基于变尺度格网索引与机器学习的行道树靶标点云识别

针对行道树连续喷雾施药方式严重污染环境,果园对靶施药技术难以推广至复杂城区环境等问题,应用车载2D LiDAR获取街道三维点云数据,研究行道树靶标识别方法。构建变尺度格网点云索引结构,实现邻域快速搜索及点云在线处理;提取高程、深度、密度、协方差矩阵等11个点云球域特征,分析特征分布特性,采用基于径向基核函数的支持向量机算法融合特征,学习树冠点云分类器;采用FIFO缓冲区保存点云帧序列,实现行道树靶标在线识别。实验结果表明,该方法能够实现行道树靶标精确识别,在测试集上的分类错误率小于0.8%,检出率大于99.4%,虚警率小于0.9%,鉴别力最强的4个特征从高到低依次是高程均值、深度均值、高程范围和高程方差。

Point Cloud Recognition of Street Tree Target Based on Variable-scale Grid Index and Machine Learning

Street tree continuous spray methods cause serious environmental pollution, however, the existing target spray technologies for tree are difficult to extend to complex urban environment. Aiming at the above problems, recognition method of street tree target was studied, which obtained the information of crown position and distance in real time and provided an accurate spraying basis for street tree toward-target spraying. The research results would improve the intelligent level of medical equipment for prevention and control of street tree and provide theoretical and technical support for street tree pest control, which had low injection, fine spraying, less pollution and high efficiency. Vehicle-borne 2D LiDAR was used to capture 3D point cloud data of street, and variable-scale grid index of point cloud was constructed to process point cloud data online and search neighborhood fast. Height, depth, density and covariance matrix features were extracted from spherical neighborhood of point cloud data, and an 11-dimensional feature vector was constructed. Distribution characteristics of features were analyzed and support vector machine algorithm based on the radial basis kernel function was used to fuse features and learn a point cloud classifier of crown. FIFO buffer was used to save point cloud frame sequences, and then street tree target can be recognized on-line. The classification error rate on the test set was less than 0.8%, with a detection rate more than 99.4% and a false alarm rate less than 0.9%. Four most discriminative features were selected, which were height mean, depth mean, height range and height variance.