基于三维重构的哈蜜瓜均瓣雕花算法

为解决哈密瓜雕刻速度慢、花瓣大小不一致等问题，该研究提出了一种基于三维重构的哈蜜瓜均瓣雕花算法。对多角度拍摄得到的哈蜜瓜照片进行滤波处理，提取其图像特征并进行稀疏重建，通过点云坐标得出哈蜜瓜的特征参数；接着在稀疏点的基础上利用CMVS/PMVS算法进行稠密重建；最后调节八叉树算法与泊松表面重建，得到哈密瓜精确三维空间坐标。根据哈密瓜体型特征及设定花瓣数量，将点云三角网格化在深度优先算法的基础上结合粒子群算法，规划雕刻起点、终点及雕刻深度，使每个花瓣体积相同。采用48个哈密瓜，雕刻花瓣数取15~30，雕刻深度为1.5、2.0、2.5 cm。其中切割花瓣数为28这一组精度最低，测得最大与最小花瓣体积分别为3.40与3.25 cm3，最大体积差为0.15 cm3，误差小于5%。结果表明，该研究提出的基于三维重构的哈蜜瓜均瓣雕花算法精度高，研究结果可为机器人雕刻哈密瓜提供技术支持。

Algorithm for the uniform petal carving of Hami melon based on three-dimensional reconstruction

Higher carving speed and uniform petal size of Hami melon are critical for the robot carving Hami melon. It is very necessary to plan the cutting path of the execution terminal (carving knife) in real-time, according to the three-dimensional coordinates of different processing objects. In this study, a uniform petal carving of Hami melon was proposed using point cloud splicing. The image features were extracted and reconstructed sparsely. The feature parameters of melon were firstly obtained by point cloud coordinates. Secondly, CMVS/PMVS algorithm was selected for dense reconstruction using the sparse points. Finally, the octree and Poisson surface reconstruction were used to obtain the accurate 3D spatial coordinates of melon. Different shapes of Hami melon was led to different reconstructions. Each piece of flesh presented the same volume after carving. The specific procedure was as follows. Firstly, the cutting height and depth of melon were determined to extract the point cloud. An arc function was then fitted to determine the center of the circle, according to the point cloud of the outermost circle of Hami melon. The number of carving petals was divided 360° to determine the pre carving start point, end point, and path. Specifically, the initial triangle was formed to search for the two closest points from any point in the numerous point clouds as the benchmark, and then to expand the triangle outward with the three sides of the triangle as the baseline, where the equal volume of each petal was taken as the objective function, while the equal cutting depth and cutting angle of each petal as the limiting conditions. Until all the point clouds were included in the three-dimensional triangle network, the area of the projected triangle was calculated by the Helen formula, where the average value for the Z coordinates of three projected points was taken as the height, and then to calculate the volume of the triangular pyramid. After depth-first and particle swarm optimization, the optimal solution was found in the coordinates of Hami melon point cloud through continuous recursive iteration. Finally, better cloud coordinates were stored as new datasets and then marked on the outside of Hami melon. As such, the manipulator was controlled to evenly carve the Hami melon. Specifically, the cutter first adjusted to the appropriate posture angle as posture point 2, then moved along the cutter ridge to a certain depth to posture point 1, and retreated to posture point 3, and finally, the cutter moved along the outer surface of Hami melon to the next adjacent posture point 2. These steps were repeated to complete the overall carving of the Hami melon. The regular and irregular models were also selected to verify the accuracy. The calculated volumes of cube, pyramid, and irregular body were compared with the real. Forty eight Hami melons (16 groups, 3 in each group) were divided, where the number of carved petals N was 15-30, and the carving depth H was 1.5, 2.0, and 2.5 cm. It was found that the precision of the group was the lowest with the number of cut petals N equal to 28. The maximum and minimum petal volumes were measured as 3.40 and 3.25 cm3, respectively, where the maximum volume difference was 0.15 cm3, and the error was less than 5.00%. Consequently, the melon petal carving using point cloud splicing presented a higher precision than before. The findings can provide strong technical support for robot carving Hami melon.