基于U-Net网络的果园视觉导航路径识别方法

针对视觉导航系统在果园环境中面临的图像背景复杂、干扰因素多等问题,提出了一种基于U-Net网络的果园视觉导航路径识别方法。使用Labelme对采集图像中的道路信息进行标注,制作果园数据集;基于U-Net语义分割算法在数据增强的基础上对全卷积神经网络进行训练,得到道路分割模型;根据生成的道路分割掩码进行导航信息提取,生成路径拟合中点;基于样条曲线拟合原理对拟合中点进行多段三次B样条曲线拟合,完成导航路径的识别;最后进行了实验验证。结果表明,临界阈值为0.4时,语义分割模型在弱光、普通光以及强光照条件下的分割交并比分别为89.52%、86.45%、86.16%,能够平稳实现果园道路像素级分割;边缘信息提取与路径识别方法可适应不同视角下的道路掩码形状,得到较为平顺的导航路径;在不同光照和视角条件下,平均像素误差为9.5像素,平均距离误差为0.044 m,已知所在果园道路宽度约为3.1 m,平均距离误差占比为1.4%;果园履带底盘正常行驶速度一般在0～1.4 m/s之间,单幅图像平均处理时间为0.154 s。在当前果园环境和硬件配置下,本研究可为视觉导航任务提供有效参考。

Path Recognition of Orchard Visual Navigation Based on U-Net

Aiming at the visual navigation system works in the orchard environment, a visual navigation method based on U-Net for path recognition was proposed. Labelme was used to label the road mask in the collected images and made the orchard dataset. Based on data enhancement, the convolutional neural network was trained to obtain the orchard road segmentation model which could identify the road region. The road segmentation mask was used to get navigation information and generate keypoints for path fitting. Using the midpoints as control points, the navigation path was recognized by multi-segment cubic B-spline curve fitting method. Experiments of semantic segmentation and path recognition were carried out respectively, when the critical threshold is 0.4, the results showed that IoU of semantic segmentation model in weak light, ordinary light, and strong light was 89.52%, 86.45% and 86.16%, respectively. The method of edge information extraction and path recognition could adapt to various visual angles of the road mask and get a smooth navigation path. Under different illumination and visual angle conditions, the average pixel error was 9.5 pixel, and the average distance error was 0.044m. It was known that the width of the road was about 3.1m, so the average distance error ratio was 1.4%. The normal speed of the tracked vehicle in the orchard was mostly 0~1.4m/s, and the average processing time of a single field image was 0.154s. Under the current orchard environment and hardware configuration, it was proved that this method had a good performance in accuracy and real-time. This research can provide an effective reference for visual navigation task in the orchard environment.