Geometric constraints and semantic optimization SLAM algorithm for dynamic scenarios

Traditional visual SLAM systems are predominantly designed for static environments, where they encounter challenges in dynamic scenes, leading to increased system errors and redundancy. This paper introduces a dynamic feature detection and filtering algorithm. Through a feature point selection and optimization strategy within quadtree nodes, high-response feature points are prioritized. Semantic information is leveraged to remove features on prior dynamic objects, and geometric constraints are applied to filter truly dynamic features. For unmatched features, an extension method is used, and high-confidence points are weighted to obtain feature point status information. Compared with the standard ORB-SLAM2 algorithm, our improved algorithm achieves over a 90% performance increase in highly dynamic environments, with absolute trajectory error performance improvements up to 96.84% in low-dynamic settings. Overall, our algorithm demonstrates superior adaptability and robustness in dynamic environments.