Since their debut at the DARPA Grand Challenge in the early 2000’s, mechanical LiDARs have been the mainstream standard for autonomous vehicles for almost two decades. In early 2020, Livox unveiled the first-ever hybrid solid-state LiDAR, Livox Horizon, at the CES. Besides its accessible price tag in the $1,500 range, the product introduced a new technology for self-driving sensor algorithms: non-repetitive scanning with rotating-polygon mirrors.

 

Livox Horizon’s Scanning Solution with hybrid-solid Rotating-polygon Mirrors

 

In theory, point clouds from non-repetitive scanning patterns are achieved by adding two to three rotatable polygon mirrors on the laser path, then controlling the LiDAR’s scanning range and location in the physical space using the refraction of light. Over the years, Livox has conducted extensive R&D and made remarkable breakthroughs in motor bearings, glue, and production processes, which enabled it to deliver performance equivalent to high-beam mechanical LiDARs with fewer laser transceivers. Meanwhile, the solution is able to concentrate point clouds in ROIs through precise control of the motor rotation mode, to provide more focused vision. With non-repetitive scanning pattern, a LiDAR’s point cloud resolution improves over the integration time. Therefore, the multi-frame overlay method is becoming the new trend for sensor algorithms due to its ability to achieve denser point clouds.

 

Livox HAP Point Cloud Simulations

 

Despite the widespread attention received by Livox Horizon since its inception, we have been aware of some users’ concerns over the past two years regarding the use of non-repetitive scanning LiDARs:

Does non-repetitive scanning work? Is it really effective?

Soon after the release of Livox HAP, we reached out to a few self-driving algorithm engineers who have been using Livox LiDARs in the past two years. Through their feedback, we hope to offer the industry diverse perspectives on the technology from real users.

 

 

01.  

"An impressive limit range."

 

Interviewee: Mark
Occupation: SLAM algorithm engineer
Professional Experience: 3 years

 

Previously, Mark had been using mechanical LiDARs to develop SLAM-related algorithms. Two years ago, his team purchased Livox Horizon, and its new scanning method piqued his interest and led him to start using the device.

 

“Because I had always been working with mechanical LiDARs, it took me a bit of time initially to get used to non-repetitive scanning with Livox Horizon. At the beginning, when I tested it on on edge and surface feature extraction algorithms, the shapes of the features extracted were irregular. Then I spent one to two days making simple adjustments to the parameters and strategies, and realized they actually worked better than mechanical LiDARs. Plus, non-repetitive scanning can generate maps with much denser point clouds.” Mark added: “The more I learn about Livox LiDARs, the more I realize that non-repetitive scanning has a very high upper limit, which however does not compromise its lower limit. The most important thing is to do some adaption at the beginning, and you’ll find that it actually involves a lot less work than you imagine.”

 

Mark’s experience is likely similar to most LiDAR engineers’. Over the past 10 years, the market only had mechanical LiDARs, hence the large amounts of data and open-source algorithms available on the technology today. As the new kid on the block, non-repetitive scanning will certainly have to go through a transitional period before it is fully embraced by the industry. For this reason, Livox has released a number of open-source tools and algorithms over the past two years (https://github.com/Livox-SDK) to help new users adjust to the new device quickly and bolster the whole Livox ecosystem.

 

In 2021, we created a robust LiDAR-inertial odometry named the Lio-Livox. With this new advanced solution for LiDAR SLAM, Livox LiDAR can achieve high-precision localization and mapping in extreme scenarios. Users only need to configure the device for a day before testing it on-site.

 

The Lio-Livox Algorithm: Works in Traffic Jams, over Highways, and through 4km Long Tunnels

 

Compared to other similar algorithms, the Lio-Livox algorithm’s distinct advantages include its high robustness, efficiency, accuracy and short development cycle. As Livox HAP is being shipped out to customers, Livox has made HAP compatible with the Lio-Livox algorithm released last year (https://github.com/Livox-SDK/LIO-Livox) to help users quickly integrate it with the new device. 

