A Methodology for Ground Targets Detection in Complex Scene Based on Airborne LiDAR - Multimedia and Signal Processing - page 285

(a) (b) (c)

Fig. 1. A case of point cloud filtering (a) original point cloud. (b) non-ground points. (c) ground
points

Fig.1 (a) is a point cloud scene with building, trees, ground targets and ground. By
applying ETEW filtering algorithm, the point cloud is divided into two parts: ground
points and non-ground points. From Fig.1(c) objects such as building and trees and
ground vehicles are removed correctly. It can also be seen that a large number of ar-
eas of the ground point cloud have no points.

2.3 Ground Estimation

The task in ground surface estimation is to identify data points that correspond to bare
earth and then, based on this information interpolate the ground level at nearby posi-
tions. In this paper we use inpaint method proposed by Criminisi [7] to estimate the
ground surface.

Fig. 2. Ground estimation from Fig. 1(c)

Fig.2 shows the result of the estimated ground surface. Even in the presence of a
large area of the hole, the inpaint algorithm can also handle very well.

2.4 Objects Removal

Once the ground is estimated, for any point on the ground, given its X and Y coordi-
nates, you can check out the corresponding elevation values.Usually only given the
absolute elevation of a point there is no way to determine whether this point is a
ground target point, the trees point or the buildings point. However, if given the rela-
tive elevation to the ground of a point, it will be easier to determine which class that