LiDAR Data Stitching: How to Improve SLAM with Precision Location

Developers often face challenges when stitching together LiDAR scans for accurate mapping. However, new GNSS technology now offers an efficient alternative to this labor-intensive process.

In just one afternoon, the Point One team built a scalable technical demonstration for globally georeferencing LiDAR scans without stitching, showcasing how high-precision GNSS solutions can develop a universal reference frame for global mapping at scale. So why are developers still doing things the hard way by data stitching without GNSS?

This post will review how SLAM LiDAR functions and highlight how Point One Navigation’s advanced solutions can streamline and improve the accuracy of global mapping efforts.

What is SLAM LiDAR? 

LiDAR (Light Detection and Ranging) is a remote sensing technology that uses laser light to measure distances to the Earth’s surface. By emitting laser pulses and measuring the time it 

takes for them to return after hitting an object, LiDAR systems can create precise, three-dimensional maps of the environment. Originally developed in the 1960s for military and aerospace applications, LiDAR has since found widespread use in fields including topographic mapping, agriculture, surveying, and autonomous vehicles.

SLAM (Simultaneous Localization and Mapping) is a computational problem that involves creating a map of an unknown environment while simultaneously keeping track of an agent’s location within it. In the context of LiDAR, SLAM algorithms process the raw point cloud data generated by the LiDAR sensor to build a coherent map and estimate the sensor’s trajectory.

By using LiDAR sensors, SLAM algorithms can generate accurate and detailed maps in complex environments. However, while the data gathered is comprehensive, translating the local positioning information into globally georeferenced features requires developers to stitch together disparate scans, which is a time-consuming process. 

Point One’s technology addresses this challenge by providing high-precision GNSS solutions that eliminate the need for manual data stitching. By leveraging advanced sensor fusion algorithms, the Polaris RTK correction network and Atlas INS deliver centimeter-level accuracy with <5 second convergence times. 

Thanks to the comprehensive calibration procedures, you can ensure that your location data is consistently accurate down to the centimeter without wasting time stitching together complex data. 

Get in touch today to discover how Point One can revolutionize your mapping and positioning solutions with unmatched precision and efficiency.

What is SLAM used for?

SLAM LiDAR technology is used in a variety of applications, including:

• Autonomous Vehicles: SLAM LiDAR ensures precise navigation and obstacle detection.
• Robotics: It enables robots to navigate and interact with their surroundings.
• Surveying and Mapping: Using SLAM LiDAR, developers can create detailed maps for urban planning, construction, and environmental monitoring.
• Augmented Reality: SLAM LiDAR can enhance AR experiences by providing accurate spatial awareness.

Different SLAM algorithms 

SLAM algorithms are diverse, and each offers a unique approach to the challenges of mapping and localization. Let’s explore a few of them:

• Graph-Based SLAM: This approach uses a graph structure where nodes represent sensor poses and edges represent constraints, optimizing the entire graph to minimize error and ensure accurate mapping. This algorithm is particularly effective for large-scale mapping due to its efficient handling of numerous poses and constraints.
• Extended Kalman Filter (EKF) SLAM: EKF SLAM employs a probabilistic framework that uses Gaussian distributions to model the state and its uncertainties. The algorithm alternates between prediction (estimating new states using motion models) and update steps (correcting estimates using sensor measurements). EKF SLAM is well-suited for systems with a moderate number of landmarks, though it can be computationally intensive.
• Particle Filter SLAM (Monte Carlo Localization): This method uses a set of particles to represent the probability distribution of the sensor’s state. Each particle represents a possible state, and as new data is received, particles are weighted and resampled based on their likelihood. Particle Filter SLAM is highly flexible and can handle non-linearities and multi-modal distributions, making it ideal for complex environments, though it requires a large number of particles for high accuracy, leading to increased computational costs.

Each algorithm provides a different method for achieving precise mapping and localization. However, the computational constraints posed by data stitching can be significant, which is why high-precision GNSS solutions are slowly replacing LiDAR SLAM as industry standard. 

How does LiDAR SLAM work? 

