ParkingTwin: Training-Free Streaming 3D Reconstruction for Parking-Lot Digital Twins

TL;DR

ParkingTwin introduces a training-free, streaming 3D reconstruction pipeline for parking-lot digital twins. It leverages OpenStreetMap semantic priors to initialize a metric $TSDF$, bypassing expensive optimization, and integrates geometry-prior based dynamic filtering with illumination-robust texture fusion in LAB space. The framework yields vehicle-free textures with SSIM $=0.87$, PSNR $=30.1$ dB, and LPIPS $=0.13$, at $30+$ FPS on a GTX $1660$, while drastically reducing VRAM compared to state-of-the-art baselines. It outputs explicit triangle meshes compatible with Unity/Unreal, enabling real-time AVP simulation and perception-validation workflows; the accompanying ICPARK dataset provides open-source data for reproducibility.

Abstract

High-fidelity parking-lot digital twins provide essential priors for path planning, collision checking, and perception validation in Automated Valet Parking (AVP). Yet robot-oriented reconstruction faces a trilemma: sparse forward-facing views cause weak parallax and ill-posed geometry; dynamic occlusions and extreme lighting hinder stable texture fusion; and neural rendering typically needs expensive offline optimization, violating edge-side streaming constraints. We propose ParkingTwin, a training-free, lightweight system for online streaming 3D reconstruction. First, OSM-prior-driven geometric construction uses OpenStreetMap semantic topology to directly generate a metric-consistent TSDF, replacing blind geometric search with deterministic mapping and avoiding costly optimization. Second, geometry-aware dynamic filtering employs a quad-modal constraint field (normal/height/depth consistency) to reject moving vehicles and transient occlusions in real time. Third, illumination-robust fusion in CIELAB decouples luminance and chromaticity via adaptive L-channel weighting and depth-gradient suppression, reducing seams under abrupt lighting changes. ParkingTwin runs at 30+ FPS on an entry-level GTX 1660. On a 68,000 m^2 real-world dataset, it achieves SSIM 0.87 (+16.0%), delivers about 15x end-to-end speedup, and reduces GPU memory by 83.3% compared with state-of-the-art 3D Gaussian Splatting (3DGS) that typically requires high-end GPUs (RTX 4090D). The system outputs explicit triangle meshes compatible with Unity/Unreal digital-twin pipelines. Project page: https://mihoutao-liu.github.io/ParkingTwin/

Figures (10)

• Figure 1: The complete pipeline of the ParkingTwin system. The system operates in three stages: (1) OSM-Prior Driven Geometric Initialization directly generates a metric-consistent TSDF mesh from semantic priors; (2) Geometry-Prior Based Dynamic Filtering utilizes multi-modal constraints to remove vehicles without training; (3) LAB Perceptual Fusion ensures seamless texturing under varying illumination conditions.
• Figure 2: Dynamic challenges in the ICPARK dataset. Captured frames exhibit high occlusion rates (>40% screen space) and diverse vehicle appearances. These factors render traditional semantic segmentation slow and unstable, necessitating our geometry-based removal strategy.
• Figure 3: Trajectory overview and global reconstruction comparison. (a) The robot executes a typical sparse forward-facing inspection path (blue lines). (b) ParkingTwin (Ours) generates a clean, vehicle-free floor plan. In contrast, (c) 3DGS and (d) ESLAM exhibit significant ghosting artifacts and geometric noise due to the lack of structural priors.
• Figure 4: Detailed comparison of texture quality on keyframes. Row 1 (Real Input): Raw captured images containing dynamic vehicle occlusions. Row 2 (Ours): ParkingTwin successfully removes dynamic vehicles (Frame 0) and reconstructs clear signage and lane markings (Frames 600/700). Row 3-4 (Baselines): ESLAM and 3DGS fail to remove vehicles (resulting in ghosting artifacts) and suffer from severe blurring or geometric holes in textureless regions (e.g., Frame 576).
• Figure 5: Failure of Traditional MVS (OpenMVS) under sparse views. Due to the lack of multi-view parallax, dense matching becomes mathematically ill-posed, resulting in >60% geometric loss and planar compression. This fundamental physical constraint validates the necessity of our OSM-driven prior.