Adaptive Multi-Frame and Density Aware Fusion for 3D Reconstruction Using a Movable Stereo Sensor

Depth perception is essential for robotic and spatial intelligence, yet conventional stereo systems are constrained by fixed baselines and static calibration. This research develops a movable stereo camera system that mimics the human eye through mechanical actuation and selective online calibration, enabling a broader and more adaptive field of view for dense three-dimensional reconstruction. A spectral-entropy based motion detection mechanism monitors overlap between consecutive frames and triggers recalibration only when required, reducing redundant computation while maintaining geometric accuracy. The system increases spatial coverage by up to 39% compared to fixed-baseline stereo and achieves precision within 5% error in static regions and 7 percent in extendedmovable ranges, comparable to commercial Intel RealSense sensors. Building on this hardware foundation, the research introduces a multi-frame 3D detection framework that integrates spatial and temporal reasoning through two modules applied sequentially: the Adaptive Density-based Region Voting (ADRV) mechanism for spatial adaptation, followed by the Probability-Based Proposal Feature Fusion (PPFF) module for temporal integration. ADRV adaptively re-weights local features based on point-density distributions to address sparsity in point clouds, while PPFF fuses object proposals across frames using probabilistic gating and attention-based weighting. Together, these modules enhance intra-frame representation and inter-frame consistency, strengthening detection robustness under motion and occlusion. Experiments on a custom movable-stereo indoor dataset and on public benchmarks such as ScanNet show that integrating ADRV and PPFF improves geometric completeness and increases 3D detection accuracy by up to 13.4%, 7%, and 2.9% mean average precision for baseline detectors including VoteNet, GroupFree3D, and V-DETR, respectively. Qualitative results demonstrate improved reconstruction of thin structures and spatial coherence across sequential frames. In summary, this thesis establishes a cohesive pipeline linking adaptive hardware sensing with density-aware multi-frame fusion. The approach provides a foundation for self-calibrating and learning-enhanced 3D perception, advancing the scalability of stereo vision for robotic mapping, inspection, and autonomous navigation in dynamic environments.