Design Of Autonomous Landing System Of Quadrotors On Mobile Platform Based On Instance Segmentation Networks

The autonomous landing of a quadrotor unmanned aerial vehicle（UAV）on a movable platform is one of the hotspots in the collaborative research of air-ground robots.With the development of machine vision algorithms and the improvement of computing power of edge processors,it has become a feasible solution to guide drones to land autonomously through vision sensors.In this paper,an autonomous landing system of quadrotors on mobile platform based on instance segmentation networks is designed.The research work mainly includes the following aspects.First of all,in view of the problem of insufficient aerial field of view of UAVs,an airborne gimbal camera is used to replace the usual airborne fixed camera,and a double closed-loop control system structure including a gimbal target tracker and a body gimbal tracker is proposed.It can actively search for the location of the landing platform to improve the success rate of autonomous landing of the UAV.Second,a new type of nested composite landing platform sign that is not based on any universal identification code is proposed,which can provide eye-catching visual signals and effective landing yaw angle guidance information at the same time.Solved the problem of deviation between the physical position of the landing platform relative to the UAV and the pixel position of the landing platform sign relative to the image.Third,an image sequence instance segmentation network using mask cropping and optical flow estimation is proposed for the recognition of landing platform sign.In order to eliminate image-independent motion information,PWC-Net is directly used to predict the inter-frame optical flow motion field before instance segmentation,and based on pixel-by-pixel The optical flow motion distance uses the segmentation result of the previous frame to crop the current frame image as an additional input to the instance segmentation network;in order to optimize the segmentation result,the predicted inter-frame optical flow field is used to reverse the segmentation result of the previous frame to the current frame.The frame positions serve as additional network supervision,thereby improving the segmentation accuracy of the current frame.This will reduce the waste of inter-frame information,and can suppress the instability phenomenon of landing platform sign segmentation in the image sequence.Fourth,a multi-period and multi-stage quadrotor UAV autonomous landing visual servo strategy based on active disturbance rejection control（ADRC）algorithm is proposed.The ADRC algorithm is used to design the UAV airborne pan-tilt tracking control system and the body attitude control system including five control loops,so as to improve the tracking response speed of the UAV to the mobile landing platform.Finally,the landing platform sign dataset was collected and the instance segmentation network model was evaluated.The mean intersection over union（m Io U）reached 92.1%,the F1 evaluation reached 89.3%,the time consistency（TC）reached92.9%,the mean average precision（m AP）reached 77.5%,and the FPS reached 26 on the onboard computer DJI Manifold.The DJI M100 quadrotor UAV was used as the flight platform to carry out the UAV autonomous landing system experiment.According to the different motion states of the landing platform,the autonomous landing experiment of the fixed landing platform and the autonomous landing experiment of the mobile landing platform were carried out respectively.The experimental results show that the vision scheme proposed in this paper can output smooth landing platform sign recognition results,and achieve a balance between accuracy and speed on the airborne computer,which verifies the effectiveness of the quadrotor UAV autonomous vehicle landing system in this paper.It can meet practical application needs.