关键词: 肉鸽养殖场, 机器人, U-Net, 视觉导航, 语义分割, 直线拟合

Abstract: In order to address the issues of difficulty for visual navigation in indoor breeding farm in distinguishing between the main road and background environments, as well as interference from pigeon feathers and feces, a UNet networkbased visual navigation path recognition method for meat pigeon breeding farms is proposed. The road datasets are created by using mapping software, and the corresponding training datasets of convolutional neural networks are selected to generate semantic segmentation models. The pixel coordinates of the edge points are obtained based on the generated segmentation mask area,  and the fitting navigation points are obtained by the equal proportion navigation method, and the navigation lines are formed by using the least squares fitting line principle to fit the navigation points. The experimental results show that the accuracy, intersection to union ratio, and recall of the test set trained for generating the optimal semantic segmentation model are 98.48%, 96.21%, and 99.05%, respectively. When the road width is 1 m, under different lighting and road conditions, the average lateral standard deviation and lateral absolute deviation of autonomous navigation are 0.019 5 m and 0.027 m, and the average heading standard deviation and heading absolute deviation are  1.657° and 1.95°. In the field environment of a pigeon house, this method can achieve the effective segmentation of roads in images, providing a technical solution for visual navigation path recognition of breeding farm robots.

Key words: meat pigeon breeding farm, robots, UNet, visual navigation, semantic segmentation, straight line fitting