SLDF: A semantic line detection framework for robot guidance

Abstract

Semantic line is a high-level feature in image segmentation and is of great significance in image understanding. This paper proposes a semantic line detection framework shorten as SLDF in addressing robot guidance or navigation tasks. First of all, this paper cast semantic line detection task as a special object detection task with oriented bounding boxes, and then discusses the diversity between several kinds of label coding categories. There are two alternative backbones provided by SLDF: A two-stage model in pursuit of higher precision, or a single stage model that reaches the best trade-off between speed and accuracy. Besides, a geometry-based loss function is designed for SLDF to exploit the benefits of localization accuracy and brings significant improvement. Finally, considering the lack of a comprehensive line evaluation standard including angle, offset, length and other factors, this paper respectively designs corresponding evaluation function based on the characteristics of frame-independent, frame-dependent and streaming perception model, and then evaluate the performance of our framework on SEL-Dataset, NKL-Dataset. Finally, we migrate the designed model to the agricultural robot and realize the robot guidance task in the farmland.