Imagine how easy computer vision would be if the world were two-dimensional? But no, it’s in 3D, which leaves many of us flailing around with ridiculous algorithms, expensive sensors, and a little thing called ‘3D point cloud data’.
Let’s zoom in on that last one for a minute. A 3D point cloud is quite the popular problem in the realm of computer vision. So let’s take a good, hard look at it and how it ties in with the top surveying method: LiDAR.

What is a 3D Point cloud?

For starters, a point cloud is a set of points in space. So, a 3D point cloud is a collection of data points analogous to the real world in three dimensions.
Here are a few key things to understand about point clouds:

  • Each point is defined by its own position and colour. 
  • The points can then be rendered as pixels to create a highly accurate 3D model of the object.
  • Point clouds can describe objects measuring just a few millimetres or objects as large as trees, buildings, and even entire cities.

To collect geographic point cloud data, scanners such as LiDAR (also known as ‘laser scanning’) are the most commonly used thanks to their scarily accurate mapping. If you’re not all that familiar with LiDAR, not to worry, we’re about to dig into everything you need to know about it.

What on earth is LiDAR?

In a nutshell, LiDAR is a remote sensing technology which emits lasers to collect measurements that can later be used to create 3D models and maps of objects and environments.
The major advantage of LiDAR is its unbeatable accuracy, which makes it particularly attractive to those working with driverless vehicles.

How does LiDAR work?

Basically, the LiDAR sensor acts like a bat using echolocation to determine where objects are and how far away they are.

081805 LIDAR

LiDAR shoots pulses of light (up to a million pulses per second), which bounce off a surface and return to the sensor. The sensor then calculates the time it took for the ray of light to return and which direction it came from.

This process ultimately creates a point cloud map of the scanned surroundings. The best part is that the point cloud will store every detail down to the last millimeter, while capturing all the different contours and shapes.

The problem with LiDAR

Like with any state-of-the-art technology, LiDAR has a few rough edges that may give data scientists and researchers a few grey hairs.

1. LiDAR requires visible access to real objects

LiDAR collects data by bouncing rays of light off surfaces, so if there are any objects beyond the line of sight of the sensor, well, it’s going to be a problem.

2. Reflective surfaces can cause problems for the laser

Laser being light rays at the invisible end of spectrum, they follow the rules of reflection. If it targets a reflective surface, the ray will bounce off in a weird angle and can create confusing data.

3. It’s not great at capturing moving objects

Consider a self-driving car taking a turn on the road. While zooming along, it records a bus overtaking it during the turn. This could result in the bus being recorded as twice the length.

Not to mention turning at high speeds tends to add noise to the data.

4. Bad weather can interfere with data collection

LiDAR is not a fan of bad weather. Thick fog and raindrops can interfere with the lasers by causing reduced range and refraction, meaning the LiDAR won’t be as effective.

5. It’s a costly investment

Nothing good comes cheap, and that principle certainly applies to LiDAR sensors. While units are currently priced around 80,000 USD, the cost can potentially tumble down as low as 1000 USD due to the newer developments by companies such as Velodyne, Quanergy, LeddarTech.

Applications of 3D point data

Now that you’ve gotten the 101 on point clouds and LiDAR, let’s take a look at how you can use this cool data.

1. Autonomous vehicles

3D point data is ideal for self-driving cars. These nifty vehicles need to understand their environment and perform SLAM in real time in order to be able to move.
A real life example includes Google’s Waymo. Waymo uses high frequency LiDAR sensors mounted on vehicle tops to create maps and understand their environment. (Fun fact: Google reportedly cut project costs by 90% after adopting LiDAR.)
If you want to know more about how sensors are used in self driving cars read our post on the Visual Recognition Sensors Self-Driving Cars Use To Map Out Their Environment.


2. Applications needing highly accurate data

LiDAR is mostly known for its ability to deliver measurements with a tremendous degree of detail. If your project is headed for a real world application where accuracy is practically a life or death situation, then you’ll want to use LiDAR to capture even the most minute details.

Here’s an animation where you can see the richness of the mesh model generated by LiDAR.

3. Exploration and dangerous situations

Did you hear about the discovery of an ancient civilization hidden in Northern Guatemala? Well, that was only possible due to LiDAR sensors on a plane which mapped the dense forest and revealed subtle signs of the archaeological wonder.

While archeologists are rejoicing at the potential of LiDAR technology, the same benefits can be enjoyed in other situations where the ground isn’t safe for humans to map manually. With LiDAR, robots can scan, generate and help us navigate risky environments like quarantined areas or a town after a natural disaster.

Challenges with 3D point cloud data

1. Huge file size

If you’ve already researched point cloud data sets, you’ve likely come across the following:

Those are just to name a few. These data sets are publicly available and easily downloadable.

2. Designing a User Interface

Since we are dealing with 3D instead of just 2D, you need a much more complex user interface (UI) that is both usable and comprehensive. That’s not an easy balance to achieve. The learning curve for 3D-focused UIs tends to be much steeper too.

3. Expensive Sensors

Point cloud data is collected using different kinds of lasers, and lasers are expensive. We can expect the costs to go down as time passes and newer technologies arise, but for now, lasers are still a hefty investment.

Low-cost options are definitely available, but you do get what you pay for when it comes to sensors. You may want to dig deeper into your wallet if you’re searching for a sensor with high resolution that will give you superior data.

Speaking of high quality data, if you’re in the market for custom annotation services to suit your project’s specific needs, drop us a note to find out how our team of dedicated professionals can help you bring your training model to the next level.

Originally published on September 03, 2018 Topics: Machine Learning Deep Learning Autonomous driving Computer vision


You may also like: