CPU环境下多传感器数据融合的机器人3D目标检测方法

Multi-Sensor Data Fusion for Robotic 3D Target Detection in CPU Environment

实时、准确的3D目标检测算法能提供目标的位置和形态信息,为移动机器人实现高效导航、有效避障等各项任务提供保障。现有的3D目标检测算法对硬件设备运算能力的依赖较为严重,为了在确保检测精度的同时降低方法对硬件设备的要求,提出一种能部署在移动机器人CPU环境下的多传感器融合3D目标检测方法。方法结合了2D目标检测和点云聚类技术,利用2D目标检测技术从图像中获取目标的检测信息,根据相机与雷达的空间映射关系对检测框内的点云进行分割,并对分割后的点云进行聚类和信息提取,从而实现3D目标的检测和定位功能。通过与经典的多传感器3D目标检测算法MVX-Net的对比,该算法有更优的检测精度,同时具有更小的计算复杂度。此外,该方法在实际移动机器人CPU设备的边缘终端上进行部署分析,算法的处理速度达到0.069s/帧,满足10Hz激光雷达频率的需求。

Real-time and accurate 3D object detection algorithms can provide target position and shape information,en-suring efficient navigation and effective obstacle avoidance for mobile robots,among other tasks.Existing 3D object de-tection algorithms heavily rely on the computational capabilities of hardware devices.A multi-sensor fusion 3D object de-tection method that can be deployed in mobile robot CPU environments has been proposed to reduce hardware require-ments while ensuring detection accuracy.The method combines 2D object detection and point cloud clustering tech-niques.It utilizes 2D object detection technology to obtain object detection information from images,then performs point cloud segmentation within the detection bounding boxes based on the spatial mapping relationship between cameras and lidars.The segmented point clouds are further clustered and processed for information extraction,achieving 3D object de-tection and localization capabilities.By comparing with the classical multi-sensor 3D target detection algorithm MVX-Net,the algorithm in this paper has better detection accuracy with smaller computational complexity.Furthermore,the method is deployed and analyzed on actual mobile robot CPU devices at the edge terminal,achieving a processing speed of 0.069 seconds per frame,satisfying the 10 Hz laser radar frequency requirement.