A) Scene understanding in autonomous driving

Autonomous driving has received widespread concern and development recently. Scene understanding is one of the key building blocks of autonomous driving in dynamic, real-world environments. Scene understanding tasks are mainly divided into object detection and semantic segmentation. Earlier works [Reference Maturana and Scherer19–Reference Wang and Jia23] have made great progress in the detection task of autonomous driving. Unfortunately, the bounding box representation can just acquire rough localization information, but lacks semantic details to distinguish different objects. Semantic segmentation provides annotation for each 3D point, which is essential for the visual perception task, because a self-driving vehicle needs to give different attention to different objects during driving, therefore semantic segmentation can provide not only the important information for vehicle decision-making, but also a powerful auxiliary role for precise positioning, which is essential in many autonomous driving applications. Previous approaches have achieved promising results in 2D semantic segmentation [Reference Teichmann, Weber, Zoellner, Cipolla and Urtasun24–Reference Yang, Zhao, Shi, Deng and Jia26]. They do make good use of the texture information in 2D images. However, it is a great challenge to apply them to the unstructured 3D points.

B) Semantic segmentation of point clouds

Traditional 3D semantic segmentation approaches [Reference Chen, Tian, Shen and Ouhyoung27,Reference Sun, Ovsjanikov and Guibas28] mainly focus on extracting hand-crafted features which contain complex post-processing operations as presented in [Reference Feng, Taguchi and Kamat29,Reference Himmelsbach, Mueller, Lüttel and Wünsche30]. Meanwhile, these approaches usually need a lot of parameters and are difficult to tune, and thus the performance is relatively low compared to the recent deep neural networks-based approaches. Deep learning has been widely explored in 2D image segmentation. But until recently, significant progress was achieved by learning comprehensive and differentiated characteristics of 3D point cloud. [Reference Li, Bu, Sun, Wu, Di and Chen4,Reference Qi, Su, Nießner, Dai, Yan and Guibas20,Reference Wu, Song, Khosla, Yu, Zhang, Tang and Xiao31] propose to represent the point cloud as a high-dimension volumetric form which can be applied with 3D convolutional neural networks. This representation is constrained by its resolution due to data sparsity and the redundant computation of 3D convolution operations which is not suitable for LiDAR point clouds. PointNet [Reference Qi, Su, Kaichun and Guibas7] introduces a new architecture for the unordered and irregular raw 3D point cloud. By learning high-dimensional features with several MLPs and extracting local and global features with max-pooling, it is the first attempt to use the un-ordered point clouds as the input and apply symmetrical operators that are able to deal with ordering problem of the 3D point cloud. Finally, max pooling is used to extract global and permutation-invariant features. However, it is difficult to extract local information, which limits its generalizability to more complex environment. This problem was tackled by designing a hierarchical structure to capture more local structures in PointNet++ [Reference Qi, Su, Kaichun and Guibas6]. By applying individual PointNet architectures in a local area, PointNet++ captures local information and then applies this hierarchical architecture to exploit increasingly contextual architectures in a high-dimensional space. Although high performance has been demonstrated in the indoor scenes, neither PointNet nor PointNet++ can be well generalized to the sparse outdoor point clouds captured from a 32-beam LiDAR sensor equipped on a self-driving vehicle. PointCNN [Reference Li, Bu, Sun, Wu, Di and Chen5] introduces a new transfer module to learn weighting information from the original data which can benefit from the irregular and unorder 3D data, then typical 3D architecture is applied to obtain final segmentation results. Inspired by the SIFT feature extractor [Reference Lowe32], PointSIFT [Reference Jiang, Wu and Lu2] employs local octant directional vectors as feature extraction layers which show good results in indoor scenes. OctNet [Reference Riegler, Ulusoy and Geiger8] uses octree architecture to express input point clouds. The memory cost of this method is too high for the autonomous driving application.

