IDMF-VINS: Improving Visual-Inertial SLAM for Complex Dynamic Environments With Motion Consistency and Feature Filtering

Abstract

The detection of dynamic feature points presents a substantial challenge to dynamic scene analysis for simultaneous localization and mapping (SLAM). Conventional methods based on semantic segmentation, which are capable of producing complete object outlines, are expensive and not compatible with applications that run in real time. This study proposes a novel method combining YOLOv5 object detection information with motion consistency results to accurately differentiate between dynamic feature points and the corresponding states of predefined objects. To roughly distinguish background and dynamic objects within the object detection bounding boxes, a deep clustering approach is employed. The cluster centers have been optimized through iterative computation. In addition, a depth-based anomaly outlier filtering algorithm is employed to exclude stationary points in extremely close proximity to dynamic objects, thereby enhancing the capacity to distinguish between dynamic objects. The proposed method effectively minimizes the distortion resulting from dynamic feature points throughout pose estimation, which enhances the overall performance of the system while preserving a comparable quantity of feature points.