Using Machine Learning to Classify Unique Features in Point Cloud Data: Bridge Infrastructure
With advancements in technology, point clouds have emerged as a popular survey method for infrastructure management. They provide the ability to access difficult-to-reach areas remotely. Through analysis, visualization tools, and machine learning, point clouds can be analyzed to understand and estimate various conditions and characteristics of a survey site. Point cloud surveys often contain thousands, if not millions of points, capturing detailed positional information, attribute data, and more for a survey site.
Manually deriving information from these large data sets can be an arduous task. Applying a classification type to each point adds structural meaning to the data. Automatic classification methods for identifying standard features such as ground and vegetation are commonly available and incredibly useful for finding common features in any point cloud. When it comes to surveys of structures and many other niche objects, automatic classification tools may not be sufficient in such industry-specific scenarios. In such cases, machine learning techniques can be applied to customize automatic unique features that are not part of the standard feature set. Global Mapper Pro’s award-winning Custom Classification Training tool enables the creation of custom classifications that can automatically identify target features within a point cloud.
The terrestrial lidar scan of a bridge used in this example was collected by one of our users, Southeastern Surveying and Mapping Corporation. This high-resolution dataset cleanly captured all angles of the bridge, providing a clear representation of the features.
Automatic Classification Tools
For years, Global Mapper has provided an array of built-in automatic classification tools to identify ground, vegetation, building, pole, and wire points. Within the goal of creating additional custom classifications, these automatic tools can be used before training a classification to help filter points that aren’t part of our target features. Different methods are available for each tool to fit different point clouds and feature types better. The Max likelihood method, a machine learning method, was designed for high-resolution and terrestrial points clouds often used in infrastructure management. Max Likelihood is a segmentation-based method. For each classification type, the tool has been tailored to find clusters of points that have the common shapes and characteristics of these features in the point cloud.
In this point cloud, the vegetation classification tool was used to identify the grass and bushes. Classified points can be easily filtered out, exposing the infrastructure underneath.
*Tip: Apply a classification and want to try again with different settings? Use the undo functionality (CTRL+Z) to undo changes in the workspace.
Filtering Point Clouds
Excluding irrelevant points from custom classification processing not only increases the accuracy of the classifications but also decreases processing time as there are fewer points to be considered. There are many methods for filtering points, with two being the most common: filtering by classification and specifying bounds.
The Filter Points tool excludes points from processing based on their classification at a workspace level. This not only prevents the points from being processed but also disables them from the data view. Turning off classifications in a workspace is also useful elsewhere in the software for processes such as creating grid layers and exporting data. To exclude points from classifications without turning them off in the workspace, use the Filter Points button from within the Automatic Point Cloud Analysis tool itself.
Another filtering option in the Automatic Point Cloud Analysis tool is to geographically filter using Specify Bounds. This will limit the tool to a specific area, allowing the tool to only process needed data. Additionally, focusing on a small subsection of data is a great method for troubleshooting classification settings for any classification.
The Custom Classification Tools
Default automatic classifications identify objects in point clouds on the basis of their attributes and structures. Each tool is built to look for specific types of structures in the data to find what they are classifying. For instance, the Ground Classification tool will look for characteristics in points such as areas of low elevation change, the last return, etc. Custom classification tools work on a similar principle, but the attributes and structures they search for in the data rely on the point clusters they were trained on. This training data is identified by the user from within the point cloud. Running a Geometric Segmentation prior to training is recommended.
Classifications Based on Segmentation
Custom classification uses the same machine learning segmentation-based analysis as built-in Max Likelihood classifications to assess point cloud characteristics and find commonalities among the points that make up an object. For example, to segment paint stripes on the road, a user would look for points that make up a flat surface, have the same color, normal values, etc. This method operates on the assumption that each object in the point cloud, each cluster of points identified with segmentation analysis, has a signature made up of attributes and/or structures that differentiate it from its neighbors. The Custom Point Cloud Classification tool takes advantage of these signatures to classify specific and unique features in a point cloud. When training the classification, selecting points by segment will help to ensure that all points have similar signatures.
It’s common to run different segmentation settings to identify different types of objects since they will most likely have unique signatures. The dialog in the image above displays the signature attributes that are common between all of the points that comprise the pylons. Curvature is the most significant attribute, as all of the pylons are flat and end at the 90-degree angle where they connect to the bridge.
For more information on segmentation, see the blogs listed below. The dialog boxes reflect earlier versions, but the settings values are unchanged:
- Point Cloud Segmentation Analysis in Global Mapper Pro
- Classifying Transmission Towers in Point Clouds using Global Mapper Pro
Creating Training Samples
A custom classification tool is trained based on selected segments or clusters of points. These selected clusters are assessed to record patterns in the attributes and the bounds or overall shape. When a tool is trained, its signature can be viewed as numerical data or as a bar chart in the feature model plot. Multiple attributes are assessed.
The Principal Component attribute measures the 3D shape of the points at the neighborhood level (neighborhood being set by the Resolution field toward the top of the dialog). These look at the feature or segment of points as a whole. For example, in the image below, a new streetlight classification is created to identify these curved, tall structures in the data. When applied, the custom tool will segment the data to find features with a similar shape as determined by Principal Component measurements.
Eigentrophy measures the entropy of the point cloud. For example, entropy in a flat surface is very low, but it is high in vegetation. Curvature analyzes the curves created by the points in a local neighborhood. Consistency in curvature values will help indicate that points likely belong to the same object. Normal measures the direction perpendicular to the surface the point is representing. Similar normal values over a local area indicate the points are representing a consistently shaped feature.
Feature Model Plot
Classification signatures are visually displayed in the Feature Model Plot. This bar graph highlights a few of the variables assessed when searching for features in a point cloud. Point cloud segments with value characteristics that fall within the described ranges will be classified. Here, we can see how the Pylon and Streetlight classifications differ from each other per attribute.
Also available with the machine learning functionality of custom classifications is the ability to collect samples and train existing classifications: ground, building, etc. Instead of starting from scratch with a new classification, you can train the existing classes to identify specific feature types, such as buildings with unique shapes. In the Classification and Extraction Shared Settings section, expand any Feature Model and click Train to begin the process.
Blue Marble Geographics was awarded the Outstanding Innovation in Lidar Award for developing a groundbreaking new functionality in training custom automatic point cloud classifications. Custom Classification through machine learning in Global Mapper Pro opens the door for increasing the accessibility and application of lidar and point clouds in many industries without a high-cost barrier.
WORK MADE EASY WITH GLOBAL MAPPER
Want to try Global Mapper? Sign up for a 14-day free trial. You can also request a demo from one of our experts to see this workflow or other Global Mapper processing abilities.
References:
https://www.advancednavigation.com/tech-articles/insights-into-lidar-technology-and-lidar-based-surveying/
https://www.linkedin.com/pulse/lidar-survey-technology-road-infrastructure-santosh-kumar-bhoda/
