Marker-free coregistration of UAV and backpack LiDAR point clouds in forested areas

Abstract

Unmanned aerial vehicle Laser Scanning (ULS) and Backpack Laser Scanning (BLS) are two emerging mobile mapping technologies applicable for monitoring forested environments in unprecedented detail from complementary perspectives. Although ground-based backpack techniques provide detailed information about the forest understory and terrain, the measured point clouds based on SLAM techniques are stitched together gradually and normally expressed in a less-accurate arbitrary coordinate system. Conversely, ULS point clouds are acquired from above and usually georeferenced, yet the point density and penetrability near the ground may still suffer from dense overstory despite the low attitude operation. Coregistering the ground and aerial point clouds in the ULS coordinate system therefore provides a method for fusing understory and overstory information at single tree level without the time consuming procedure of applying ground control points. Since the ULS and BLS acquisition viewpoints differ greatly, standard coregistration methods requiring 3D point-level correspondences are likely to fail. This paper presents an object-level coregistration approach which instead operates on two sets of tree positions, with the goal of finding the optimal 3D transformation (consisting of rotation, translation and scaling) between the respective coordinate systems. The entire task is decomposed into separate problems of computing the common Z axis, estimating the scale, and 2D coregistration. In contrast to existing methods, our approach does not require additional information such as tree diameters or heights. We evaluated our method on real test plots involving diverse stem densities and tree species situated in forest farm of the eastern coastal region of Jiangsu, China. The tree positions for ground and aerial data were obtained respectively by cylinder fitting and tree segmentation. On 3 coniferous (dawn redwood) plots, 46–81% trees were matched with a distance below 50 cm, and mean position deviation of 27–36 cm. For 4 broadleaf (poplar) plots, no more than 50% trees were matched below a 1 m threshold and mean error of 54–67 cm, which can be attributed to the broadleaf trees’ more irregular shape and lack of a well defined tree top. Moreover, we show that the introduction of scaling into the transform can increase the matched tree count by up to 20 percentage points and decrease the mean matched distance by up to 13% compared to a strictly rigid transform.