How “Magically” Increase LiDAR’s Range?
In my previous article “How far should a LiDAR “see”?”, I raised two questions:
- What is the required detection range for a LiDAR system in an automotive application?
- How should we define the target detection range for a LiDAR system, and what are the appropriate test conditions?
Now is the time to pay my debt and answer the second question, which is essential when we wish to compare “apples to apples,” evaluating the performance of two different systems.
One may wonder why is this such an important question? For example, when you wish to compare the acceleration rates of various cars, you can look at the numbers in the datasheets. So why can’t you do the same with LiDARs?
The reason is that there is a standard set of requirements both from the vehicle and test scenario when measuring the acceleration. However, when measuring the performance of a LiDAR system, there is currently no standard at all.
Occasionally, the marketing department rightfully requests R&D engineers to present performance results that surpass the competition. However, not all LiDARs are made equal, and the physics of some of them don’t allow full compliance with such performance requirements.
How is called something that contradicts physics? Right, Magic.
Accordingly, these R&D engineers need to apply some “magic” to fulfill marketing requests and support the prosperity of their company, unless they have enough time to go back to the drawing board. However, since I’ve not yet seen a LiDAR company requiring graduation from Hogwarts's magical school in its job description, these engineers need to look for another kind of magic. Accordingly, they practice the magic of mathematics and leverage the lack of LiDAR standardization for their cause.
This article will cover the main characteristics that a reader should look for in a datasheet to verify that he gets an accurate picture of the LiDAR system range detection performance, uncovering any “magical” occlusions.
This article has two main goals:
- Describe the main parameters one should seek between the lines of a LiDAR system datasheet to distill the actual detection range performance.
- Encourage the industry experts to agree upon a common standard in testing and reporting LiDAR systems performance values.
Let’s assume a general LiDAR system that can detect a Lambertian target up to 150m but is required to detect at 200m. The nominal test conditions are:
Target:
- Reflectivity: 10% (returns 10% out of the incident light)
Ambient conditions:
- Sunlight illumination: 100klux, at 37 degrees tilt angle
Detection Confidence level:
- Probability of detection: >95%
- False alarm rate: <0.1%
LiDAR system mode:
- Frame rate: 25Hz
- Angular resolution: 0.1deg x 0.1deg
In the following chapters, I’ll describe how changing each of the parameters above influences the resulting detection range, in most cases without any need of changing a single bit in the LiDAR system itself.
Target
The detection range of a target is proportional to the returning laser power from that target. Therefore, we can estimate the implications of the various parameters on the detection range through the change of the reflected power. We can easily calculate the returning power using the General LiDAR Equation:
Accordingly, if we assume a target with a reflectivity of 50%, instead of 10%, we’ll improve by x5 the received power, or by sqrt(5) the detection range R, improving it from 150m to a whopping 335m, without changing anything in the system!
Conclusion: By changing the reference target reflectivity from 10% to 18%, we can “magically improve” the reported detection range of our LiDAR from 150m to 200m.
Ambient Conditions
Unlike camera systems that need an external light source to detect the view, the opposite is true in LiDAR systems. Ambient light is one of the primary noise sources for LiDARs, decreasing the system's overall SNR (Signal to Noise Ratio) and the maximal detection range accordingly.
The sunlight is reflected from the target and enters the LiDAR system at the same angle as the laser beam. Since the system sensor can’t distinguish between sun photons and laser photons, the signal might be non-distinguishable from the noise. This is the case of a direct sunlight.
Since you would probably want to use your LiDAR system during sunny days, it must comply with the requirements of direct sunlight scenario. The luminous flux per unit area of the ambient noise is measured in Lux units. Full direct sunlight is considered to be 100klux:
Although the random sunlight noise influences the SNR only through a square root of 2, the difference in LiDAR performance in full direct sunlight (100 Klux) and an overcast day (1 Klux) is enormous:
Meaning, the performance of a LiDAR during an overcast day is ten times better compared to the required conditions of full direct sunlight.
Conclusion: By changing the reference ambient light intensity from 100 Klux to 56 Klux, we can “magically improve” the reported detection range of our LiDAR from 150m to 200m.
Detection Confidence Level
The detection confidence level depicts the probability of detecting a target vs. the probability presenting a false alarm. For example, suppose we require a 95% probability of detection. In that case, it means that 950 out of 1000 times we fire a laser pulse at the target in a specific pixel, we’ll adequately distinguish its returning signal. On the other hand, a false alarm rate of 0.1% means that only in 1 out of these 1000 measurements our system presents a wrong distance at some random range, which is distant from the actual target range. In the rest 49 times, the system doesn’t show any target at all in that pixel.
The confidence level is defined according to the noise threshold. A measurement above this threshold is considered a target and presented in the point cloud of the LiDAR. While below the noise threshold, the measurement is not shown.
Accordingly, we can loosen our confidence requirements by changing the height of the noise threshold. For example, if we require only a 50% probability of detection, along 5% false alarm rate, we can accept much lower SNR values, which correspond to longer detection ranges.
The relation between the change confidence level and improvement in detection range is a bit less straightforward than in previous cases, therefore out of the scope of this article, but the trend is similar.
Conclusion: By loosening the confidence level requirements, allowing lower detection probability and higher false alarm rate, we can “magically improve” the reported detection range of our LiDAR from 150m to 200m.
LiDAR System Mode
In terms of “system mode,” we’ll consider possible changes of two parameters: system frame rate and angular resolution (binning configuration).
Frame rate
Let’s assume our system fires N laser pulses per pixel. Reducing the system frame rate by x2.5 means increasing the duration of each pixel by the same factor. If we don’t change the laser's Duty Cycle (DC), we can fire 2.5*N pulses per pixel. Accordingly, the accumulated signal power is x2.5 higher, increasing the SNR of the system by sqrt(2.5).
Therefore, once we evaluate system performance at a 25Hz frame rate, we can’t directly compare it to the performance of another system working at a 10Hz frame rate.
Conclusion: By changing the system's frame rate from 25Hz to 14Hz, we can “magically improve” the reported detection range of our LiDAR from 150m to 200m.
Angular resolution
Some LiDAR systems include a feature of angular binning. Angular binning means an SW summation of several neighbor pixels:
That way, it’s possible to increase the total received signal per pixel (by four times in the figure above). Accordingly, since we sum the signal and the noise (which is random), we improve the SNR only by sqrt(4)=2.
Therefore, once we evaluate a range performance of a system with a specific angular resolution, we can’t directly compare it to the range performance of another system with another angular resolution.
Similarly, for example, a system with a FOV of 120x25 degrees can’t be compared with another system that provides a FOV of only 40x25 degrees. This is because the required power to cover a whole FOV of 120deg is x3 more than for 40deg FOV. For that reason, some systems offer different configurations. One configuration with a narrow FOV but longer detection range. Another configuration with a wide FOV but short-range detection.
Conclusion: By reducing the angular resolution by 2 (for example, by summing each two neighbor pixels), we can “magically improve” the reported detection range of our LiDAR from 150m to 212m.
Since the scope of this article is limited, I left out additional system/test-setup parameters, such as:
- Overlap ratio between the laser beam and the target
- Ambient temperature
- Positioning of the target in the FOV and its tilt angle
Summary
In this article, I showed that it’s straightforward to “adjust” the LiDAR system performance per marketing needs by playing with the test conditions without the need to change the system design itself.
Therefore, until all industry players agree on a common standard for presenting and measuring LiDAR performance, we need to look in between the datasheets' lines to understand the actual performance of each system.
Feel free to share your thoughts in the comments and contact me on LinkedIn.
