어안 카메라를 이용한 근접 감시환경에서 차량 추적을 위한 학습된 메시 기반 객체 전역 좌표 추정

초록

Artificial intelligence-based methods for detecting and tracking objects in videos, typically represented by bounding boxes, are well established. However, accurately determining the global coordinates of these bounding boxes, especially in environments monitored by fisheye cameras, presents unique challenges. These challenges stem from significant variations in the bounding box's size and position, influenced by the vehicle's direction and its distance from the camera. Traditional camera calibration techniques struggle to accommodate the diverse shapes and sizes of vehicles, making precise global coordinate estimation difficult. This paper introduces an innovative approach that addresses these challenges by estimating the global coordinates of the bounding box through a learned mesh. This mesh, generated from parabolic lines aligned with the calibrated camera's horizontal and vertical coordinates, is specifically trained to minimize errors in a dataset of vehicles traversing a parking lot. The effectiveness of the proposed algorithm was validated by comparing the discrepancies in global coordinates of objects derived from bounding boxes, calculated using the suggested learning mesh model, against those determined with reference to the road surface.

키워드