 

 

02.  

”The dense point clouds mean that non-repetitive scanning is more than up to the job.“

 

Interviewee: York 
 Occupation: Detection algorithm engineer 
 Professional Experience: 4 years

 

Before learning about Livox LiDARs, York had been using 64-line mechanical LiDARs to detect and classify high-speed objects with deep-learning algorithms.

 

“I might be different from other algorithm engineers, but when I first saw the point clouds generated by a Livox LiDAR, I thought it would be up to the job. At the time, our team was not too concerned with the issue of affordability. To me, as long as the LiDAR’s point clouds were dense enough, we could count on it to work.” In York’s view, deep learning-based target detection depends mainly on data and network algorithms. So long as the neural network can learn effectively, it will be able to master any repetitive or non-repetitive scanning patterns. “The rest depends on the data accumulated,” he said. As a test, York applied a model trained by a 64-line mechanical LiDAR on Livox Horizon. He discovered that the Livox LiDAR could detect vehicles, pedestrians and other objects instantly and effectively. Out of curiosity, York paired a model trained by Livox Horizon with Livox Tele-15 which has a longer detection range. The test showed that the device was able to accurately detect a vehicle at 300 meters away. York added, “I shared the result later with the team, and everyone agreed that point cloud patterns were only a small factor considered by algorithms. Detection algorithms would be able to work with a LiDAR as long as its dot frequency or point cloud density has reached a basic threshold.”

 

Applying Algorithm Models Trained by 64-line Mechanical LiDARs on Livox Horizon

 

Another challenge to non-repetitive scanning is the blurring of images. On the issue, York said: “We treat it as a part of the process. For object detection, with the nuScenes dataset from 32-line mechanical sensors and Waymo Open Dataset from 64-line sensors, we would normally improve precision by overlaying several frames of point clouds. The trade-off for such density would be a longer blurring of the object, which was a common practice in the industry. Besides, a dynamic object produces a more obvious afterimage compared to a static scenario. This actually helps differentiate moving objects from still ones, and thus determine the object’s motion at the time.”

The overlaying of point-cloud frames described above works the same as the concept of integration in non-repetitive scanning, which is intended to capture more detailed spatial information to produce clearer images.

 

“Based on our tests on Livox HAP, we found that the device would cause an afterimage of less than 2m long when identifying high-speed objects. In a low-speed scenario, the blurring was less than 0.5 m. Such a small difference will not affect the detection accuracy of the system in any substantial way,” York added.

 

Blurring in High/Low-speed Scenarios

 

Xpeng has been launching a host of improved functions recently, such as their NGP and enhanced ACC/LCC-L, which shows that non-repetitive scanning LiDARS can support advanced detection algorithms for autonomous driving. For the benefit of HAP users, Livox has released the open-source Livox_Detection V2 algorithm (https://github.com/Livox-SDK/livox_detection). Available on Github, the algorithm is compatible with HAP and all other Livox product series.

 

Successful Integration of SLAM and Target Detection Algorithms with Livox HAP 

 

Users can migrate other common 3D detection algorithms such as SECOND, Pointpillars, SA-SSD and train them directly in datasets captured by HAP, with nearly identical results. Horizon users can even move data previously captured by the older device and utilize them directly on HAP. With technological change comes a shift in user habits, and as LiDARs evolve in the future, newer forms of point clouds will continue to emerge. As these changes occur, Livox believes that an ecosystem based on open-source algorithms will make transitions more hassle-free and bearable for users. Besides making LiDARs more accessible and affordable through product design and supply chain advantages, we will continue to develop algorithms that help users take advantage of new technologies easier and faster. Livox also looks forward to seeing more engineers like Mark and York work together to promote cross-industry use of hybrid solid-state LiDARs.

 

Click Here to go to the official Github page of Livox, where you can learn about and download the latest open-source algorithms from Livox.