[A-13] ITS

11:30 AM - 11:45 AM JST(2:30 AM - 2:45 AM UTC)

[A-13-12] Semantic-Guided Soft Reliability Weighting for Dynamic-Scene Visual Odometry

〇JIAXIANG QIU¹, Shunsuke Kamijo¹ (1. Univ. Tokyo)

Keywords:

Visual Odometry,Dynamic Scene,Reliability Weighting

This study proposes a framework called Semantic-Guided Soft Reliability Weighting to improve the robustness of monocular Visual Odometry (VO) in scenes containing dynamic objects. Conventional dynamic-scene SLAM methods have several limitations: geometric robust kernels and RANSAC may degrade in strongly dynamic environments, while semantic filtering methods such as DynaSLAM remove dynamic regions in a hard manner and may discard useful visual information. Moreover, objects belonging to dynamic categories are not always actually moving. In this study, a frozen DROID-SLAM backbone is combined with semantic priors from YOLOv8 and a scene-dependent Adapter. The Adapter outputs weighting parameters from scene context and adjusts edge-wise weights consumed by Dense Bundle Adjustment. This enables scene-adaptive soft reliability control without hard removal of dynamic regions. The method is evaluated on the KITTI Odometry dataset using Absolute Trajectory Error (ATE), together with behavioral analysis of the weighting strategy.