Lightweight object-level monocular data association system

Mobile robots can generate point cloud maps based on continuous image sequences to realize the perception of the environment. However, in the lightweight application scenario, the sparse point cloud map generated based on monocular camera is difficult to carry out accurate object data association due to the lack of sufficient object information. Therefore, this paper proposes a lightweight object-level data association method based on the monocular camera to realize the robot’s more accurate perception and positioning of environmental objects. In this method, feature points of monocular images are tracked, semantic information is introduced to parallel data association, and map points are distinguished in three stages, and a mutually exclusive object table is proposed to improve the accuracy of data association. The method is extensively tested on several public datasets and a real robot. The results show that data association is superior to similar methods in terms of accuracy and computation cost. In addition, this paper implements a lightweight robot system, which can achieve a frame rate of 20 FPS when using only half of the system’s computing resources, achieving lightweight real-time performance.