基于自适应核带宽度Mean Shift算法的单木识别研究

Study on Single Tree Recognition Based on Adaptive KernelBand Width Mean Shift Algorithm

为提高多树种森林中单木识别的精度,利用机载激光雷达点云数据作为研究对象,提出一种基于自适应核带宽度Mean Shift算法的单木识别方法。该方法先采用直方图分析法分离树冠点云和冠层下点云,再采用基于二维增量网格投影的区域生长法,估算单木冠幅有效半径,然后以单木冠幅有效半径作为自适应核带宽度,对树冠点云进行自适应Mean Shift聚类分析,得到树冠点簇,最后采用包络盒方法根据树冠点簇和树干点云的空间关系识别单木。试验结果表明,检测树与实际树的位置、树冠形态近似一致,单木召回率达到86.1%,准确率达到91.5%,高于2个对比试验的结果。研究证明设置的自适应核带宽度能够自动调整以反映局部树冠的实际大小,在多树种森林的单木识别中表现良好。

To improve the accuracy of single tree recognition in multi tree forests,a single tree recognition method based on adaptive kernel band width Mean Shift algorithm was proposed using airborne LiDAR point cloud data as the research object.Firstly,this method used histogram analysis to separate the crown point cloud and the point cloud below the canopy.Then,a region growth method based on two-dimensional incremental grid projection was used to estimate the effective radius of a single tree crown.Then,the effective radius of a single tree crown was used as the adaptive kernel band width to perform adaptive Mean Shift clustering analysis on the crown point cloud,obtaining the crown point cluster.Finally,the envelope box method was used to identify a single tree based on the spatial relationship between the crown point cluster and the trunk point cloud.The experimental results showed that the position and crown shape of the detected and actual trees were approximately consistent,with a single tree recall rate of 86.1%and an accuracy rate of 91.5%,which was higher than the results of two comparative experiments.It was proved that the adaptive kernel band width can be automatically adjusted to reflect the actual size of local tree crowns,and performed well in single tree recognition in multi tree forests.