Keywords

1 Introduction

In the last few years, extensive research and development have been conducted towards the widespread adoption of autonomous vehicles. With the anticipated proliferation of autonomous driving, there is an expectation that occupants, including drivers, will engage in activities other than driving, further increasing the demand for enhanced ride comfort. In these considerations, studies focused on improving ride comfort of autonomous vehicles have prominently explored active suspension control. Recently, attention has shifted towards preview control using on-board sensors that have evolved in conjunction with autonomous driving technologies.

For preview suspension control, various systems have been proposed to estimate road surface profiles. These include a system achieving high-precision road surface profile acquisition by mounting LiDAR at the front of the vehicle [1], and a system using optical sensors mounted on the roof of the vehicle to obtain road surface profiles [2]. While these preview sensing systems can independently detect the undulations of the road surface with high precision, their implementation often involves expensive sensors for preview control. Particularly, LiDAR stands out as one of the most costly sensors among those integrated into autonomous vehicles, and its use should be limited to the minimum necessary quantity and specific mounting positions for ensuring safe autonomous driving. For example, as point cloud dataset [3] shown in Fig. 1, frequently LiDAR installed for acquiring information about the surroundings of autonomous vehicles is a roof-mounted 32 channels. LiDAR used for environmental perception is often positioned on a high elevation such as on the roof, and the point cloud density is relatively low for use in preview control. Therefore, this study clarifies the differences in road surface profile estimation from the data obtained from roof-mounted LiDAR, as typically proposed for autonomous vehicles, contrasted with front-mounted, as suggested in conventional preview control studies. Moreover, to obtain a sufficient road surface profile for preview control from the sparse point cloud from LiDAR installed on autonomous vehicles, proposes that utilizes RGB image information obtained from onboard cameras.

Fig. 1.
figure 1

An example of point cloud data from roof [3]

Full size image

2 Experiments of Measuring Speed Bump

In this paper, to investigate the differences on road surface measurement by the mounting position and the number of channels of LiDAR, experiments were conducted using a system shown in Fig. 2a to measure speed bumps. Both roof-mounted and front-mounted 128 channels LiDARs (Ouster) were utilized, with an RGB camera installed near the LiDAR mounted on the roof. By using the system, a bump that has a cross-section shape as shown in Fig. 2b placed on a flat road surface was measured. In addition to the obtained 128 channels raw point cloud data, the point clouds were downsampled to 32 and 64 channels to create comparative data. For example, the point cloud data obtained from the front-mounted 128 channels LiDAR, which is suitable for preview control when both LiDARs capture the bump simultaneously, were compared with the point cloud data obtained from the roof-mounted LiDAR assumed to have 32 channels for use in autonomous vehicles, as shown in Fig. 3. Note that both plots have the distance from the front wheels as the X-axis. As shown in Fig. 3, there is a significant difference in point cloud density due to the different mounting positions and the number of channels. Additionally, it is considered highly disadvantageous from the perspective of capturing road surface shapes if the density decreases to the level shown in Fig. 3b, even when using LiDAR.

Fig. 2.
figure 2

Experimental Setups

Full size image
Fig. 3.
figure 3

Examples of Point Cloud from the System

Full size image

To conduct a specific comparison, data with the width of 1 m capturing the bump from the obtained point clouds were extracted, and 2D road surface profiles with the origin at beneath the front wheels were presented for each condition in Fig. 4. The red dashed lines in the figure represent the moving average of the raw point clouds, simplified estimating the road surface shape. It can be confirmed from Fig. 4 that front-mounted LiDARs capture the undulating shape better than roof-mounted ones, and larger number of channels LiDARs captures more detailed undulations. However, in the case of roof-mounted installation, except for 128 channels, the obtained point clouds are insufficient for capturing the undulating shape adequately, making it difficult to estimate the road surface profile. Besides, the Root Mean Square Error (RMSE) of the bump shape estimated by moving average compared to the true value and the ascending order based on the RMSE including the results for 64 channels are shown in Table 1. This study aims to enable the estimation of 2D road surface profiles with the same degree of precision as those obtained from high-density point clouds by front-mounted LiDAR, as typically handled in conventional studies on preview control, even from such sparse point clouds.

Fig. 4.
figure 4

Road Surface Profile from Raw LiDAR Point Cloud of Each Condition

Full size image
Table 1. RMSE of Bump Shape Estimation
Full size table

3 Interpolation Method by RGB Using FusionNet

In this study, to densify point clouds for road surface profile estimation, FusionNet that is a supervised machine learning framework proposed by Gansbeke et al. [4] is employed to perform depth completion from sparse depth images generated from low-density point clouds and their corresponding RGB images. This framework has an architecture as shown in Fig. 5. By combining part of the dataset used for training the original model with the data obtained from the current experiments, a new training dataset is created. This aims to construct a model more focused on depth of road surface than the original one.

Fig. 5.
figure 5

Architecture of FusionNet [4]

Full size image

4 Training and Results

In this training, depth map scaled to 15-m range from 32 and 64 channels front-mounted LiDAR, along with their corresponding RGB images, were provided as input data, with the ground truth data based on the 128 channels data as supervision. An example frame used in the training data is shown in Fig. 6. Our dataset provides 6650 frames for training, 2490 frames for evaluation, and 1091 frames for testing. In this paper we tried to achieve densification from 32 channels front-mounted LiDAR data. However, directly using the raw 128 channels data as the ground truth can make it challenging to estimate planes accurately due to noise. Therefore, denoising was performed as shown in Fig. 7. This involved segmenting the planes from the point clouds, removing the points recognized as planes, and inserting grids based on plane equations to denoise the data.

The results of point cloud densification by using our trained model at the same location as in Fig. 4 are shown in Fig. 8. The best RMSE between the dense depth map as an output and the ground truth was 0.023 m. Additionally, the RMSE of the undulating shape in estimated road surface profile from predicted dense depth map by the same method as those in Fig. 4 was 0.0118 m.

Fig. 6.
figure 6

One Frame of our Dataset

Full size image
Fig. 7.
figure 7

Ground Truth

Full size image
Fig. 8.
figure 8

Prediction Results with Trained Model

Full size image

5 Conclusion

In this paper, we compared measurement accuracy differences due to LiDAR installation position and the number of channels. The results indicate significant challenges in estimating undulating shapes with LiDAR mounted on autonomous vehicles. To address this, we a trained a model to densify low-density road point clouds. Future work will focus on enhancing accuracy by applying object detection techniques.