|Yang, Fuzeng -|
|Liu, Shan -|
|Chen, Liping -|
|Song, Huaibo -|
|Wang, Yuanjie -|
Submitted to: Transactions of the Chinese Agricultural Machinery
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: April 20, 2012
Publication Date: May 10, 2012
Citation: Yang, F., Liu, S., Chen, L., Song, H., Wang, Y., Lan, Y. 2012. Detection method of various obstacles in farmland based on stereovision technology. Transactions of the Chinese Agricultural Machinery. 43(5):168-172. Interpretive Summary: Camera vision is often used to distinguish objects for unmanned and automatic work. Special recognition methods are needed for each object, but there are many kinds of obstacles in the field. A general method based on stereovision like a pair of human eyes is needed to recognize different objects quickly. A binocular stereovision system was built based on two cameras and a computer programmed with a detection method for farmland obstacles. The computer workflow of automatic obstacle recognition was constructed by the following five steps: camera calibration, image acquisition, stereo rectification, stereo match, and depth (distance) calculation. Five different obstacles such as human, human in the field, brick, shovel, and hole in the field in different environments and their precision position information were obtained in real-time during the experiment, especially for objects within a range of two meters. This general method to recognize different objects quickly could be used as vision guidance to help unmanned or automatic agricultural machines such as harvesters and planters to work properly and safely.
Technical Abstract: Camera calibration, image acquisition, stereo rectification, stereo match and depth calculation were developed to study detection method of various obstacles in farmland, but using the methods such as Bouguet algorithm, area match, triangulation, and so on. Five different obstacles and their position information in different environments were obtained, and open computer vision library Open CV was used to improve the real time property. The experiment showed that the accuracy rate of obstacles detection reached 96 percent and the absolute error of depth was maintained within +/- 30 mm (that is under 1.6 percent of the relative error) when the depth was less than 2,000 mm, and it takes less than 100 ms to finish an obstacle detection.