Simultaneous Localization and Mapping (SLAM) is a fundamental capability for autonomous mobile robots operating in unknown environments. However, most visual SLAM systems assume static scenes and incur substantial computational overhead in front-end processing, limiting real-time performance in dynamic environments. This paper presents ELY-SLAM, an efficient SLAM framework that integrates lightweight YOLO object detection for semantic perception. Lucas-Kanade (LK) optical flow is employed for inter-frame tracking, while Progressive Sample Consensus (PROSAC) is adopted for robust outlier rejection. The primary contribution lies in a low-latency heterogeneous GPU-CPU cooperative architecture, where deep learning inference is executed on the GPU and geometric tracking and optimization are processed on the CPU in a parallel and asynchronous manner. This design reduces front-end latency and improves robustness in dynamic environments. Extensive evaluations on the TUM-D and Bonn-D datasets demonstrate that the proposed system achieves improved real-time performance, trajectory accuracy, and robustness, while producing cleaner static maps through effective removal of dynamic objects. These results underscore the potential of ELY-SLAM to enable reliable localization accuracy and real-time for robots in indoor scenarios, thereby contributing to the development of intelligent robotic systems for smart cities and inclusive urban environments. • An efficient ELY-SLAM algorithm is developed by integrating LSR-YOLO (a lightweight object detection model based on YOLOv8n), a new LK optical flow method, and the PROSAC algorithm to improve dynamic feature point handling. • Dual-thread architecture in ELY-SLAM (frame construction and tracking) overcomes ORB-SLAM3's real-time limitations. • An innovative keyframe selection and redundancy removal strategy enables high-quality 3D static mapping with reduced ghosting and data accumulation.
Zhao et al. (Wed,) studied this question.