-generation SVM and ADAS applications.

키워드