What is sensor fusion in reality capture?

Sensor fusion means combining measurements from different instruments into one coordinated dataset. In practice it usually means registering terrestrial LiDAR (accurate geometry) with photogrammetry (dense colour and texture), so the result carries both properties rather than trading one for the other.

How do you register LiDAR and photogrammetry into the same coordinate system?

We use survey control — ground control points or targets measured with GNSS or total station — plus cloud-to-cloud and ICP refinement between the photogrammetric reconstruction and the LiDAR cloud. Control points anchor both datasets to the same coordinate frame so accuracies can be verified numerically.

Do fused Gaussian splats stay accurate enough for engineering use?

The splat layer is a visual product; the fused LiDAR cloud remains the measurement reference. Because both are georeferenced to the same control network, dimensions are taken from the point cloud while the splat provides appearance. Accuracy claims are reported with an RMSE against control, which is the number that matters.

Sensor Fusion: Combining LiDAR and Photogrammetry into Gaussian Splats

Every capture method has a blind spot. LiDAR knows exactly where a surface is but renders it in muted colour; photogrammetry reproduces colour beautifully but drifts where texture is poor. Fusing them removes both weaknesses at once — and the fusion is what makes Gaussian splat deliverables defensible.

Anyone who has reviewed scan data for long enough has seen both failure modes. A point cloud that measures beautifully and looks like gravel. A photogrammetric model that looks like a photograph and drifts visibly across a flat facade.

Neither is a defect of the instrument. Each method simply optimises for something different. Fusing them is how you stop choosing.

What each sensor actually contributes

Terrestrial LiDAR (FARO Focus S 350, ±1 mm ranging) measures positions directly. It works in near-darkness, on textureless surfaces, and at range. Its weaknesses are appearance — colour is derived from integrated imagery or added later — and the fact that it captures geometry only along lines of sight from each scan position.

Photogrammetry reconstructs geometry and colour from overlapping photographs through structure-from-motion. Its colour is photographic because it is photography. Its weaknesses are texture dependence, sensitivity to lighting and movement, and the tendency of error to accumulate across large, flat, uniform surfaces.

Put them together and the fused product holds measured positions and photographic appearance.

The registration problem

Fusion is fundamentally a registration problem, and registration is where projects succeed or quietly fail.

Four anchors matter:

Survey control. Ground control points or targets observed with GNSS or total station. These define the coordinate system both sensors are pulled into.
Scan-to-scan registration. LiDAR positions are registered together, typically with target-based alignment refined by cloud-to-cloud matching, with the residual error reported.
Sensor-to-sensor alignment. The photogrammetric reconstruction and LiDAR cloud are aligned, checking both global fit and local behaviour on planes, facades and features.
Independent verification. Control points withheld from the fit are used to test, not train. This is the difference between a defensible result and an optimistic one.

The number we report is an RMSE against held-out control. A deliverable without that figure is asking to be trusted rather than proven.

Convergence into Gaussian splats

Once registered, the fused dataset feeds the splat training step — and this is where fusion becomes visibly worthwhile.

Camera poses come from the photogrammetric alignment, giving the optimisation views to train against.
Depth and geometry priors come from the LiDAR cloud, constraining the optimisation where photographs are ambiguous.
Colour fidelity comes from the imagery, which is what makes the result look real rather than sampled.

The outcome is a splat scene with substantially fewer floaters, cleaner surfaces on large flat areas, and correct global shape — the three places splats trained from imagery alone most often disappoint. For a deeper look at what splats are and are not, see 3D Gaussian splatting for visualization.

A decision framework

Fusion is not always the right answer. Our default guidance:

Site condition	Recommended approach
Complex industrial plant, tight tolerances	LiDAR-led, photogrammetry for colour; fusion for visual layer
Heritage facade or monument	Fused — LiDAR for geometry, UAV and ground photography for material and detail
Large open site, terrain-focused	UAV photogrammetry with LiDAR checkpoints where accuracy demands it
Rapid stakeholder visualisation only	Photogrammetry or splats alone, with accuracy expectations stated up front
Interiors with poor texture	LiDAR-led; photography adds little where surfaces are uniform

The pattern is consistent: whenever dimensional decisions follow the capture, LiDAR leads and photography serves appearance. When the deliverable is purely presentational, the cheaper single-sensor route can be entirely appropriate.

Quality control that holds up

Fusion projects fail for mundane reasons. A short list of the checks we run as standard:

Density and overlap review before leaving site, not after.
Double-scan comparison on selected geometry as an independent consistency check.
Control point residuals reported per dataset and per phase.
Plane flatness tests on representative surfaces to catch systematic drift.
Visual inspection of the splat against source imagery at matched viewpoints.

None of this is exotic. It is simply the discipline that distinguishes a survey from a demonstration.

Deliverables and formats

A fused capture typically ships as:

Registered LiDAR point cloud — LAS, E57 or RCP, classified.
Photogrammetric products — dense cloud where relevant, orthomosaics, DEMs.
Gaussian splat scene — .ply plus an optimised runtime format for the intended viewer.
Control report — coordinate system, control point table, residuals, and stated accuracy.

Clients then measure from the cloud, review from the splat, and model in BIM from the classified geometry.

Where this is heading

The convergence is real and moving quickly: sensors are cheaper, splat optimisation is more robust, and viewers are increasingly browser-native. What has not changed, and will not, is that fusion value flows from registration discipline. The instrument mix is the easy decision; the control network is the one that determines whether the data can be used.

To scope a fused capture for your site, see our reality capture and scanning service, our UAV capability or get in touch with the details of your project.