Multi-Cue Dynamic Feature Scoring and Robust Filtering for RGB-D Visual Odometry

Abstract

Feature-based RGB-D visual odometry and simultaneous localization and mapping (SLAM) commonly rely on an approximately static scene; moving objects can introduce observations inconsistent with camera motion and reduce pose-estimation reliability. We present DynamicScore-VO, a multi-cue dynamic-risk-aware front end that retains ORB extraction while assessing individual keypoints before ORB-SLAM2 Tracking. A dense semantic prior, epipolar inconsistency, background-relative motion residual, and temporally accumulated evidence are fused into a continuous dynamic-risk score rather than a calibrated probability. A frame-adaptive threshold based on the median and median absolute deviation (MAD) removes high-risk keypoints without redesigning the downstream pose-estimation pipeline. On five Freiburg 3 dynamic sequences from the TUM RGB-D benchmark, the five-run median absolute trajectory error (ATE) root-mean-square error (RMSE) is lower on four sequences. Relative to Original ORB-SLAM2, median ATE RMSE decreases by approximately 98.44 percent and 94.93 percent on the two highly dynamic walking sequences. The results show pronounced gains under strong dynamics but scene-dependent behavior under weaker dynamics, including a mild degradation on sitting_xyz, indicating that dynamic-feature filtering is not uniformly beneficial across motion regimes.