Research On Visual Navigation System Of Greenhouse Inspection Robot

In plant seedling rearing,the application of advanced technologies such as computer vision and artificial intelligence makes it possible to nondestructively diagnose seedling growth conditions,providing key real-time data for precise environmental regulation based on plant growth and development.Inspection robots are a more flexible and economical solution than the extensive deployment of sensor networks in greenhouses.The inspection robot needs a kind of automatic navigation technology suitable for the large area and complex structure of the seedling greenhouse environment.This paper proposes a visual navigation solution for the seedling greenhouse inspection robot,which can realize the automatic cruise and obstacle avoidance function in the seedling greenhouse environment.Specific research contents and conclusions are as follows:(1)An image segmentation algorithm for greenhouse pavement based on VGG16-Unet is proposed.Aiming at the special environment of seedling greenhouse,a road depth image segmentation method based on Otsu threshold was proposed firstly,which can extract the near information of greenhouse road surface well,but has poor robustness to illumination.In order to solve the influence of different lighting on pavement segmentation,a pavement color image segmentation method based on deep learning was proposed.Firstly,the greenhouse pavement images under different lighting were marked,expanded and divided with label sets.Then,Mask-rcnn network and improved VGG16-Unet network were used to train the training set.The trained model is used to segment the prediction set.The experimental results show that VGG16-Unet has a better segmentation effect on the greenhouse pavement image under different lighting,with the accuracy of 99.22% under strong light and 99.38% under weak light,which lays a foundation for the subsequent navigation line detection.(2)A visual navigation scheme for seedling greenhouse was constructed.First of all,the walking sequence of the inspection robot is constructed for the seedling greenhouse environment.According to the walking sequence and the current road condition,the robot can realize the switch of three modes: straight running,turning and turning.In the straight mode,Hough transform method and least square method were used to fit the pavement boundary.The least square method could achieve better fitting effect,with the average transverse deviation of only 3.8pixel and the Angle deviation of only 0.62°.The detection accuracy reached 98%,and the average time was only0.026 s,which could meet the accuracy requirements.Then,the camera projection model was used to solve the theoretical conversion accuracy,and the road boundary was converted into the navigation line of the inspection robot,so as to realize the extraction of navigation parameters.(3)Obstacle recognition and obstacle avoidance algorithm of greenhouse inspection robot.In order to avoid the collision between the greenhouse inspection robot and obstacles in the process of inspection,an obstacle recognition algorithm was proposed,and the VGG16-Unet pavement image segmentation algorithm was optimized through the transfer learning method.The segmentation accuracy of pavement image with obstacles reached 99.25%.The depth map containing only obstacle information is extracted through the mask,then the depth image is transformed into a three-dimensional point cloud,and the statistical filtering is adopted to filter the point cloud.Finally,the point cloud containing only obstacle information is obtained.The minimum surrounding box is obtained for the point cloud,and the height of the obstacle can be estimated.The height estimation error is less than 2.6cm,which meets the accuracy requirement.The threshold is set as the chassis height.When the obstacle is less than the threshold height,the obstacle can be crossed.When the obstacle is greater than the threshold height,the walking will stop and wait for the obstacle to be cleared manually.(4)Human-machine software design and walking performance experiment of greenhouse inspection robot visual navigation system.Py Qt5 is used to establish the human-machine software of the vision navigation system of greenhouse inspection robot,which can realize the function of manual control of inspection robot and autonomous walking of seedling bed.In addition,walking experiments were carried out for each walking mode of the inspection robot.In the straight mode,initial transverse deviations of 0mm,100 mm and 200 mm were set,and the inspection machine could reach stable running after a period of time,and the average transverse deviation under the stable running state was only 12.2mm.Setting three different speeds of0.06m/s,0.09m/s and 0.12m/s,the inspection machine can achieve stable running per capita,among which 0.06m/s has the best navigation effect and the average lateral deviation is only 9.8mm.In the steering mode,the average distance between the inspection robot and the edge of the seedbed after completing the steering was 21.8cm,and the average yaw Angle was 3.6°.In the U-turn mode,the average distance between the U-turn position of the inspection robot and the front end of the seedling bed was24.9cm,and the average yaw Angle after U-turn was 3.4°.According to the size of the obstacle,the inspection robot can stop to recognize the obstacle within a safe distance of 0.95 m,and realize two control commands of obstacle crossing and stopping according to the height of the obstacle.To sum up,this paper constructs a complete set of greenhouse inspection robot visual navigation system.Through the greenhouse pavement image segmentation,visual navigation scheme construction,obstacle recognition and obstacle avoidance,the automatic greenhouse inspection function of the greenhouse inspection robot is realized,and the navigation system is tested and verified to meet the requirements of greenhouse inspection.The visual navigation system has the advantages of low cost,high precision and strong real-time performance,which provides reference value for greenhouse agricultural navigation.