In the applications of LiDAR, rain mist and dust are common sources of noise. To a certain extent, the noise will influence the results of the algorithm, for example, false detection of objects, so we need to filter out the point cloud noise according to the actual situation. Considering this, Livox has built in the point cloud data output by the LiDAR to identify the tag information of multi-echo and noise information to help users cope with the point cloud noise more efficiently.

 

I. What is a Lidar Tag?

 

A Tag is the data that indicates multi-echo and noise information in the LiDAR point cloud data. We define the format of Tag information in the following way:

 

 

 

II. What does the Tag contain?

 

Each Tag information consists of 1 byte and the information contained in that byte are grouped:

Group 1: bit7 and bit6

Group 2: bit5 and bit4

Group 3: bit3 and bit2

Group 4: bit1 and bit0

 

Among them, the second group in the Tag information indicates the echo sequence of the sampling point.

 

Featuring a coaxial optical path, the Livox LiDAR, even if there is no detectable object around, will generate an echo, which is recorded as Echo 0. After that, if there is any object within the detectable range, the first returning laser is recorded as Echo 1, then Echo 2, and so on. If the target object is too close (the distances varies with products, take Horizon as an example it’s 1.5m, Echo 1 will be overlapped and fused with Echo 0 to form a wave and be recorded as Echo 0.

 

When the lidar is in single echo mode, the echo number can only be 00 and 01, indicating that the point cloud is calculated based on Echo 0 and Echo 1 information. When in multi-echo mode, there are four possible types: 00, 01, 10, and 11, indicating that the point cloud information is calculated based on Echo 0, Echo 1, Echo 2 and Echo 3 information respectively. The echo sequence number is the number of pulse signals received by the receiver, not the order in which the laser is sent to the object. Therefore, the echo information in the second group of tags generally cannot be directly used for noise removal. A user can apply it as auxiliary information to the denoising algorithm according to the applications.

 

So, how to use the third group and the fourth group of echo information in Tag to handle the point cloud noise?

 

III. Use Tag to deal with the dust, rain, and fog noise

 

 

For dust, rain, and fog noise, we can use the third group of information in Tag (i.e. bit3 and bit2, to determine if the sampling point is a noise point based on echo energy intensity) to determine if the current point is dust, rain, or normal.

 

There are three confidence levels from high to low: 11, 10, and 01. The lower the confidence level is, the higher the echo energy is, indicating the point is less likely to be a noise. Normally, the laser beam is less affected by the echo energy of the noise from dust, rain, fog, snow, etc. As such, the noise has two confidence levels based on the intensity of echo energy:

01 indicates the weak echo energy and a higher probability of noise, such as dusty point;

10 indicates the medium echo energy and a medium probability of noise, such as rain or fog noise.

Note: The colors in the figure below are marked with tag, instead of reflectivity

 

Tag can be used as important information to distinguish dust, rain and fog noise, but if by relying on it alone, there will still be false filtering or under filtering in some complex scenes. Take the following scene as an example.

 

Actual Scene: Medium or heavy rain, outside the window of the building.

According to the third group information of Tag, dust could be correctly marked in green and rain or fog in red.

 

Actual Scene: Medium or heavy rain, and the waters splashed by the car on the road ahead.

The dust is marked in green and rain or fog in red. However, it could be seen from the figure that there are points on the ground that are mistakenly recognized as dust and fog. In such cases, other point cloud spatial features may be considered for better filtering.

 

Actual Scene: Medium or heavy rain, at an intersection.

The dust is marked in green and rain or fog in red. The same situation in the figure above also exists.

 

IV. Dealing with spatial thread-like noise with Tag

 

Light spots falling on objects with different front and rear distances at the same time may make depth calculations inaccurate and cause spatial thread-like noise.

 

In this case, the fourth group of Tag is used to judge whether it is noise based on the spatial position of the sampling point, and can be used to filter out noises like thread-like noise. Take the point cloud of the ceiling scan in the following figure as an example. It can be seen that there are a lot of threads on the edges of the pendent lamp because part of the light spots fall on the pendent lamp while others are on the ceiling behind it, making the depth calculation inaccurate.

 

 

Test Scenario  (* The image angle is slightly different from the point cloud due to camera FOV’s limitation)

 

The effect is shown by the figure below after filtering with different confidence levels according to the values of Group 4 of Tag.

Filter the point clouds with confidence level 01

Further filter the point clouds with confidence level 10

Further filter the point clouds with confidence level 11

It can be seen from the scene above that the effect is very obvious after step-by-step filtering with Tag.

 

V. Processing the sunlight noise with Tag

 

For sunlight noise, the intensity of ambient light will first be detected by the Livox LiDAR. The corresponding point cloud coordinates will be set to 0 (i.e. filtered out) under excessive ambient light. However, if further processing is required in the algorithm, sunlight noise can be filtered out based on their characteristics, such as lack of obvious local structural features, and large depth changes of the point cloud.

 

As shown in the figure below, sunlight noise in the right image is marked in yellow (when noise is not filtered, LiDAR will first judge it as sunlight noise and set the depth coordinate value of the point cloud to 0 to filter it out). Meanwhile, products such as Horizon and HAP have also been introduced with the sunlight noise self-filtering function.

 

 

However, it is well known that actual application scenarios of LiDAR are usually more complex. Tag provides rich identification information, but relying on Tag only can sometimes lead to incomplete filtering. In this case, better filtering can be achieved in combination with other point cloud spatial features, so as to obtain the point cloud map of actual scene more accurately.