3D Point Cloud Multi-target Detection Method Based on PointNet++
3D object detection is an important research direction in the fields of computer vision and pattern recognition in recent years. This technology can provide important technical support for unmanned driving and intelligent robots. Aiming at the challenges
- PDF / 1,821,696 Bytes
- 12 Pages / 439.37 x 666.142 pts Page_size
- 11 Downloads / 257 Views
2
Liaoning Shihua University, Fushun 113001, China [email protected] ISDCT SB RAS, Ac. Lavrentieva ave. 17, Novosibirsk 630090, Russia
Abstract. 3D object detection is an important research direction in the fields of computer vision and pattern recognition in recent years. This technology can provide important technical support for unmanned driving and intelligent robots. Aiming at the challenges of object detection caused by the sparseness of 3D point clouds in outdoor scenes, this paper designs a 3D point cloud multi-target detection method based on pointnet++. The method first preprocesses the collected original point cloud; after obtaining the point cloud of the region of interest, the point cloud is clustered, and then the 3D target detection is performed by pointnet++ to obtain the object category. Finally, get the size and orientation of the target object through the 3D boundingbox. In order to verify the effectiveness of the method in this paper, the point cloud data of real outdoor scenes were collected using lidar, and a sample set was produced for network training. The final results verify that the method can achieve higher detection accuracy and meet the requirements of real-time performance. Keywords: 3D object detection
PointNet++ Point clound Deep learning
1 Introduction Object detection is one of the traditional tasks in the field of computer vision. Unlike image recognition, object detection not only recognizes objects on the image, but also gives the corresponding category of the object. It also needs to give the object through the minimum bounding box Location information [1]. According to the different output results required for target detection, the method of using RGB images for target detection to output the type of the object and the smallest bounding box on the image is generally referred to as 2D target detection. The detection using RGB images, RGB-D depth images, and laser point clouds to output object types, and the object’s length, width, height, and rotation angle in three-dimensional space is called 3D target detection. With the development of deep learning in recent years, the two-dimensional detection technology has become very mature, such as FasterRCNN [2, 3] and MaskRCNN [4] of the RPN series, and YOLOv1-YOLOv3 [5] of the OneShot series. In recent years, 3D object detection for outdoor scenes has been widely used, such as driverless and home intelligent robots. However, ordinary 2D detection in application scenarios such as unmanned driving, intelligent robots, and augmented reality © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2021 Q. Liu et al. (Eds.): CENet 2020, AISC 1274, pp. 1279–1290, 2021. https://doi.org/10.1007/978-981-15-8462-6_147
1280
J. Li et al.
cannot provide all the information needed to perceive the environment. It can only provide the position information of the target in the two-dimensional picture. Currently, the mainstream sensors used in outdoor scenes are camera and lidar. The
Data Loading...
