Nav2CAN: Achieving Context Aware Navigation in ROS2 Using Nav2 and RGB-D sensing

摘要

This paper presents a real-time human interaction detection system using RGB-D cameras to enable context-aware navigation for mobile robots. The system employs a convolutional neural network (CNN) architecture optimized for efficient inference on embedded GPUs. Using a keypoint detection based human detector on RGB-D images, interactions are localized in the 3D scene using object detection of humans. The detected human interaction zones are integrated into the robot’s navigation costmaps to modify planned paths accounting for social spaces. The system is validated through simulated and real-world tests showing reliable interaction detection at over 10 Hz. The modular system, called Nav2CAN, can be added to mobile robots operating in ROS2 (Robot Operating System 2) and achieve easy integration and compatibility with other packages. By combining deep learning-based perception with semantic navigation costmaps, socially-aware robot navigation in human environments is achieved.