스테레오 비전을 이용한 합성곱 기반 실시간 깊이맵 생성 기법

초록

In autonomous driving systems, accurate perception of the surrounding environment and appropriate decision-making require reliable acquisition of 3D spatial information. Although LiDAR-based 3D mapping technologies have been widely used, they face several limitations and difficulty in integrating with vehicle platforms. To address these issues, this study proposes a real-time depth map generation method that relies on stereo vision, offering a cost-effective alternative to expensive sensors. The proposed approach effectively detects object boundaries, removes unnecessary non-object regions connected to the background, and calculates depth by directly matching disparity between frame elements within object regions. In the process, a convolution filter is used to enhance pixel uniqueness. This approach achieves superior real-time performance over traditional and deep learning-based techniques without requiring complex computation or costly sensors. Thus, it serves as a low-cost, high-efficiency alternative suitable for autonomous vehicles and embedded vision systems.

키워드