Cluster in radar data

#1

Hi,
In the radar documentation, it says that ‘The reflected signals are
processed and after multiple steps they are available in form of clusters and objects.’ Does one cluster here actually means one radar scatter point? I have this doubt because from the radar fields provided in https://github.com/nutonomy/nuscenes-devkit/blob/master/python-sdk/nuscenes/utils/data_classes.py#L294, the basic elements of radar data structure are positions of points and their RCS/velocities in array. My understanding is a cluster means a group of radar scatter points which belong to one object, it’s done by clustering algorithm like K-means, etc. I am confused with this. Could you clarify this for me please?

Regards,
Kathy

#2

Hi,

Does one cluster here actually means one radar scatter point?

Yes, the terminology is confusing. Every radar (scatter) point/return is a cluster. From the manual:

“The ARS sensor uses radar radiation to analyze its surroundings. The reflected signals are
processed and after multiple steps they are available in form of clusters and objects.
Clusters are radar reflections with information like position, velocity and signal strength.
They are newly evaluated every cycle. In contrast to this, objects have a history and
dimension. They consist of tracked clusters.”

My understanding is a cluster means a group of radar scatter points which belong to one object, it’s done by clustering algorithm like K-means, etc.

There is no actual clustering algorithm (e.g. k-means) happening here, just filtering steps.

#3

Thanks for your reply, that’s very helpful!

#4

Hi,
I have thought the whole acquring process of radar data should be like this:
1). the sensor recevied the reflected signal and processed it with DSP; then original point cloud is produced;
2). the original cloud will be further processed by the radar manufactory with clustering algorithm (e.g. k-means) and other criteria ( e.g. RCS); then output the object information .
3). we received the object information via CAN bus in the form of bit-wise operations

I’m not sure whether the process is right . And how about the radar data in nuScenes, is it final object information or original radar point cloud?

#5

I admit that this is confusing, but as quoted from the manual above, the so-called “clusters” are the “original radar point cloud” (after some filtering).

#6

Hi Cubloc,

I think you can understand as follows: the radar data from nuScenes dataset is called cluster, not because it uses clustering algorithm in pattern recognition theory like k-means etc., but clustering algortihm in radar signal detection field, one cluster is generated by combining multiple radar peaks to one point, this cluster then gets a position, RCS value and a velocity derived from the doppler effect. In a further step multiple clusters might be combined to an object, which you can use clustering algorithms in pattern recognition theory. On this level tracking algorithms/feature extraction algorithm can be used to determine some properties of the object. I hope this is helpful for you to understand.

2 Likes
#7

The whole process is right, but there is a little mistake.

Semiconductor manufacturer(such as Ti, NXP) will supply some core hardware(RF, DSP, and others)。Then radar manufactory(Boash, Conti, Veoneer, Aptiv …) integrate these on PCB and sell the finished product which is also called Radar.

The Signal Processes in Radar have serval phases:

  1. convert the IF-Sigal from RX to original points(relative-vel, range, angle, and so on)
  2. cluster the points from 1. I guess this step cluster some very closely original points into one, which may due to the resolution of FFT. I think we can call the output of 2 to optimized IF-Signal points, which is also the raw data.
  3. cluster the raw data of 3 to object.
  4. transform the objects in 3 and raw data in 2 to controller via CAN.

The Radar of Boash, Conti, Veoneer, Aptiv will supply the Info which come from 4 to consumer. But the radar pcd files in nuScenes only supply the raw data of step 2.

I think that’s reasonable, because of the algorithm of cluter in step 3 is secret for every company.

We can use the raw data to test cluster algorithm.But if you’re focus on tracking、controlling, you have to choose one way to do that youself.

1 Like