Efficient long-term visual simultaneous localization and mapping via sparse prior embedding and nonlinear score-guided sparsification under memory constraints

Abstract Efficient memory management is essential for the stability and long-term performance of mobile robots in Simultaneous Localization and Mapping (SLAM). However, existing methods often struggle to control redundancy in keyframes and map points, leading to reduced efficiency, increased latency, and potential system failure due to resource constraints. Achieving high accuracy in both mapping and trajectory estimation while maintaining a compact state representation remains a key challenge for scalable and efficient SLAM systems. To address this issue, this paper proposes an efficient long-term visual SLAM method based on sparse prior embedding and nonlinear score-guided sparsification for memory-constrained environments. The approach embeds keyframe information into sparse prior factors, avoiding global coupling while preserving system sparsity and consistency. Additionally, a nonlinear scoring function combining parallax and descriptor uniqueness is introduced to guide map point sparsification within the sliding window. This strategy enables efficient state graph management, achieving compact global map representations and effective observation constraints. The proposed method has been implemented in a complete visual SLAM system and evaluated through long-term real-world mapping experiments on an embedded robotic platform. Experimental results demonstrate that the approach significantly reduces memory consumption while maintaining trajectory and mapping accuracy. Furthermore, the method ensures real-time execution and deployment potential, indicating its suitability for large-scale SLAM tasks in resource-constrained and long-duration operational scenarios.