C) Semantic segmentation of large-scale LiDAR point clouds

The primary challenge of LiDAR perception on the whole may be the sparse and large-scale characteristic of 3D point cloud. SPG [Reference Landrieu and Simonovsky3] manages to summarize the local relationships in a similar fashion to PointNet by coarsely segmenting a point cloud into a superpoint graph (SPG). A SPG is designed for the large-scale 3D outdoor point clouds. Superpoints are locally coherent, geometrically homogeneous groups of points that get embedded by feature extractors. SPG outperforms the previous methods designed for dense point clouds. Due to the heavy workload for data annotating, a large-scale LiDAR dataset captured by a low-resolution LiDAR sensor for semantic segmentation was not available. Approaches were proposed by projecting the 3D LiDAR point clouds to a number of 2D images from different viewing directions [Reference Wang, Shi, Yun, Tai and Liu33,Reference Wu, Wan, Yue and Keutzer34] or using synthetic data [Reference Chen, Ma, Wan, Li and Xia35,Reference Varga, Costea, Florea, Giosan and Nedevschi36]. SqueezeSeg [Reference Wu, Wan, Yue and Keutzer34] uses a spherical projection of the point cloud enabling the usage of 2D convolutions and conditional random fields (CRF) for semantic segmentation. This representation allows to perform the segmentation by utilizing a light-weighted full convolutional network, and the last step will join the 2D segmentation results to form the 3D segmentation results. PointSeg [Reference Wang, Shi, Yun, Tai and Liu33] improves the CRF procedure of SqueezeSeg to provide more local details. SqueezeSegv2 [Reference Wu, Zhou, Zhao, Yue and Keutzer37] improves the architecture of SqueezeSeg with a new module named Context Aggregation Module and uses batch normalization and focal loss strategies to improve its robustness and effectiveness. However, the performance of these approaches is relatively low for the autonomous driving application, especially on small objects such as pedestrians and cyclists. Additionally, there is a gap between the actual sparse 3D point cloud and their 2D projections or the corresponding synthetic data Fig. 2.

Fig. 2. Multi-section ground plane segmentation. The ground plane was divided by the scanning beams. Each green dashed line represents a beam from the Velodyne HDL-32E Lidar used in this paper.

A) Ground point estimation

There is a visual discrimination in data distribution between the objects on the ground and the ground. The distribution of ground points is anisotropic, but points in the obstacle area are linearly distributed. This non-uniform, anisotropic distribution makes it difficult to directly adopt the state-of-the-art semantic segmentation approaches which were proposed for isotropically distributed point clouds. The 3D structure of a real-world driving scenario is extremely complex, a ground surface is not a simple 3D plane. As a result, using a single plane model to model the ground surface is not sufficient in the autonomous driving application. In our framework, a multi-segment plane fitting approach was proposed to model the ground points, and ground points are partially divided from a point cloud. The ground point estimates were then used as the pseudo labels for feature extraction. Our approach is motivated by the assumptions that most objects such as vehicles and pedestrians will be above the ground. We suppose that when the ground can be well extracted, then the segmentation accuracy of the other categories can be improved as well. Meanwhile, we empirically find that the ground plays a significant role in the segmentation of sparse point clouds, the objects consisted the number of points, mutual interference among the points leads to segmentation ambiguities.

We first separate the input point cloud into a number of regions along the driving direction (and let it be the x-axis) of a self-driving vehicle.

The variance of the angle interval $\triangle \theta$ between two beams of the LiDAR is normally very small. As a result, the density of point mainly changes on the horizontal distance of the traveling vehicle. As shown in Fig. 2, the input point cloud is mainly divided according to the angular interval of the rays. Take part of the point cloud in front of a moving vehicle for example, we will divide these 3D points into $N_{sec}$ parts by calculating a set of area boundaries $\{B_k\}_{k=0,\dots ,N_{sec}}$ as

