Landscape environment space extraction based on YOLOv5 and distance information metrics

Abstract

The current methods for extracting spatial information in classical gardens suffer from low extraction accuracy, low efficiency, and difficulty in dealing with complex and ever-changing situations. Given this, the paper proposes a garden environment space extraction method based on You Only Look Once version 5 (YOLOv5) and distance information measurement. The research first introduces context transformation and multi-scale strategies to improve the YOLOv5 algorithm. Subsequently, the research combines the target detection results with distance information. Through camera calibration technology, the target position is accurately calculated to achieve efficient extraction of spatial information. Based on the distance information, obtain the proportional coefficient of the actual length of each pixel and calibrate the system. Calculate the distance between the new position and the original position, and calculate the three-dimensional coordinates of the target based on this distance and the camera parameters. Finally, by using these three-dimensional coordinate information, a three-dimensional spatial model of the garden environment is constructed, thereby achieving the extraction of environmental space. Experiments have shown that when the proposed method was applied to 3D modeling of garden environments, its chamfer distance value was only 0.02, the integrity ratio value reached 0.98, the F1 score was 0.98, the average accuracy value was 0.97, and the frame rate was 28.6. Under different interference scenarios, the average accuracy value was less affected and remained above 94.9%, with a root mean square error of only 0.05. The proposed method for extracting spatial information from garden environments can achieve high accuracy and efficiency in complex and varied garden environments, providing reliable data support for garden design, planning, and management.