Physically realizable laser radar 3D point cloud confrontation sample generation method and system

Abstract

The invention discloses a physically realizable laser radar 3D point cloud confrontation sample generation method and system, and belongs to the technical field of confrontation machine learning. The method comprises the following steps: acquiring original 3D point cloud data containing a target area, and modeling according to a physical law of laser radar point cloud; randomly injecting a preset number of confrontation points in the spherical coordinate range of the target area; and designing a'target hiding attack 'loss function and a'target creating attack' loss function, substituting the functions into the cloud confrontation sample simulation model, optimizing coordinate information of confrontation points which are randomly injected, and taking optimal output of the point cloud confrontation sample simulation model as a finally generated 3D point cloud confrontation sample. According to the method, the vulnerability of an existing 3D point cloud target detector is utilized, modeling is innovatively performed on the physical rule met by the point cloud data acquired by the laser radar, so that a physically realizable 3D point cloud confrontation sample facing two different types is constructed, and new guidance is provided for machine learning safety analysis and protection.

Description

technical field [0001] The invention belongs to the technical field of adversarial machine learning, and in particular relates to a method and system for generating a physically realizable laser radar 3D point cloud adversarial sample. Background technique [0002] Self-driving cars are developing rapidly in recent years, and some of them are already operating on public roads, providing driverless taxi services for passengers, such as Google's Waymo One and Baidu's Apollo Go. In autonomous driving perception, point cloud object detection is crucial to ensure the safety of autonomous driving. LiDAR is generally considered more reliable because it operates in harsher weather and lighting conditions than cameras. Most autopilot manufacturers now use lidar, which can provide a 360-degree viewing angle, and use infrared lasers to measure the distance to surrounding obstacles, store point cloud information in the form of coordinates and intensity, and then get the road environmen...

AI Technical Summary

Problems solved by technology

The literature [Xiang C, Qi C R, Li B. Generating 3d adversarial point clouds[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition.2019:9136-9144.] proposes a point cloud against sample generation method, by adding disturbances imperceptible to the human eye on the original point cloud to achieve the purpose of deceiving the PointNet network (3D point cloud target detector), but the point cloud adversarial sample generation method proposed by this method can only deceive 3D point clouds in the digital domain The target detector does not consider the principle of lidar to collect point cloud information, and cannot be practically applied in the real world

Images