Noise Ring under Clear Weather Condition ?

[OuYaozhong]

Hi, there, @MarioBijelic

I found a strange phenomenon of this DENSE Seeing through Fog Dataset.

Normally, under the bad weather, such as snow, fog and rain, there will be a noise ring around the LiDAR sensor, which can be seem under the collected point cloud.

However, I also found that noise ring (a ring around (x, y): (0, 0)) under the point cloud under clear weather (or to say under clear split).

And, the noise ring disappear in some point cloud under bad weather split (e.g., fog split).

Here I give some example, where I plot the data from camera and LiDAR, and the file name is shown as title in the picture.

(dense) USER@HOSTNAME~/SeeingThroughFog/splits$ cat train_clear_day.txt | grep 2019-09-11_18-53-00,00270
2019-09-11_18-53-00,00270
(dense) USER@HOSTNAME:~/SeeingThroughFog/splits$ cat dense_fog_day.txt | grep 2018-10-08_08-18-59,00120
2018-10-08_08-18-59,00120

According to my observations,

I guess that this ring appears maybe caused by the laser hit the roof of the car or the auxiliary for mounting, since the LiDAR sensor is not mount high enough. And there are some point clouds that noise ring is missing, I guess maybe the position of the LiDAR sensor has been ever adjusted during the data collecting.

Could you confirm that if my observation is truly existed ?

And any explanation can be given for this strange phenomenon?

Thank you in advance.

[MarioBijelic]

Hi @OuYaozhong ,

Thank you for your detailed observation and for providing the examples—great analysis!

You're right in noticing the ring-shaped artifact near (x, y) = (0, 0) in some point clouds. This is not due to any external obstruction like the car roof or mounting accessories. Instead, the phenomenon is caused by how the LiDAR handles non-returned points in dual return mode. Most often those points correspond to laser rays being either directed towards the sky or low SNR objects. Instead of not returning any point in the operation mode we chose here, the laser sets it to a minimum distance which corresponds to the laser calibration parameters.

Hope this brings a bit more clarity.

Best,
Mario

[OuYaozhong]

Hi, @MarioBijelic ,

Thanks for your explanation.

Though this explanation seems reasonable, I still confuse with some points.

🌟Question 1:

According to your explanation, it can still be divided into several situations,

Your explanation seems to imply that in dual-echo mode, when there are no object, or a low signal-to noise ratio (SNR) object, in the laser's direction within detection range,

1. Sensor will return the FIXED minimal detection distance instead
2. Sensor will return some special signal that indicates it is not detecting an object (no object or a low SNR), and Seeing through Fog team set them as minimal detection distance in post-processing ?
3. Due to the low SNR or the absence of any object, and the sensor must select a peak as the distance, it will select the most credible peak (of course it is also fake) as the result among a bunch of fake peaks, but it often appears close to the sensor for some reason, thus forming a circle of artifacts.

Which one is the ultimate reason for the appearance of this artifacts ? (1, 2, 3 or other)

🌟Question 2:

Regardless of the answer to the Question 1, it seems that from your explanation, it can be inferred that regardless of whether the laser hits the object with not too low SNR, it will return a distance, the only difference is whether it is real or an artifact. If this is the case, then the number of points in each frame of the point cloud should be roughly the same, but I counted it an it seems not to be the case.

I use the following code to visualize the histogram of the number of point of each point cloud frame.

from pathlib import Path
import numpy as np
from matplotlib import pyplot as plt

dataset = "STF" # STF or KITTI object

if dataset == "STF":
    lidar_root_path = [
        Path("$STF_ROOT/lidar_hdl64_strongest")
    ]
    num_features = 5 # the -1 dimension of the point cloud, for STF is 5, for KITTI is 4
elif dataset == "KITTI":
    lidar_root_path = [
        Path("$KITTI_OBJ_ROOT/training/velodyne"),
        Path("$KITTI_OBJ_ROOT/testing/velodyne")
    ]
    num_features = 4 # the -1 dimension of the point cloud, for STF is 5, for KITTI is 4
else:
    raise NotImplementedError



pc_point_nums = []
for path in lidar_root_path:
    for file in path.iterdir():
        pc = np.fromfile(file, dtype=np.float32).reshape(-1, num_features)
        pc_point_nums.append(pc.shape[0])

hist, bin_edges = np.histogram(np.asarray(pc_point_nums, dtype=int), bins=100, )

print(f"Sum of hist: {hist.sum()}")


fig = plt.figure(dpi=300, figsize=[10, 6])
ax = plt.subplot()
ax.plot(bin_edges[:-1], hist / hist.sum())
ax.set_xlabel("Point Cloud Point Count")
ax.set_ylabel("Percent")
ax.text(x=0.01, y=0.99, s=f"Total: {hist.sum()}", transform=ax.transAxes, ha="left", va="top")
ax.set_title(f"{dataset}")
fig.savefig(f"./{dataset}_point_cloud_point_count_histogram.png", bbox_inches="tight")

The result (count for Seeing Through Fog and KITTI object):

It seems that the number of point in each frame varies obviously.

Though I also check the histogram in KITTI object, and shows that KITTI also have the same problem, but it is normal for KITTI since not any artifacts appear when the laser hit nothing or low SNR object.

🌟Question 3:
Most important and most concerned thing is that, how can we remove this kind of artifacts ?

Thank Mario in advance.

Best,