改进随机样本一致性算法的弯曲果园道路检测

Detection of winding orchard path through improving random sample consensus algorithm

果园道路检测的目的是为农业采摘机器人鲁棒实时地规划出合适的行走路径,因果园环境的复杂性,例如光照变化、杂草和落叶遮挡等因素的影响造成视觉检测算法鲁棒性差,为此提出融合边缘提取和改进随机样本一致性的弯曲果园道路检测方法。首先,根据果园道路的颜色分布特征和几何形状特征,使用有限差分算子提取图像边缘,再使用灰度值对比度约束和霍夫直线检测去除噪声,实现道路边缘点提取。然后,提出多项式函数描述直线和弯曲道路,使用改进的随机样本一致性算法和线性最小二乘法拟合道路边缘点,以估计多项式函数的参数,实现果园道路检测。在华南农业大学果园采集240张道路图像作为试验对象。试验表明:在光照变化、阴影和遮挡背景的影响下,该方法能有效地提取果园道路边缘点,并能正确地拟合道路以实现道路检测,平均正确检测率为89.1%,平均检测时间为0.2639 s,能够满足视觉导航系统的要求。该研究为农业采摘机器人的视觉导航的鲁棒性和实时性提供指导。

Agricultural mobile robot and sightseeing agriculture is a direction of agricultural development in recent years. Agricultural mobile robot, a kind of efficient transportation equipment and means of transport, was of great significance in the orchard sightseeing agriculture. Road detection is the key technology and an important prerequisite for mobile agricultural robot to achieve autonomous navigation. In practical applications, the complexity of the orchard environment, e.g., the impact of illumination changes, shadows and occlusion, has resulted in poor robustness of vision detection algorithm. Therefore, the orchard road detection algorithm is required to be improved. So a method fusing edge detection and improved random sample consensus for winding orchard path detection was proposed. The proposed algorithm was consisted of orchard road edge detection algorithm（REE） and improved RANSAC algorithm（IRANSAC）. Because the orchard road image contained a lot of noise, such as shadows and occlusion, the REE was aimed at extracting road edge as well as removing noise according to the color distribution and geometry characteristics of the orchard road. First, using the finite difference operator to extract image edge may contain noises. Then a basic assumption that road edges had striking gray contrast among their neighborhood was proposed, so we used the constraint of contrast of gray values to removed noises. However, some noises satisfied the constraint condition, hence another assumption that a curved road could be seen as straight road in a certain scale was proposed, therefore, the image was divided into n regions, if n was large enough, a linear curve could approximate to curve in sub-image. On this basis, an improved hough line detection algorithm was executed to remove noises which were not lying on the lines. The REE could dramatically remove noises and keep the road edge points. However the REE could not remove all the noises, so the linear segments in the image could not represent curve. The spline curve model was proposed to describe the line or curve road, so the remaining problem was how to find the true spline among the edge points. IRANSAC was aimed at fitting the spline curve. The IRANSAC combining the advantages of linear least square method and RANSAC could correctly estimate model parameters of the spline curve, and achieve detecting orchard road. In order to test the proposed algorithm, we collected 240 Orchard Road images as test objects, including straight roads, curved roads, roads disturbed by illumination change and blocked roads in the South China Agricultural University. The result showed that： under the influence of illumination change and occlusion, the REE algorithm can effectively extract the edge of orchard image, and reduce 96.5% residual noises effectively, with the average computation time of 0.1658 s; The IRANSAC can correctly fit the road edge, and the correct fitting rate of the four roads are 93.3%, 86.7%, 85.0% and 91.7% higher than RANSAC respectively, with the average correct rate of 89.1% and the average detection time of 0.1834 s; Sometimes the IRANSAC failed to fit the right spline curve because the complex environment may cause road edge points missing, or the REE algorithm failed to get sufficient edge points. In brief, the proposed algorithm can satisfy the robustness of navigation system and real-time requirements, and ensure the effectiveness of the visual navigation system to achieve orchard road detection. In order to further improve the algorithm robustness under the clutter background, the key point is to improve robustness of REE.