LiDAR SLAM operates by using LiDAR sensors to capture detailed 3D point clouds of the environment. These sensors emit laser pulses that bounce off surrounding objects, creating precise distance measurements. SLAM algorithms process these point clouds to build a map and simultaneously estimate the sensor’s position within that map. Let’s explore each component:

SLAM Algorithms 

SLAM algorithms, such as those outlined above, process the data from LiDAR sensors to ensure accurate and real-time mapping. Graph-Based SLAM constructs a graph where nodes represent sensor poses and edges represent constraints between poses. 

EKF SLAM uses probabilistic models to predict and update the sensor’s position. Particle Filter SLAM employs a set of particles to represent the possible states of the sensor, updating these particles based on observed data.

LiDAR Sensors

LiDAR sensors are pivotal in SLAM applications due to their ability to provide high-resolution 3D measurements. They emit laser pulses in rapid succession, which reflect off objects and return to the sensor, allowing it to measure the distance to various surfaces. 

This results in a comprehensive 3D point cloud representing the environment.

Map Generation

The core of LiDAR SLAM lies in map generation. As the sensor moves, the SLAM algorithm continuously updates the map by integrating new point cloud data. This involves matching new measurements with existing map features, correcting the sensor’s position, and refining the map structure. Effective map generation ensures that the sensor’s location and the environmental layout are accurately represented, facilitating navigation and further data collection.

In sum, LiDAR SLAM integrates LiDAR sensors and sophisticated algorithms to produce real-time, highly accurate maps of environments. However, LiDAR SLAM alone has its limitations. To achieve the utmost precision, RTK (Real-Time Kinematic) technology can be integrated to provide centimeter-level accuracy in real-time.

Challenges in LiDAR SLAM Data Interpretation

Interpreting SLAM LiDAR data presents various challenges, including balancing quality and efficiency, managing precision limitations, and addressing the time-consuming process of data stitching. 

Leaders in the industry, including Tesla, are ditching LiDAR given the challenges presented by interpreting LiDAR SLAM data. In this section, we’ll explore why and how each of these issues emerges, and how alternatives such as Atlas INS can mitigate these challenges.

1. Quality vs. efficiency

Despite advancements in geospatial technology, translating local positioning information into globally georeferenced features remains a challenge for developers.

The need to stitch together disparate LiDAR scans into a cohesive map still often requires the development of a tech stack for accurate data interpretation and referencing -- a time-consuming process for developers who want to build quickly and focus on their main objectives. At the same time, developers need to integrate HD maps to stay competitive and create truly innovative solutions to some of the world’s most complex problems.

This dynamic is forcing many developers to choose between quality and efficiency as they build products and applications with LiDAR data.

2. Limitations in precision

Local referencing of LiDAR data is sufficient for mapping tools where “lose enough” is good enough, such as consumer navigation applications that help individuals find directions to a storefront. But this approximation is inadequate for producing HD maps that can power exciting new developments in autonomous vehicles, robotics, and other solutions requiring high precision.

For AVs and similar devices to navigate safely and accurately in the real world, they need to be built with seamless, globally referenced data.

Any developer working on these types of projects knows that LiDAR point clouds of the same area captured by different sensors will produce disjointed data, usable after first manually stitching them together. Only then can developers begin creating a complex suite of hardware, software, and algorithms that align the data points to each other (local referencing) and the rest of the Earth (global referencing), resulting in a unified reference dataset that precisely represents the physical world.

3. Manually stitching LiDAR data is time consuming and expensive

Traditionally, developers use Simultaneous Localization and Mapping, or SLAM, which aims to estimate sensor poses and reconstruct traversed environments. Using traditional methods to stitch together just two disparately captured LiDAR point clouds is resource-intensive, but nowhere near the effort required to produce HD maps at the scales needed for global solutions. 

If stitching together only two scans can take a developer a few seconds time, stitching together a map of an entire city to safely and accurately power AVs could take years. 

As high-precision maps and LiDAR data increasingly fuel complex solutions at global scales, developers need a tech stack that enables them to build with absolute accuracy quickly and efficiently.

Changing the Way We Georeference Objects With Precision Location

Recent advancements in GNSS solutions are facilitating the high-precision global referencing developers need, eliminating the need to stitch together disparate sensor data. 

