Bilateral filter denoising of Lidar point cloud data in automatic driving scene

Abstract

3D Lidar has been widely used in auto drive systems owing to its high resolution, strong imaging ability, and long detection distance. The point cloud data (PCD) is generated by scanning the surrounding environment with 3D Lidar, providing a basis for decision-making and control of unmanned driving. However, the PCD is inevitably polluted by noise because traffic condition is complex and changeable, reducing the accuracy of environmental perception. The research on bilateral filtering of PCD in the automatic driving scene was carried out. The automatic driving experiment system is constructed. The campus scene is scanned by 16 Lidar lines. Four PCD traffic scene maps are obtained, consisting of bicycles, vehicles coming from the opposite side, many pedestrians, and vehicles parking on both sides of the road, respectively. Then, four point cloud maps are added with Gaussian noise. The effect of the parameters, i.e., the number of nearest neighboring points N, European distance weight Δd, and normal direction distance weight Δn on denoising, are researched. The experimental results show that the filtering effect is the best when N, Δd, and Δn are set to 10, 0.05, and 20, respectively. The signal-to-noise ratio is reduced by 5.6 dB on average. The values of N have little effect on the denoising results. However, the values of Δd and Δn have significant effects on denoising results. The results of this research can be used as an important reference for point cloud signal processing in other fields.