Fast, robust and adaptive strip adjustment

BayesStripAlign 1.3
Goals:

Perform strip adjustment faster and more accurately than existing software using an advanced, spatially adaptive algorithm. No need to filter the data: results are robust to vegetation, buildings and noise.

Input:

– LAS/LAZ 1.4 or TXT LiDAR point clouds (separate swaths),
– SBET/SOL/TRJ/POF/TXT trajectory and attitude data

Output:

– Corrected LAS/LAZ/TXT strips
– [TIFF/JPEG Z-difference maps, histograms] [calibration parameters]

Supported platforms:

Windows, MacOS – 64-bit – Request a demo

Documentation (Version 1.3) [pdf, 244 kB, 10/6/16]

Advantages:
  • Fast point cloud registration, compared to ICP and least-squares surface matching
  • No pre-filtering required: built-in robustness to vegetation, buildings, point of view, changes and noise
  • High relative AND absolute accuracy through spatially adaptive attitude and position corrections
  • Boresight, internal geometry and attitude drift calibration from regular flight lines helps reduce absolute errors
  • Automatic boresight calibration (angles, leverarm, scan angle scale) module, accepting arbitrary geometry overlapping swaths
  • Multi-channel capability to calibrate and correct forward/backward looking scanners, e.g. Riegl Q1560
  • Large-scale and complex projects: correct strips by groups, choose which ones are fixed, to handle various scenarios
  • Flexibility: no accurate trajectory required, even post-production data can be fixed, trajectory segment text files accepted
  • QA/QC: vertical error analysis exporting TIFF/JPEG z-difference maps, color maps, hill-shaded elevation, and text histograms
Vertical difference between overlapping strips.  Left: original data. Right: geometric corrections applied (drifts, biases and local fluctuations).  The spatially adaptive registration is robust to vegetation, mismatched surfaces and natural differences between points of view.

LiDAR swath alignment.
Vertical difference between overlapping strips. Left: original data. Right: geometric corrections applied (drifts, biases and local fluctuations).
The spatially adaptive point cloud registration is robust to vegetation, mismatched surfaces and natural differences between points of view.

Histograms of z-differences.  Top: original swaths. Middle: boresight correction with drifts only, solves only part of the problem. Bottom: linear biases and local corrections applied, to achieve maximum accuracy , residual 1.5 cm RMS.

Histograms of z-differences between overlapping parallel strips.
Top: original swaths, affected by insufficient boresight calibration and IMU errors. Middle: boresight correction with linear drifts (solves only part of the problem). Bottom: linear biases and local corrections applied, to achieve maximum accuracy: residual 1.5 cm RMS.

Your benefits

Accuracy improvement

Correcting systematic errors or biases by treating them as parameters and applying Bayesian inference is a powerful way to improve the data accuracy. Strip adjustment is a good example of significant improvement, where both relative and absolute uncertainties can be dramatically reduced, depending on the number of strips and the amount of overlap.

Optimal data fusion, or swath combination, helps produce highest possible accuracy elevation models given the collected data. This is possible once the swaths have been properly aligned, so that most biases are eliminated.

automatic point cloud registration

Automation

Manually tuning essential processing parameters is no longer necessary. Grid size, number of iterations, internal parameters, that’s the algorithm’s problem, not yours!

Strip adjustment requires accurate alignment of overlapping point clouds from neighboring swaths. Probabilistic approaches provide automatic ways of computing the geometric transformation between such datasets while being robust to occlusions and noise.

convergence, automatic parameter estimation, unsupervised algorithm

Uncertainty management

Managing uncertainty properly allows not only to propagate errors, but to preserve information by using error structure (e.g. covariances). Understanding how uncertainty affects 3D points helps perform the alignment in optimal conditions, as most accurate points are given the largest weight. Uncertainty estimates do not come with the data, so we have to compute them when needed.
kernel regression, nonparametric model


Automatic boresight calibration. Vertical differences between two overlapping cross-strips, before (left) and after (right) calibration using n strips simultaneously. A standard calibration cross, or any set of overlapping lines can be used, not necessarily parallel or perpendicular. The calibration module estimates one set of parameters (boresight angles, leverarms and scan angle scale) for the whole dataset.  While boresight can account for a significant part of geometric errors, it is not sufficient to compensate time-dependent IMU errors, this is why we developed drift calibration and local corrections.

Automatic boresight calibration.
Vertical differences between two overlapping cross-strips, before (left) and after (right) calibration using n strips simultaneously. A standard calibration cross, or any set of overlapping lines can be used, not necessarily parallel or perpendicular. The calibration module estimates one set of parameters (boresight angles, leverarms and scan angle scale) for the whole dataset.
While boresight can account for a significant part of geometric errors, it is not sufficient to compensate time-dependent IMU errors, this is why we developed drift calibration and local corrections.

Absolute geometric correction, optimal combination

Relative alignment is not enough! Combining inter-swath transformations correctly is paramount to a real, measurable correction.
Chaining relative transformations estimated from overlapping swaths
(forward / backward propagation) allows us to compute optimal absolute corrections. This way, absolute geometric uncertainties can be reduced by a significant factor, up to the square root of the number of swaths.