Point One’s Atlas is an end-to-end inertial navigation system (INS) that leverages high-precision sensor fusion libraries to deliver accuracy ranging from 10cm to 1cm, and is easily integrated with ROS for streamlined application and product development. This drastic increase in precision is thanks to Polaris, Point One’s RTK correction network that models additional sources of GNSS error to provide map alignment 100x better than standard GNSS.

To demonstrate how developers can use this high-precision data to globally georeference LiDAR point clouds and produce a seamless reference frame, the computer vision team at Point One dedicated an afternoon to developing a proof of concept. First, the team collected data from a Point One Atlas INS. 

Atlas leverages Point One’s Polaris RTK and FusionEngine libraries to capture positioning data with centimeter-level accuracy derived from the industry’s most precise GNSS cloud correction service. They then integrated Point One’s FusionEngine API with ROS to enable the sensor data to be both locally and globally referenced. The resulting data produced a unified, globally referenced 3D map across an area of a city block without using SLAM. 

You’ll notice from the video above, that the output data is a clean map with minimal fuzziness or overlap.

Because this demonstration registers data to an absolute frame at every time step, there is no drift that usually encumbers SLAM using scan matching alone.  Providing the globally referenced data as an “initial guess” for SLAM algorithms, will dramatically cut compute required for place recognition, loop closing, and overall optimization.

We expect that this approach will dramatically cut operational costs related with map creation and updates.

Unlocking Opportunities to Rapidly Scale Global Georeferencing

In just one afternoon, Point One’s engineers were able to develop a scalable tech stack for globally georeferencing point clouds without data stitching.  While this was an internal research and development project, the results show how accessible high-precision GNSS solutions, easily integrated with commonly used frameworks like ROS, can dramatically reduce the time and effort needed to develop a universal reference frame for global mapping. 

This enables developers worldwide to augment traditional data stitching methods, unlocking exciting opportunities for innovation at scale.

The following are just a few examples of solutions developers can build leveraging Point One’s Polaris, Atlas, & ROS tech stack:

• Path planning: Centimeter-level accuracy map data is essential to power navigation systems in not only AVs, drones, and robotics products, but also provides deeper insights to inform logistics and delivery fleets, precision agriculture, and transportation networks
• World mapping & measurement: Global georeferencing LiDAR data at scale eliminates the need to survey for position, replacing traditional mapping methods that require ground reference points or interpolation
• Progress monitoring: Physical details captured by LiDAR and rendered digitally while retaining accuracy can be used to monitor construction sites, land use, and any other area stakeholders need to observe regularly
• Mining operations: Data captured by sensors in open pit mines can be easily georeferenced to enable engineers to take volumetric measurements against reference points in real-time

And of course, the simplification and accessibility provided by this tech stack means LiDAR hobbyists can easily create living models of whatever they capture data for, whether it be their own neighborhood or a race track.

Getting Started With Point One’s High-Precision GNSS Solutions

Thanks to the latest in GNSS technology, it only took the Point One team an afternoon to build this flexible tech stack that eliminates the need for developers to manually stitch together disparate LiDAR datasets. 

The best part? This workflow is easily repeatable by developers working on similar projects. Instead of devoting hours to an extremely tedious process, developers can now focus on building innovative solutions with confidence in the precision of the data they are using.

Making the latest advancements in technology accessible to developers through easy integrations is critical, particularly in the geospatial industry which holds so much potential for having a positive impact on real-world problems. 
Point One makes it easy for developers to build their own tech stack by offering Polaris RTK and FusionEngine APIs that are easily integrated with common systems like ROS, as well as easy-to-use hardware like the Atlas INS to capture data points with precision time and location.

FAQs on LiDAR SLAM

How does SLAM work? 

SLAM works by using sensors to create a map of an environment while simultaneously tracking the sensor’s location within that environment.

What is a SLAM algorithm using LiDAR?

SLAM algorithms use LiDAR process 3D point clouds to build maps and estimate sensor positions in real-time.

Do you need LiDAR for SLAM?

While LiDAR is not strictly necessary for SLAM, it provides high-precision data that significantly enhances mapping accuracy.

What is the difference between LiDAR and visual SLAM?

LiDAR SLAM uses laser sensors to capture 3D point clouds, while visual SLAM relies on cameras to process visual data for mapping and localization.