| In the process of moving, intelligent agricultural robots need to perceive the obstacles inthe changing environment, so how to realize the real-time accurate obstacle detectionbecomes a hot topic in the agriculture research field. Since binocular stereo technology is fast,informative and has a strong capability of environment perception, it has been widely used inthe field of agricultural robot navigation. Existing research generally requires priorinformation of both the background and obstacles in the environment. However, consideringsuch annoying factors as the variation of background and irregularity of obstacles, it isdifficult to extract obstacles from background effectively.In response to this issue, while focusing on binocular technology, this paper proposed adisparity-based method for obstacle detection in unknown environment and achieved thereal-time and valid obtainment of the3D information of obstacles in farmland. This researchis expected to provide decision-making basis for agricultural robot navigation. The chief workand major innovations are as follows:(1) Analyzed the overall design of binocular vision-based system for obstacle detection,and set up the hardware and software platform of the system.(2) Realized the calibration method based on MATLAB toolbox, accompanied byexperiments and accuracy analysis for different baseline distances and calibration distances,which contributed to the selection of structural parameters. Results of experiments indicatedthat calibration error reached the lowest point without being beyond [0.17;0.17] when thebaseline and calibration distance were237mm and1300mm respectively. Meanwhile, basedon previous calibration parameters, Bouguet algorithm was employed to rectify originalimage pair.(3) Developed an improved stereo matching algorithm. Based on traditional region stereomatching technology, Box Filtering and multi-resolution matching were adopted to speed upmatching. In addition, on the basis of OpenCV, the reliability and robustness of stereomatching were enhanced by reducing disparity searching range. In order to evaluate theproposed algorithm, experiments were conducted on both standard testing images and realscene images. Experimental results indicated that the improved stereo matching algorithm could achieve high-precision disparity images as well as high matching speed.(4) Proposed a disparity-based obstacle detection algorithm. Combining area anddistance constraints, initial disparity image was processed by region labeling and threshfiltering techniques, resulting in the acquisition of obstacle disparity information. After that,the distance and size information of obstacles were gained via binocular measurementprinciple. Eventually, the method was experimentally verified by capturing images of realcornfield. Experiments demonstrated that the absolute error percent of the calculated size anddistance were upper-bounded by3.87%and4.37%respectively within the detection scope of5m. |
