Alibaba has made a new breakthrough in the field of autonomous driving. ?
On March 19, Alibaba Dharma Institute announced that a paper was selected for CVPR? 2020。 In this paper, a universal high-performance autopilot detector is proposed, which can give consideration to the detection accuracy and speed of three-dimensional objects and improve the safety performance of autopilot system. This is the first time in the industry. ?
The team in this paper said that the innovation of detectors is a key breakthrough in the field of autonomous driving. The detector proposed this time combines the advantages of single-stage detector and two-stage detector, so it realizes the intensive reading and speed improvement of 3D detection at the same time. The innovative research of future detectors can also solve more problems in the autonomous driving industry. Its team authors are all from Ali Dharma Institute, and the first author is Chen Hang, a research intern of Ali Dharma Institute. He, other authors include senior researcher of Dharma Institute and IEEE? Academician Hua Xiansheng, Senior Research Fellow of IEEE Dharma Institute? Researcher, Zhang Lei, etc. ?
It is understood that the autopilot detector is the core component of the autopilot, which needs to quickly process and analyze the multi-dimensional information collected by sensors and lidar, so that the vehicle can identify the surrounding objects and accurately locate the position of the objects in the three-dimensional space. This process needs the assistance of three-dimensional target detection.
Use RGB images for target detection, and output object categories and 2D images? Jump? The 2D detection method of the box is different, and the 3D target detection needs to be realized by using RGB images, RGB-D depth images and laser point clouds, and finally the information such as object category, length, width, height and rotation angle in 3D space can be output. ?
For autonomous driving, it needs to estimate more 3D bounding boxes from the real world to complete advanced tasks such as path planning and collision avoidance. In order to ensure the safety of automatic driving, the accuracy and speed of 3D detection are essential. However, as far as the two main architectures of 3D object detection based on point cloud are concerned, neither single-stage detector nor two-stage detector can take these two indicators into consideration.
In this regard, Dharma Institute proposed to use auxiliary network in training to solve the above problems. Specifically, it can transform the voxel feature in a single-stage detector into a point-level feature, and apply a certain supervision signal, so that the convolution feature also has the ability of structure perception, thus improving the detection accuracy. At the same time, the auxiliary network does not participate in the calculation (separation) when inferring the model, which ensures the detection efficiency of the single-stage detector.
In addition, dharma institute also proposed engineering improvement, which is partly sensitive? Warp? (PSWarp)，？ Which is used for solving the problem of' frame confidence mismatch' existing in a single-stage detector.
At present, the detector is the authoritative data set in the field of autonomous driving. BEV ranks first in the rankings. The test results show that KITTI, the authoritative data set in the field of autonomous driving? In the BEV ranking, this detector ranks first, its accuracy exceeds that of other single-stage detectors, and its detection speed reaches 25FPS? It is more than twice that of the current second-ranked scheme. ?
Compared with Baidu and Tencent, Alibaba has always been quite low-key in the field of autonomous driving. Since the automatic driving was announced in April 20 18, few voices have been heard. On the technical route, Alibaba chose L4-level autonomous driving road, trying to reduce the physical dilemma and cost obstacles of the existing scheme of autonomous driving through collaborative intelligence. The research was led by Wang Gang, the chief scientist of the artificial intelligence laboratory. At the Yun Qi conference last September, Zhang Jianfeng, CTO of Alibaba Group and president of Alibaba Cloud Intelligent, said that the automatic driving of Dharma Institute had reached L4 level, but no more information flowed out.
This external voice may mean that Alibaba will make more efforts in this field.
This article comes from car home, the author of the car manufacturer, and does not represent car home's position.