关键词：
双目视觉;相机标定;SGBM算法;BM算法;深度学习
摘 要：
随着人工智能时代的到来,计算机视觉领域被广泛应用到各个行业中。同样的,人工智能改变着传统林业的研究方法,林业信息工程技术日渐成熟。针对传统树高测量方法中存在的结果准确性不高、操作困难、专业知识转化为规则困难等问题,采用了一种基于双目立体视觉理论计算树高的方法,实现了树木高度的无接触测量。以双目相机作为采集设备,基于MATLAB、VS2015开发平台,采用张正友单平面棋盘格相机标定方法进行单目标定和双目标定,从而获取双目相机2个镜头的参数。通过SGBM算法和BM算法立体匹配后获得视差深度图像,进而获取树木关键点的三维坐标信息并以此来计算树木高度。将深度学习与双目视觉相结合可以实现树木同时在二维和三维空间的信息提取。在VS2015上的试验结果表明,该方法操作相对简单,并且能够较为准确地测量树木高度,SGBM算法树高测量结果的相对误差范围为0.76%～3.93%,BM算法相对误差范围为0.29%～3.41%。结果表明:采用双目视觉技术测量树木高度可以满足林业工程中对于树高测量的精度需要。
译 名：
Research on measurement method of single tree height using binocular vision
作 者：
ZHANG Zhenwei;ZHAO Peng;HAN Jincheng;College of Information and Computer Engineering, Northeast Forestry University;
关键词：
binocular vision;;camera calibration;;SGBM algorithm;;BM algorithm;;deep learning
摘 要：
With the advent of the era of artificial intelligence, computer vision is widely used in various industries. Similarly, the artificial intelligence has changed the traditional forestry research methods, and the forestry information engineering technology has become more and more mature. Aiming at solving problems in traditional tree height measurement methods, such as low accuracy of results, difficulty in operation and difficulty in converting professional knowledge into rules, we proposed a method to calculate tree height based on binocular stereo vision theory, which would realize the contactless measurement of the tree height. This method used binocular camera as an acquisition device, and MATLAB and VS2015 as the development platform. The single target setting and binocular calibration were carried out using the calibration method of the Zhang Zhengyou Single Plane Checkerboard camera to obtain the parameters of two lenses of the binocular camera. After stereo matching with the SGBM algorithm, the parallax depth image was attained, and then the 3 D coordinate information of key points of trees was obtained to calculate tree heights. This study introduced the model of camera and calibration plate used in the calibration process for reference. In the aspect of deep learning, this study added a method to detect and identify tree species according to tree shape, which was of great significance for future forestry engineering applications. The combination of deep learning and binocular vision could realize information extraction of trees in 2 D and 3 D space at the same time. In this study, the elimination of the interference in the case of tree occlusion and the results of different kinds of trees in the process of image correction were analyzed in detail. In the experiment, single and multiple tree measurements were carried out, and the detailed analysis was conducted on improving the accuracy of tree height measurement results. The experimental results on VS2015 showed that this method was relatively simple and could measure the tree height more accurately. In the future, the application of binocular vision technology on smart phones can reduce the cost of measurement and make it more convenient operation. Similarly, the binocular vision technology can be applied to drones to measure the height of trees that are difficult to be photographed because of complex terrain. In addition, the binocular vision technology had certain value and significance for tree height measurement of sloping trees. The relative error range of tree height measurement results of the SGBM algorithm was 0.76%-3.93%, and that of the BM algorithm was 0.29%-3.41%. The results showed that the binocular vision technology could meet the precision requirements of tree height measurements in forestry engineering.
