FVNet: 3D Front-View Proposal Generation for Real-Time Object Detection from Point Clouds

Zhou, Jie; Tan, Xin; Shao, Zhiwen; Ma, Lizhuang

doi:10.1109/cisp-bmei48845.2019.8965844

Cited by 42 publications

(32 citation statements)

References 33 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…FVNet [39] and RangeRCNN [40] are two-stage methods. FVNet first generates proposals on FV feature maps using a CNN and uses a parameter estimation network extended from PointNet [26] to regress the final bounding box.…”

Section: ) Methods Converting the Point Cloud Into Fvmentioning

confidence: 99%

Real-Time 3D Object Detection From Point Cloud Through Foreground Segmentation

et al. 2021

View full text Add to dashboard Cite

This paper aims to apply real-time light-weight high-precision 3D detection for autonomous driving. We propose LIDAR-based 3D object detection based on foreground segmentation using a fully sparse convolutional network (FS 2 3D). We design a sparse convolutional backbone network and a sparse convolutional detection head to efficiently use the computing and memory resources and accelerate the inference. Instead of using the anchor-based method, we convert the detection problem into a foreground segmentation problem on a bird's-eye view. The sparse convolutional detection head predicts the objectness and bounding box on each active point on the sparse feature map. We design a new oriented bounding box coding method and corresponding loss functions. We predict the endpoints of two mutually perpendicular lines that pass through the foreground active points and indirectly predict the objects' oriented bounding box from these four endpoints. We use the indirectly calculated object center, size, and orientation as inputs of loss functions in the training step. Experiments on the KITTI dataset show that the sparse backbone network we designed is 2.2 times faster and 18.4 times fewer FLOPs than the dense backbone network. The average improvement of the loss functions based on the bounding box code is 1.1% and 0.8% on the BEV and 3D detection, respectively, compared to no addition of these losses. Moreover, FS 2 3D outperforms the state-of-the-art LIDAR-based method in speed and precision for both cars and cyclists.INDEX TERMS 3D object Detection, LIDAR, real-time, sparse convolutional neural network.

show abstract

Section: ) Methods Converting the Point Cloud Into Fvmentioning

confidence: 99%

Real-Time 3D Object Detection From Point Cloud Through Foreground Segmentation

et al. 2021

View full text Add to dashboard Cite

show abstract

“…Object detection based on the 3D point cloud falls into two categories: traditional methods and deep learning methods. In recent years, various deep learning methods [11,12] have continuously refreshed the detection accuracy rankings for the KITTI dataset [13]. However, these methods have still not been able to eliminate dependence on the dataset [14].…”

Section: Of 30mentioning

confidence: 99%

Optimal Vehicle Pose Estimation Network Based on Time Series and Spatial Tightness with 3D LiDARs

et al. 2021

View full text Add to dashboard Cite

Vehicle pose estimation is essential in autonomous vehicle (AV) perception technology. However, due to the different density distributions of the point cloud, it is challenging to achieve sensitive direction extraction based on 3D LiDAR by using the existing pose estimation methods. In this paper, an optimal vehicle pose estimation network based on time series and spatial tightness (TS-OVPE) is proposed. This network uses five pose estimation algorithms proposed as candidate solutions to select each obstacle vehicle's optimal pose estimation result. Among these pose estimation algorithms, we first propose the Basic Line algorithm, which uses the road direction as the prior knowledge. Secondly, we propose improving principal component analysis based on point cloud distribution to conduct rotating principal component analysis (RPCA) and diagonal principal component analysis (DPCA) algorithms. Finally, we propose two global algorithms independent of the prior direction. We provided four evaluation indexes to transform each algorithm into a unified dimension. These evaluation indexes’ results were input into the ensemble learning network to obtain the optimal pose estimation results from the five proposed algorithms. The spatial dimension evaluation indexes reflected the tightness of the bounding box and the time dimension evaluation index reflected the coherence of the direction estimation. Since the network was indirectly trained through the evaluation index, it could be directly used on untrained LiDAR and showed a good pose estimation performance. Our approach was verified on the SemanticKITTI dataset and our urban environment dataset. Compared with the two mainstream algorithms, the polygon intersection over union (P-IoU) average increased by about 5.25% and 9.67%, the average heading error decreased by about 29.49% and 44.11%, and the average speed direction error decreased by about 3.85% and 46.70%. The experiment results showed that the ensemble learning network could effectively select the optimal pose estimation from the five abovementioned algorithms, making pose estimation more accurate.

show abstract

“…For point clouds, the widely used dataset is KITTI [102]. Considering that only a few models consume raw point clouds directly, we provide the related works, i.e., PointRCNN [103], VoxelNet [104], MVX-Net [105], FVNet [106], F-PointNet [107], and a deep Hough voting model [108].…”

Section: Applications Of Point Clouds Using Deep Learningmentioning

confidence: 99%

Deep Learning on Point Clouds and Its Application: A Survey

Liu

Sun

et al. 2019

Sensors

180

View full text Add to dashboard Cite

Point cloud is a widely used 3D data form, which can be produced by depth sensors, such as Light Detection and Ranging (LIDAR) and RGB-D cameras. Being unordered and irregular, many researchers focused on the feature engineering of the point cloud. Being able to learn complex hierarchical structures, deep learning has achieved great success with images from cameras. Recently, many researchers have adapted it into the applications of the point cloud. In this paper, the recent existing point cloud feature learning methods are classified as point-based and tree-based. The former directly takes the raw point cloud as the input for deep learning. The latter first employs a k-dimensional tree (Kd-tree) structure to represent the point cloud with a regular representation and then feeds these representations into deep learning models. Their advantages and disadvantages are analyzed. The applications related to point cloud feature learning, including 3D object classification, semantic segmentation, and 3D object detection, are introduced, and the datasets and evaluation metrics are also collected. Finally, the future research trend is predicted.

show abstract

FVNet: 3D Front-View Proposal Generation for Real-Time Object Detection from Point Clouds

Cited by 42 publications

References 33 publications

Real-Time 3D Object Detection From Point Cloud Through Foreground Segmentation

Real-Time 3D Object Detection From Point Cloud Through Foreground Segmentation

Optimal Vehicle Pose Estimation Network Based on Time Series and Spatial Tightness with 3D LiDARs

Deep Learning on Point Clouds and Its Application: A Survey

Contact Info

Product

Resources

About