基于双目视觉的农田障碍物检测

Detection of obstacles in farmland based on binocular vision

针对智能农业机械自动驾驶或辅助驾驶立体视觉障碍物检测技术中,传统阈值分割不能达到目标提取精度的问题,提出基于分析扫描线上区域分割与特征匹配相结合的障碍物检测算法。在双目视觉系统得到农田场景图像对中,通过分析扫描线上像素分布情况将图像分割,进行归类整理提取目标区域;对目标区域进行快速立体特征匹配,得到目标区域的空间信息,进行障碍物的检测。对800帧(400对)图像进行检测试验,结果表明:每对图像的平均处理时间<100 ms,本次试验检测出障碍物的正确率达到95%。该算法用于农田障碍物检测具有很好的检测效果。

To solve the problem of object extraction using methods of traditional threshold value segmentation, which exists in obstacle detection technology based on stereo vision of automatic drive or assistant-driving of intelligent agricultural machinery, a new arithmetic for obstacle detection based on the combination of region segmentation and characteristic matching on the scanning line is proposed in this paper. The farmland image is acquired, centered in the binocular vision system, segmentalized by analyzing the distribution of pixel on the scanning line and extracted the object region after arrangement. Then, after a fast stereo characteristic match on object region aiming at obtain its dimensional message, it is able to detect the obstacle. An experiment of detecting the image of 800 frames （400 pairs） had been done and the experimental result indicated that the average time of processing each pair was not beyond 100 ms and the accuracy of spotting the obstacle reached 95%. The result also showed that this arithmetic could get the location of the obstacle effectively and had a good performance in farmland obstacle detection.