基于Rank变换的农田场景三维重建方法

农田场景的三维重建对于研究远程监测作物的生长形态、预测作物产量、识别田间杂草等都具有重要作用。为解决农田场景图像三维重建困难、立体匹配精度较差等问题,该文提出了一种基于Rank变换的农田场景三维建模方法。该方法运用加权平均法灰度化图像,以保留农田场景的完整特征;以灰度图像的Rank变换结果作为匹配基元,采用基于归一化绝对差和测度函数的区域匹配算法获取场景的稠密视差图;根据平行双目视觉成像原理计算场景的空间坐标,并生成三维点云图;依据所得场景的三维坐标,对场景中感兴趣区域实现三维重建。采用标准视差计算测试图像验证立体匹配算法精确性,平均误匹配率较传统的绝对差和函数算法降低约5.63%。运用不同环境下的棉田场景图像测试三维重建方法,试验结果表明,在6.8 m的景深范围内,作物及杂草的高度、宽度等几何参数计算值与实际测量值接近,各项指标的平均相对误差为3.81%,验证了三维重建方法的可靠性及准确性。

Three-dimensional reconstruction method of farmland scene based on Rank transformation

Abstract: Growth parameters measurement of plants is an important aspect of crop growth monitoring, crop yield forecast and weeds detection. Since artificial measurements are always inefficient and inaccurate, more advanced technique of automatic measurement is required. Three-dimensional (3D) reconstruction can locate the real spatial position of target inside the view based on stereo vision techniques, which plays an important role in growth parameters measurement of plants. As field plants have similar features, the farmland scene is very difficult to be reconstructed completely in 3D space. Stereo matching is the key aspect of 3D reconstruction of farmland scene, which is usually time-consuming and low-accuracy. In order to solve the difficulty of 3D reconstruction of farmland and enhance the accuracy of stereo matching for farmland image, a new method based on Rank transformation was presented in this paper. The presented 3D reconstruction method consisted of 2 modules which were stereo matching and 3D cloud point reconstruction. The stereo matching module comprised grayscale transformation and disparity calculation. To reflect complete features of farmland scene, the weighted average method was used to image gray processing from color space to greyscale. Since the grayscale image is very sensitive to image noise, the Rank transformation result of grayscale image is set to matching primitive, which can increase the robustness of matching primitive against shadows, uneven illumination and other image noises. To save time and calculate dense disparity map, a region matching algorithm based on normalized sum of absolute difference (NSAD) measurement function was adopted to obtain the optimum disparity. The 3D cloud point reconstruction was composed of 3D coordinate calculation and color rendering. As the binocular camera used was assembled with 2 parallel and uniform monocular cameras, 3D coordinates of farmland scene were computed based on the parallel ranging method. Intrinsic and extrinsic parameters of the binocular camera were obtained with Zhang's calibration method. Global 3D coordinate of cloud point was obtained after being transformed from the camera coordinate system, which could describe the practical position in the farmland. After obtaining 3D points cloud of farmland image, the 3D reconstruction of interested region of total scene was accomplished. To test the accuracy of presented stereo matching algorithm, standard images of Teddy, Aloe and Cones, which were downloaded from the Middlebury website, were used to calculate disparity maps. The simple sum of absolute difference (SAD) stereo matching algorithm based on grayscale image was used as a contrast. The window sizes of the Rank transformation, the measurement function and the SAD algorithm were assigned as 5×5 pixel, 11×11 pixel and 11×11 pixel, respectively. Results of stereo matching test validate that the presented algorithm is accurate enough, which decreases bad matching ratios by 5.63% compared to the SAD algorithm. Images of farmland scene of cotton in different situations were used to test the presented 3D reconstruction method. Due to the limited view of the binocular camera, top regions of obtained disparity maps contained some errors. To reduce the effect of disparity errors, the regions with the depth less than 6.8 m on farmland scene were set to the interested region. Results of 3D reconstruction test showed that geometrical parameters such as the height and width of crop and weed were close to practical measurements. Moreover, the average relative error of total tested items was 3.81%. Although, only cotton farmland image is tested, the presented method will be efficient for more varieties of crops. The presented 3D reconstruction method of farmland scene is accurate and robust in the situations of weeds and shadows, which is available to measure outside geometrical parameters of plants for further crop growth monitoring and weed detection research.