Recently, the advancement of deep learning (DL) in discriminative feature learning from 3-D LiDAR data has led to rapid development in the field of autonomous driving. However, automated processing uneven, unstructured, noisy, and massive 3-D point clouds are a challenging and tedious task. In this article, we provide a systematic review of existing compelling DL architectures applied in LiDAR point clouds, detailing for specific tasks in autonomous driving, such as segmentation, detection, and classification. Although several published research articles focus on specific topics in computer vision for autonomous vehicles, to date, no general survey on DL applied in LiDAR point clouds for autonomous vehicles exists. Thus, the goal of this article is to narrow the gap in this topic. More than 140 key contributions in the recent five years are summarized in this survey, including the milestone 3-D deep architectures, the remarkable DL applications in 3-D semantic segmentation, object detection, and classification; specific data sets, evaluation metrics, and the state-of-the-art performance. Finally, we conclude the remaining challenges and future researches.

Autonomous vehicles were experiencing rapid development in the past few years. However, achieving full autonomy is not a trivial task, due to the nature of the complex and dynamic driving environment. Therefore, autonomous vehicles are equipped with a suite of different sensors to ensure robust, accurate environmental perception. In particular, the camera-LiDAR fusion is becoming an emerging research theme. However, so far there has been no critical review that focuses on deep-learning-based camera-LiDAR fusion methods. To bridge this gap and motivate future research, this article devotes to review recent deep-learning-based data fusion approaches that leverage both image and point cloud. This review gives a brief overview of deep learning on image and point cloud data processing. Followed by in-depth reviews of camera-LiDAR fusion methods in depth completion, object detection, semantic segmentation, tracking and online cross-sensor calibration, which are organized based on their respective fusion levels. Furthermore, we compare these methods on publicly available datasets. Finally, we identified gaps and over-looked challenges between current academic researches and real-world applications. Based on these observations, we provide our insights and point out promising research directions.

Neural networks have dominated the research of hyperspectral image classification, attributing to the feature learning capacity of convolution operations. However, the fixed geometric structure of convolution kernels hinders long-range interaction between features from distant locations. In this article, we propose a novel spectral–spatial transformer network (SSTN), which consists of spatial attention and spectral association modules, to overcome the constraints of convolution kernels. Also, we design a factorized architecture search (FAS) framework that involves two independent subprocedures to determine the layer-level operation choices and block-level orders of SSTN. Unlike conventional neural architecture search (NAS) that requires a bilevel optimization of both network parameters and architecture settings, the FAS focuses only on finding out optimal architecture settings to enable a stable and fast architecture search. Extensive experiments conducted on five popular HSI benchmarks demonstrate the versatility of SSTNs over other state-of-the-art (SOTA) methods and justify the FAS strategy. On the University of Houston dataset, SSTN obtains comparable overall accuracy to SOTA methods with a small fraction (1.2%) of multiply-and-accumulate operations compared to a strong baseline spectral–spatial residual network (SSRN). Most importantly, SSTNs outperform other SOTA networks using only 1.2% or fewer MACs of SSRNs on the Indian Pines, the Kennedy Space Center, the University of Pavia, and the Pavia Center datasets.

The mobile laser scanning (MLS) technique has attracted considerable attention for providing high-density, high-accuracy, unstructured, three-dimensional (3D) geo-referenced point-cloud coverage of the road environment. Recently, there has been an increasing number of applications of MLS in the detection and extraction of urban objects. This paper presents a systematic review of existing MLS related literature. This paper consists of three parts. Part 1 presents a brief overview of the state-of-the-art commercial MLS systems. Part 2 provides a detailed analysis of on-road and off-road information inventory methods, including the detection and extraction of on-road objects (e.g., road surface, road markings, driving lines, and road crack) and off-road objects (e.g., pole-like objects and power lines). Part 3 presents a refined integrated analysis of challenges and future trends. Our review shows that MLS technology is well proven in urban object detection and extraction, since the improvement of hardware and software accelerate the efficiency and accuracy of data collection and processing. When compared to other review papers focusing on MLS applications, we review the state-of-the-art road object detection and extraction methods using MLS data and discuss their performance and applicability. The main contribution of this review demonstrates that the MLS systems are suitable for supporting road asset inventory, ITS-related applications, high-definition maps, and other highly accurate localization services.

Recent geometric deep learning works define convolution operations in local regions and have enjoyed remarkable success on non-Euclidean data, including graph and point clouds. However, the high-level geometric correlations between the input and its neighboring coordinates or features are not fully exploited, resulting in suboptimal segmentation performance. In this article, we propose a novel graph convolution architecture, which we term as Taylor Gaussian mixture model (GMM) network (TGNet), to efficiently learn expressive and compositional local geometric features from point clouds. The TGNet is composed of basic geometric units, TGConv, that conduct local convolution on irregular point sets and are parametrized by a family of filters. Specifically, these filters are defined as the products of the local point features and the neighboring geometric features extracted from local coordinates. These geometric features are expressed by Gaussian weighted Taylor kernels. Then, a parametric pooling layer aggregates TGConv features to generate new feature vectors for each point. TGNet employs TGConv on multiscale neighborhoods to extract coarse-to-fine semantic deep features while improving its scale invariance. Additionally, a conditional random field (CRF) is adopted within the output layer to further improve the segmentation results. Using three point cloud data sets, qualitative and quantitative experimental results demonstrate that the proposed method achieves 62.2% average accuracy on ScanNet, 57.8% and 68.17% mean intersection over union (mIoU) on Stanford Large-Scale 3D Indoor Spaces (S3DIS) and Paris-Lille-3D data sets, respectively.

Although significant improvement has been achieved in fully autonomous driving and semantic high-definition map (HD) domains, most of the existing 3D point cloud segmentation methods cannot provide high representativeness and remarkable robustness. The principally increasing challenges remain in completely and efficiently extracting high-level 3D point cloud features, specifically in large-scale road environments. This paper provides an end-to-end feature extraction framework for 3D point cloud segmentation by using dynamic point-wise convolutional operations in multiple scales. Compared to existing point cloud segmentation methods that are commonly based on traditional convolutional neural networks (CNNs), our proposed method is less sensitive to data distribution and computational powers. This framework mainly includes four modules. Module I is first designed to construct a revised 3D point-wise convolutional operation. Then, a U-shaped downsampling-upsampling architecture is proposed to leverage both global and local features in multiple scales in Module II. Next, in Module III, high-level local edge features in 3D point neighborhoods are further extracted by using an adaptive graph convolutional neural network based on the K-Nearest Neighbor (KNN) algorithm. Finally, in Module IV, a conditional random field (CRF) algorithm is developed for postprocessing and segmentation result refinement. The proposed method was evaluated on three large-scale LiDAR point cloud datasets in both urban and indoor environments. The experimental results acquired by using different point cloud scenarios indicate our method can achieve state-of-the-art semantic segmentation performance in feature representativeness, segmentation accuracy, and technical robustness.