(1)\begin{equation} B_k=h\tan(\theta_{min}+k\mu\triangle\theta),\; k=0,\ldots ,N_{sec} , \end{equation}

where $h$ is the vertical distance between the ground and the LiDAR installed on moving vehicle. $\theta _{min}$ and $\theta _{max}$ represent the range of minimum and maximum scanning angles of the LiDAR. $N_{sec}$ is the default number of sections, and $\mu$ denote the number of scanning rays in each block. The 3D points with $x$ coordinates residing in $(B_{k-1}, B_k\rbrack$ will be assigned to block $S_k$. For each divided block, we use RANSAC [Reference Fischler and Bolles38] to estimate a ground plane, and compute the parameters of a plane model using different observations from the data subsets. Points of both the ground and the objects will cause ambiguity for plane fitting. In order to reject outliers, we crop the point cloud by setting a limit to the ranges of the y-axis and z-axis values. When the ground points of the point cloud were extracted, they will be used to distinguish non-ground points from ground points by distance measurement. For a point at position $\mathbf {p}=(x,y,z)$ in a section $S_k$, if its Euclidean distance to the plane $d(p,P_k)$ is larger then a threshold $\sigma$, it will be temporarily considered as an object point. Otherwise, it is classified as a ground point.

B) Feature extraction

Due to long-range scanning characteristic of the point cloud, it is challenging to represent and extract useful information from an extremely sparse point cloud. Previous works usually attempt to utilize the 2D image representation projecting from the 3D points. However, the segmentation result is often a lack of accuracy due to the limited resolution of 2D images. As a result, it is especially not suitable for small objects.

Recent approaches such as [Reference Landrieu and Simonovsky3] have shown the effective progress in the application of graph-based representation of the point clouds in large-scale urban scenes. This method does not classify individual points, but divides them into simple geometric superpoint structures, therefore the scale of the large-scale data can be reduced significantly.

The step of partition is defined as an optimization problem in [Reference Landrieu and Simonovsky3], which could be seen as L0 variant [Reference Landrieu and Obozinski39] that can quickly find an approximate solution with a few graph-cut iterations. The solution of optimization problem is defined as super point. Then a graph is defined as $G = (S; E; F)$ whose nodes are the set of super points $S$ and edges $E$ represent the adjacency between super points.

The processing method is similar to superpixel methods for image segmentation which divided the point clouds as several super points. The step is purely unsupervised and makes use of the different dimensionality feature including linearity, planarity, scattering, verticality, and elevation which can maintain the original feature of the 3D point cloud in the whole scene. In the proposed architecture, we also apply a graph-based partition step of the input point cloud. The input large-scale point cloud will be separated into ground parts and object parts. We then apply graph-based step on the ground points and object points individually.

The PointNet [Reference Qi, Su, Kaichun and Guibas7] is used for point feature extraction due to its simplicity and efficiency. It is applied for each set of points that are clustered as a superpoint. In order to learn spatial information of different parts of point clouds, each superpoint graph is rescaled to the unit sphere before the extracting steps. Figure 3 illustrates the detailed architecture in the point feature extraction module. We employ a Spatial Transform Network (STN) which is combined with T-Net and the transform matrix to align the points in each superpoint to a standard space in the position and feature level.

We also apply Multi-Layer Perceptron (MLP). The values of MLPs have been presented in Fig. 3. The MLPs are used to map the input points independently and identically to a higher-dimensional space which can extract the feature of the point cloud.

MLP returns a feature vector with dimension 64 in local feature and 512 in global feature for each point in the superpoint. For each point in whole point cloud, the local and global feature are aggregated with a $576$-D feature vector.

In our architecture, we also build the connection between the ground points and the object points above the ground by utilizing the distance above the ground surface as an extra affinity for feature embedding. The point cloud is computed against the estimated ground model for each point, then we acquire the distances from the point to its projection on the plane model. As shown in the distance feature extraction module in Fig. 3, we use the position and distance as input which is represented as $(x,y,z,d)$, where $d$ represents the distance described above. It allows us to compute a height value with respect to the ground plane. Then the (${x, y, z}$) coordinates and distances are separately analyzed by the partial PointNet with STNs and MLPs. After the embedding module, we acquire the features embedded with distance affinity in the dimension of $N\times 512$. These features extract more information from the point cloud. Finally, different features are concatenated for semantic segmentation.

C) Weighted loss function

To better combine the ground affinity, the entire framework in this paper is trained by cross-entropy loss with the ground estimates computed automatically as described in Section A. It can be seen that there is a severe imbalance in the number of categories in the point cloud dataset, which is biased toward dominant classes rather than the others from the perspective of having higher number of instances. For instance, the number of instances for vehicles class is higher than the others. The imbalance condition may lead to lower performance for the class with few instances.

