Abstract: In order to improve the sorting work efficiency of the modular manipulator, this paper studies a kind of real-time 3D semantic map generation method. This method introduces an improved region growing segmentation algorithm used for the extraction of object clusters. It obtains candidate models by feature matching method based on distance threshold, using Random Sample Consensus algorithm（RANSAC） and Iterative Closest Point（ICP） algorithm to generate transformation hypothesis and verify the candidate model. Finally it acquires position and orientation of objects, which are stored in the XML map file. It analyzes the resulting map file and the position and orientation of the tip of the mechanical arm, then the track of grasping task is achieved. The results show that the map generation method meets the operational requirements of real-time and accuracy.

Key words: 3D object recognition, region growing segmentation, semantic map building, Intrinsic Shape Signatures（ISS） algorithm, Color_Signature of Histograms of OrienTations（CSHOT） local feature descriptor

摘要： 为了提高模块化机械臂分拣作业效率，研究了一种实时三维语义地图生成方法。该方法设计了一种改进的区域增长分割算法提高了分割效率和准确性，通过基于距离阈值的特征匹配方法得到候选模型，采用随机采样一致性（RANSAC）和迭代最近点算法（ICP）生成转变假设并对候选模型进行验证，获取物体位姿信息，保存于XML地图文件中，将其与机械臂末端位姿进行推理分析，得到抓取任务轨迹。实验结果表明，该方法满足作业实时性与准确性的要求。

关键词: 三维物体识别, 区域增长, 语义地图构建, 内在形状特征（ISS）算法, 基于颜色与签名的方向直方图（CSHOT）局部特征描述子