| The automatic navigation technology is an important guarantee to realize the intellectualization of agricultural machinery.Aiming at the complex working environment of rice transplanter,this research built a set of low-cost and high-precision automatic navigation system firstly,which realized the automatic path tracking process through the positioning technology combining GPS/Beidou RTK sysytem and inertial navigation system,integrated control algorithm and sectional head turning strategy.Then,a method to acquire navigation map and identify fixed obstacle based on UAV remote sensing,satellite remote sensing,geographic information system and image processing technology was proposed.Meanwhile,after comprehensive analysis and comparison of several commonly used deep learning algorithms,the models of farmland obstacle detection platforms suitable for cloud and mobile terminals were determined individually.The study of remote sensing technology and deep learning made preparation for the function of path planning and obstacle avoidance of automatic navigation system.The main conclusions are summarized as follows:(1)This research built an automatic navigation system based on positioning technology combining GPS/Beidou RTK system and inertial navigation system,which used integrated control algorithm for path tracking and sectional head turning strategies.The system is composed of steering wheel,display,controller,RTK satellite positioning sysytem,front wheel angle sensor,inertial navigation sensor and vision sensor.Data transmission is carried out between subsystems by CAN bus.Among them,eMini 4310 display was used to display data and realize human-computer interaction.SBC6300 X development board was used as the control center of navigation system.C94-M8 P module was produced by U-blox Company and used to aquire real-time position information by GPS/Beidou satellite.DCM260 B high-precision three-dimensional electronic compass was used as inertial navigation sensor to aquire real-time attitude information of transplanter.Electric steering wheel was built to control front wheel rotation,which was made up by KFN-130 hollow reducer,PMC007C6SE-57-L closed-loop micro-stepper motor and STM32F103C8T6 core control circuit board.The vision sensor was composed of DFK 33GP1300 camera and VS-0514-5M lens,which can acquire crop row or obstacle pictures on the forward route.(2)This research determined the parameters of the integrated controller through simulation,and analayzed the accuracy of automatic navigation system through experiment.Firstly,aiming to analyze the static and dynamic positioning accuracy of C94-M8 P module,ranging test,linear motion test and circumferential motion test of its mobile station module were carried out.After data conversion,the results showed that positioning accuracy of this module is about 2 cm,which is close to the standard in the instruction book.Secondly,this chapter completed the design and parameter tuning of the PID-fuzzy control algorithm through matlab/Simulink.The conversion factors of the lateral deviation and course deviation of the input ERROR of the PID control algorithm were 10 and 1,individually.The proportional coefficient,integral coefficient and differential coefficient of the PID control algorithm were 0.8,16 and 0.4,individually.The fuzzy rules were defined and the fuzzy control table was generated according to the expert experience.Fuzzy quantization parameter of the combined controller was 0.25 and PID quantization parameter was 0.75.The simulation results of integrated controller showed that the maximum tracking error is 5.26 cm,the average tracking error is 1.36 cm,and the straightness accuracy is 2.34,which meets the requirements of transplanter operation.Then,aiming at the problem that the turning radius of transplanter was larger than the spacing between rows of crops when it turned at the edge of field,a sectional turning mode was designed.Finally,field tests were carried out on asphalt pavement and in paddy field environment.The results showed that the absolute average deviation was 2.85 cm and the straightness accuracy was 3.51 cm on asphalt pavement,which was 6.11 cm and 9.32 cm in paddy field.(3)This research obtained the information of experimental fields through remote sensing technology of UAV and satellite,designed algorithms for coordinate automatic matching and obstacle boundary automatic extraction based on image processing technology,and determined the lowest resolution of UAV remote sensing image for obstacle detection.Firstly,the boundary of obstacles was taken as the object of study,and the 31 cm panchromatic resolution and 1.24 m multi-spectral resolution images provided by WV3 satellite were compared with the UAV remote sensing images.The results showed that the average deviation of obstacles' boundary length between the high resolution satellite remote sensing results and the measured results was 17.3 cm,which was much larger than the 3.4 cm of UAV remote sensing.Secondly,the RGB image of farmland in the west area of campus was acquired by using Sony A7 RII camera on the eight-rotor UAV.The coordinate registration and obstacle boundary extraction process were completed manually by using ArcGIS.The average deviation between the geographic coordinates converted from the registration maps and the actual geographic coordinates converted from longitude and latitude of eight landmark was 4.6cm in the X direction and 5.7cm in the Y direction.Then the algorithm was designed to realize the automatic registration of coordinates and the automatic extraction of obstacle boundary.The average deviation between the converted geographic coordinates on the registration map and the actual geographic coordinates of eight landmarks was 4.6 cm in the X directionand 12.0 cm in the Y direction.Meanwhile,the manual and automatic extraction of seven geographic coordinates were carried out.The average deviation of obstacle corner coordinates was 2.9cm in the X direction and 5.4cm in the Y direction.The conclusion was drawn that the algorithm designed can be used in path planning in subsequent research of automatic extraction of farmland navigation information.Lastly,the obstacle boundary was extracted based on correlation coefficient template matching after compressing image,when the pixel reached 735x2174(the image resolution reaches 6cm),the average deviations of boundary points I of six obstacles on the X and Y were individually 0.87 and 0.95 cm and the whole process of detection only took 3s.(4)This research judged and recognized people with different postures in farmland through deep learning algorithm,and determined a model suitable for mobile embedded and cloud obstacle detection platform individually.For mobile platform and cloud platform obstacle detection system,three kinds of target detection models based on convolutional neural network were selected individually.After the above models were trained on cloud platform by Tensorflow,people in experimental fields were detected on mobile app.Firstly,the overall detection accuracy,average detection time and maximum detection time of each model were obtained.Secondly,by calculating the normalized index of each model's detection results and transforming it into final score through formula,the mobile platform models of Mobilenet-SSD,Mobilenet-PPN and cloud platform model of Mask RCNN+Inception were selected for further analysis.Then,the detection distance of the model was used as a criterion to analyze the performance of these three models mentioned above.The results showed that the detection accuracy of the model Mobilenet-SSD decreased more slowly with the increase of the distance from the camera compared with Model MobilenetPPN.And for the cloud platform,the detection accuracy of model Mask R-CNN+Inception kept 93.1% when the distance was up to 9m.Finally,it was concluded that the model Mobilenet-SSD and model Mask R-CNN+Inception were used as the deep learning model for the mobile embedded and cloud obstacle detection platform. |
PDF Full Text Request