How Rebulk Uses AI, Computer Vision, and 3D Sensors to Measure the Physical World

Measuring bulk inventory accurately is an engineering challenge that touches computer vision, 3D reconstruction, sensor fusion, and machine learning. It's not enough to point a camera at a pile and estimate a number. You need to understand geometry, account for terrain, handle occlusion, and produce measurements that operators can trust for real decisions.

In this post, we walk through the technical approach behind Rebulk — how we combine cameras, LiDAR, and depth sensors with AI-powered processing to turn raw sensor data into trusted inventory measurements.

Starting with the right sensors

Every measurement starts with data capture. Rebulk uses a combination of sensor types depending on the environment and use case:

The choice of sensor depends on the site. Outdoor stockpile yards with long range requirements typically use LiDAR and RGB cameras. Indoor sheds and bays may use depth cameras. Many deployments combine multiple sensor types to get the best coverage and accuracy for the environment.

3D reconstruction from sensor data

Raw sensor data — images, point clouds, depth maps — needs to be transformed into a coherent 3D representation of the environment. This is where reconstruction happens.

For camera-based capture, we use photogrammetric techniques to reconstruct dense 3D point clouds from overlapping imagery. Multiple viewpoints of the same scene are matched, aligned, and fused into a unified surface model.

For LiDAR-based capture, the sensor directly produces point cloud data. Our processing pipeline registers, filters, and segments this data to isolate individual piles and storage areas from the surrounding environment.

A critical step in both workflows is terrain modeling. Bulk materials sit on ground surfaces that are rarely flat. If you assume a flat base plane, you introduce systematic volume errors that grow with pile size. Rebulk models the actual ground surface beneath each pile, using surrounding terrain data to interpolate the base geometry. This terrain-aware approach significantly improves volume accuracy, especially on uneven industrial yards.

Computer vision for material understanding

Geometry alone doesn't tell you everything. Our computer vision models add a layer of material intelligence on top of the 3D reconstruction:

• Pile segmentation — automatically identifying where individual piles begin and end, even when they overlap or sit close together.
• Material classification — recognizing different material types based on visual and geometric features, so the system can apply the right density and packing assumptions.
• Packing factor estimation — analyzing the internal structure of piles (particularly for irregular materials like logs or biomass) to estimate how much of the bulk volume is solid material versus air gaps.
• Change detection — comparing measurements over time to track material additions, removals, and movement across the site.

These models are trained on data from real industrial deployments, not synthetic datasets. That matters because industrial environments have characteristics — dust, weather, variable lighting, equipment occlusion — that don't appear in controlled settings.

From volume to inventory

A 3D volume measurement is useful, but operators need inventory numbers they can act on. Converting volume to usable inventory requires understanding the material:

Density can vary based on moisture content, compaction, and material composition. Packing factor varies based on particle size, shape, and how the material was deposited. Our models account for these variables using both physics-based priors and learned adjustments from deployment data.

Edge processing and continuous monitoring

For sites with permanent hardware, data is processed on edge compute devices located at the site. This means measurements happen locally, with low latency, and without depending on continuous cloud connectivity.

Processed results are synced to the cloud when connectivity is available, feeding into the Rebulk dashboard where teams can monitor inventory levels, review trends, set alerts, and generate reports.

The system runs continuously. Instead of getting a measurement once a month from a survey, operators get updated inventory data throughout the day. That changes what's possible for operational planning, reconciliation, and logistics coordination.

Why industrial environments require a different approach

Building measurement systems for industrial environments is fundamentally different from building for controlled settings. A few of the challenges that shape our engineering:

• Dust and particulates can obscure sensors and reduce data quality. Our processing pipelines include filtering and quality assessment to handle degraded inputs.
• Weather exposure means hardware must survive rain, snow, extreme heat, and freezing temperatures year-round without maintenance windows.
• Operational interference — equipment, vehicles, and personnel move through sensor fields of view. The system needs to distinguish between inventory and activity.
• Connectivity constraints — many industrial sites have limited or unreliable internet. Edge processing and store-and-forward architectures are essential.
• Accuracy expectations — operators use these numbers for real financial and operational decisions. The system needs to be trustworthy, not just directionally correct.

These constraints aren't edge cases. They're the norm. And they're why we build everything with real-world deployment conditions as the baseline, not the exception.