Kaito NISHIZAWA

Investigation of Snow Bank Height Measurement on Road Shoulders Using Monocular Depth Estimation AI

Kazuyoshi TAKAHASHI

 In mountainous regions with heavy snowfall, the decline and aging of snowplow operators have made efficient snow removal a critical issue. While LiDAR offers high-precision 3D measurement, its high cost and lack of real-time capability remain challenges. This study investigates the effectiveness of using "Depth Pro," a state-of-the-art monocular depth estimation AI, to measure snow bank heights from drive recorder footage. The goal is to evaluate the stability of the "scale factor" required to convert AI-generated point clouds into real-world scales by comparing them with MMS  data.
 Static images were extracted from drive recorder videos captured in Nagaoka City between January and March 2025. After distortion correction, "Depth Pro" was applied to generate metric depth maps, which were then converted into 3D point clouds based on a pinhole camera model. Snow bank heights were measured relative to the road surface using CloudCompare. The "scale factor" (MMS measurement divided by AI measurement) was analyzed across three patterns: within a single image, along the depth axis, and across different shooting dates.
 In urban areas with distinct structures like buildings, the AI successfully reconstructed spatial geometry. However, accuracy deteriorated in low-visibility conditions or on featureless snow surfaces. The scale factor analysis revealed significant instability; while the coefficient of variation was as low as 7.1% under optimal conditions, it fluctuated between 21% and 37% across different dates. This suggests that while monocular AI excels at relative positioning, it struggles with absolute distance estimation on white, low-texture surfaces. Future work will focus on developing correction algorithms and methods to compensate for the lack of visual textures to achieve practical accuracy.