农用无人空气动力船的复杂环境感知与避障控制方法研究

A remote-controlled agricultural unmanned airboat has been developed in Japan to carry out weeding and fertilizing in paddy field or shallow area. However, it is very difficult for the operator to distinguish the position and attitude of the unmanned airboat in the visual range when facing the large faddy field or shallow area. The author developed the GPS-based navigation system for the unmanned airboat to meet the farming work accuracy. However, when the GPS signal is unstable or the unknown obstacles appear, such the developed navigation system cannot make the decision of obstacle avoidance, which may lead to ship collision. Based on the achieved research achievement, this project continues to study on complex environment awareness and obstacle avoidance control method of agricultural unmanned airboat. To study the online nonlinear kinematic modeling method of the agricultural unmanned airboat, to use the stereo machine vision system studies the obstacle fast awareness, recognition and positioning method, and build the obstacle classification model. Refer to the encounter situations of the unmanned airboat and obstacles, make the decision model of obstacle avoidance and research the response transfer function of the turning system and propulsion system in obstacle avoidance control. In order to improve the safety of autonomous navigation, make the effectiveness evaluation of obstacle avoidance control for unmanned airboat. This project is of great significance to realize intelligent weeding and fertilizing used by an agricultural unmanned airboat in China.

日本农机公司研发的遥控式农用无人空气动力船可以代替人力完成水田除草和施肥作业。但是，面对大区块水田时，操作员在视距范围内很难分辨无人船航行状态。因此，笔者在日本留学期间在业内首次开发了基于GPS的船载自动导航系统，满足无人空气动力船的田间作业精度要求。然而，当GPS收信不稳定或出现未知障碍物时，该导航系统无法进行避障控制。本项目拟在原有研究积累的基础上，以农用无人空气动力船的复杂环境感知和避障控制方法研究为主线，研究无人空气动力船的非线性运动模型在线构建方法，利用立体机器视觉系统，研究在非结构化复杂水田环境中障碍物的快速在线感知、识别与定位，并构建障碍物分类模型。根据无人船与障碍物的不同会遇态势，建立避障策略模型。研究避障控制中的转向系统和变速系统的响应传递函数，并对避障控制和复航效果进行科学评价，以提高无人船自动导航的安全性。该研究对实现大规模水田的智能化除草和施肥作业具有重要意义。

农用无人空气动力船等智能农机可以代替人力完成水田除草和施肥作业。基于GPS的船载自动导航系统，可以满足无人空气动力船的田间作业精度要求。然而，当GPS收信不稳定或出现未知障碍物时，该导航系统无法进行避障控制。本项目以农用无人空气动力船的复杂环境感知和避障控制方法研究为主线，研究了无人空气动力船的非线性运动模型和自动控制方法；利用立体机器视觉系统，基于Yolo算法和Deep SORT算法在非结构化复杂水田环境中实现了障碍物的快速在线感知、识别与定位，并构建了田间障碍物（农人、家畜）分类模型。根据无人船与障碍物的不同会遇态势，建立了基于Double-DQN强化学习算法的农机智能避障模型，以提高自动导航的安全性。该研究对实现大规模水田的智能化除草和施肥作业具有重要意义。