Research On Navigation Environment Perception Of Unmanned Ship

With the rapid development of artificial intelligence and deep learning,unmanned and intelligent ships have become one of the main directions of development.As a fully automatic surface robot,the unmanned ship can autonomously navigate in the complex environment to complete important tasks instead of humans.The realization of autonomous navigation depends on the ship's accurate perception of the environment.However,in complex sea conditions and high-speed driving,the preception of the surrounding sailing environment can not meet the requirements of real-time and accuracy.Based on this problem,this thesis focuses on the research of the perceptual system,and proposes the method of target detection based on depth learning,the method of the data fusion of camera and Lidar.So the methods which propose in this thesis improve the perception of unmanned ships in complex marine environment.The main contents of this thesis are as follows:(1)Aiming at the blindness and low charateristic of the trational feature extraction method in the processing of the image information obtained by the visual sensor.This thesis proposes the method of maritime target detection of intelligent ship based on improved Faster R-CNN.Deep learning is adopted to automatically acquire the deep-seated characteristics of other target ships at sea.The method is based on Faster R-CNN for target detection,and uses the residual network with better performance as the framework of classified network.In the process of model training,the difficult negative sample mining strategy is combined to improve the accuracy of model detection in the complex marine environment.(2)Aiming at the limitations to the target information acquired by single sensor.This thesis proposes the intelligent ship's perceptual method based on the sensor's data fusion of Lidar and camera.Realize the unification of space and time of sensor system through coordinate conversion and multithreading.Then the data of effective targets are fused by adaptive weighting algorithm.So as to make the position of obstacles more accurate and improve the accuracy of perceptual system.(3)Building verification platform of sensor data fusion based on ROS(Robot Operating System).The perceptual system of unmanned ships is built by Gazebo,and datas of Lidar and camera are obtained and analyzed.Lidar can obtain the coordinate and distanceof obstacles.Since the monocular camera cannot obtain the depth information of the obstacle,this thesis applies the principle of binocular camera's depth measurement,ORB(Oriented FAST and Rotated BRIEF)feature extraction and pixel matching are performed on the picture information obtained by the left and right cameras to obtain the parallax of the camera,and the distance of the obstacle can be calculated.Finally,the fusion model is verified by sensor data,simulation shows that the fusion model can improve the sensing system's ability to measure obstacles.It provides data guarantee for autonomous navigation of unmanned ships and increases safety of navigation.