基于RealSense深度相机的多特征树干快速识别方法

针对农业机器人在果园定位和导航中,环境背景复杂、光照强度变化大等问题,本文提出了一种基于RGB-D相机并利用颜色、深度、宽度和平行边特征的树干快速识别方法。首先,使用RealSense深度相机获取果园的彩色图像和深度数据;然后,将彩色图像转换为HSV颜色空间,再对HSV颜色空间中的S分量进行超像素分割,并将颜色特征和深度特征相近的相邻超像素块进行合并;随后,对深度图像进行树干宽度特征检测,对宽度置信率大于阈值的物体看作是待处理物体;最后,对待处理的物体进行平行边特征检测,在待处理物体边缘区域选择感兴趣区域窗口（ROI）进行边缘检测,搜索可能的树干边缘直边,当物体边缘的置信率R_B大于设定的阈值T_LB时,则识别为树干。通过对树干的多特征提取,有效提高了在不同环境下树干识别准确率。利用移动机器人平台在果园环境进行试验测试,以检验在强光照、正常光照和弱光照条件下树干识别算法的性能。试验结果表明,本文的树干识别算法在强光照、正常光照和弱光照条件下,树干识别的准确率分别为92.38%、91.35%和89.86%,每帧图像平均耗时分别为0.54、0.66、...

Fast Recognition Method of Multi-feature Trunk Based on RealSense Depth Camera

For agricultural robots in orchard positioning and navigation, the environmental background is complex and the illumination intensity changes greatly. In order to solve the problems, a rapid identification method of tree trunks was proposed by using the features of color, depth, width and parallel edges based on RGB-D camera. Firstly, the color image and depth image of the orchard were obtained by using RealSense depth camera. Then, the color image was converted into HSV color space, and superpixel segmentation was performed on the S component in HSV, and then adjacent superpixel blocks with similar color characteristics and depth characteristics were combined. Secondly, trunk width feature detection was carried out on the depth image, and the object whose width confidence rate was greater than the threshold value was regarded as the object to be processed. Finally, the parallel edge detection of the processed object was conducted, the region of interest (ROI) window was selected in the edge area of the object to be processed for edge detection, and the possible straight edge of the trunk edge was searched, when the confidence rate of the object edge was greater than the set threshold, the processed object was recognized as trunk, otherwise it was non-trunk. Through the extraction of multiple features of tree trunks, the recognition rate of tree trunks under different environments was improved effectively. In order to evaluate the performance of tree trunk recognition algorithm under strong light, normal light and weak light, a mobile robot platform was used to test in orchard environment. The experimental results showed that the recognition rate of this algorithm was 92.38%, 91.35% and 89.86% and the average time of each image was 0.54s, 0.66s and 0.76s under strong light, normal light and weak light, respectively, which can stably realize the trunk recognition in orchard environment.