In order to solve the imbalanced dataset problem and give more attention to the small objects, we utilize a weighted cross-entropy loss, which can not only classify different objects correctly, but also enhance mean IoU measurement metric of point-wise classification. Let

(2)\begin{equation} L_{weighted}={-}\frac1N_t\sum_{i=1}^{N_t}\sum_{k=1}^{K}\alpha_k {y}_{ik}\log {p}_{ik}, \end{equation}

where ${p}_{ik}$ represents the probability that the $i$th point belongs to the $k$th category, and the input point cloud to network has a total of $N_t$ points. We give different weights to different categories according to their frequency in the dataset $\alpha _k=f_{med}/f_k$, where $f_k=\sum ^M_{i=1} N_{ik}$ represents the total number of points of $k$th category in the entire training set. $f_{med}$ is the median of $f_k$ in all $K$ categories.

A) Datasets

A LiDAR sensor is capable of continuously transmitting and receiving scanning lasers over a 360-degree range. LiDAR is now a key component of a self-driving vehicle as it provides an accurate 3D representation of the driving environment. The 64-beam LiDAR were used in the previous dataset. For instance, the KITTI dataset [Reference Geiger, Lenz and Urtasun15] is applied for 3D object detection in autonomous driving. However, 32-beam LiDAR sensors are more popular recently due to its cost and long-range sensing ability. Meanwhile, it has a smaller size and thus is easier for installation. Due to the lack of data and the sparseness of the point cloud, it is challenging to apply deep neural network to point clouds captured from 32-beam LiDAR, which is the target of this paper.

B) Quantitative evaluation

Tables 2 and 3 present the quantitative evaluation results on the 3D-DF dataset and our dataset, which demonstrate that the proposed method outperforms the state-of-the-art for semantic segmentation on sparse point clouds.

Table 2. Evaluation on the DF-3D dataset. We show the quantitative results for each category in DF-3D dataset, our method outperforms the state-of-the-art through out the categories especially the bigmot class. 6pt

Table 3. Evaluation on the Semantic-LiDAR dataset. The quantitative results for each category are shown in the table. Note that the proposed method performs better for both small objects and big objects, resulting in 5% segmentation accuracy improvement on average comparing to the state-of-the-art.

The detection or semantic segmentation of small objects is quite challenging in the computer vision community [Reference Wang25,Reference Kampffmeyer, Salberg and Jenssen44,Reference Lin, Dollár, Girshick, He, Hariharan and Belongie45]. Tables 2 and 3 show that the state-of-the-art performs poorly on the small objects (e.g. crowds, pedestrian, and tricycle, etc.) on sparse point clouds, while the proposed method has a higher performance around the small objects. For instance, the pedestrian category generally has only a few sparse points. It is extremely difficult to identify and perceive the sparse points. However, visual evaluation shows that the better performance can be obtained from the proposed architecture.

C) Qualitative evaluation

Figure 5 presents the visual evaluation on the segmentation results on the 3D-DF dataset, which contains fine-grained semantic categories (e.g. crowds and pedestrian, big mot and small mot) that can effectively illustrate the model performance. Note that the segmentation results obtained from the proposed method are visually more accurate than the state-of-the-art, especially from the close-ups.

Fig. 5. A picture of the LiDAR and the corresponding passenger car used to captured the proposed dataset.

As can be seen from the close-ups (in green boxes) of the first row in Fig. 4, the proposed method correctly separates all small and large objects above the ground. However, the other methods either miss the small object category or cannot handle large vehicles. Meanwhile, it can be seen that our method is robust to points far from the LiDAR due to the integration of the ground information.

The close-ups (in orange boxes) of the second row in Fig. 4 show that when there are many vehicles in a small region, the current state-of-the-art methods will fail. The proposed method can better separate this type of crowd scenes. The close-ups (in green boxes) of the third row in Fig. 4 show that our approach is robust to large objects(big mot) which is easier to be misclassified in some complex urban scenes. From the visualized results, we can conclude that the proposed approach outperforms the state-of-the-art both quantitatively and qualitatively.