| Field robots for agriculture are one way to improve the automation of agriculture, which can be satisfied with the demand of future agriculture with high yield, food safety, high working efficiency, environment protection and precision horticulture. The automated guidance technology is one that must be dissolved as the basic technology. The paper has focused on the algorithms for vision based automated guidance system, which included algorithms of field road recognition, stereovision-based field environment reconstruction and guidance method, and positioning system. The main contents are:1) Set up the platform for automated guidance based on the idea of biorobot. There were several sensors to monitor the state of robot, which include a NIR stereo-camera, a tilt sensor, a turning angle sensor, two encoders, an offset measuring sensor and a heading measuring sensor. The system of Supervisory Control and Data Acquisition (SCADA) consisted of a PC, which was the center of the SCADA system, some ECUs, which were the node of the SCADA system, and the RS232 communication net. The software system was designed based on the thought of Object-Oriented program, which consisted of image class, ECU class, positioning class, recognition class, serial communication class, and etc.2) Developed some new algorithms for recognizing field road base on machine vision. As the general methods need a grey level threshold to segment an image, a correlation analysis-based approach was developed to segment field images based on the correlation coefficient threshold by constructing small windows. It can segment the rice crop from background of high dense weeds. 219 RGB rice images and 30 maize images were processed to validate it, and the results showed that the correct segmented rate was higher than 90%. A serial of models for field road location were also developed, which included a step model, a trapezium model, and a multi-trapezium model. Assisted by the wavelet multi-analysis decomposing algorithm and Mahalanobis distance (MD), a coarse-to-fine edge detection algorithm was developed. 219 RGB images and 200 NIR images were processed to validate the models. The results showed that the correct segmented rate was higher than 87% for RGB rice images, and higher than 69% for NIR images. It also showed that the model based algorithm was not sensitive to illumination, which is very important for field road recognition.3) Studied the basic methods to reconstruct the field environment based on stereovision. After comparing different stereo area matching methods, it was found that SAD, SSD, ZSSD, and MD could express the correct disparities;furthermore, the MD method could produce the best disparity images. It showed that the area size should be set based on the shape of crops. The methods of improving the correct rate of stereo matching were also studied. The 2D wavelet decomposition algorithm was used before the stereo matching process to avoid useless matching process, and comparing of minimum extremum and the second minimum extremum was used to remove the wrong matching. It succeeded to reconstructed field environment. The experiment had succeeded in reconstructing the landform of vegetable and paddy fields. The elevation was the main feature to detect crop rows, end of crop rows. We provided a stereovision-based guidance algorithm that integrated to detect crop rows, edges, and the end of crop rows.4) Developed the methods of positioning system for agriculture vehicle with high precision. The general vision-based positioning system need be rectified the camera accurately, however, a new system based on a 2D CMAC neural net was developed without camera rectified. The experiments showed thatoffset RMS was 10.5 mm, and offset STD was 11.3 mm;the heading RMS was 1.1°, and offset STD was 0.99°. A new vehicle model, which integrated a vision sensor, a turning sensor, and encoders, was developed. The positioning system was developed based on the new vehicle model and the extend Kalman filter. Results showed that the offset RMS was 8.7 mm, and offset STD was 18.3 mm.
|
PDF Full Text Request