SLAM Localization Accuracy through Dynamic Object Detection and Adaptive Feature Filtering

To address the critical challenge of localization accuracy degradation in visual Simultaneous Localization and Mapping (SLAM) systems, primarily caused by dynamic feature point interference in complex environments, we propose an advanced visual SLAM framework that integrates deep learningbased object detection with an optimized feature-point filtering strategy. The proposed methodology follows a two-stage processing pipeline. First, a YOLOv5-based object detection module accurately identifies and segments dynamic objects in the operational environment. Second, a probabilistic dynamic feature-point elimination mechanism refines localization precision by selectively retaining reliable static features. To validate the framework, we conduct comprehensive experiments using datasets collected from an Automated Guided Vehicle (AGV) system operating in port environments. Comparative results demonstrate that our approach achieves a 50% improvement in localization accuracy over the conventional ORB-SLAM3 framework in dynamic scenarios. These findings not only confirm the effectiveness of our method but also provide valuable insights for developing robust SLAM systems in real-world engineering applications.