search

robotics-upo / ars548_ros

ROS2 driver for the Continental ARS548 Radar 4D
BSD 3-Clause "New" or "Revised" License
21 stars 4 forks source link

Problem with object classification #2

Open angeladifazio opened 3 months ago

angeladifazio commented 3 months ago

Showing the topic /ObjectList to check which classes the identified objects belong to, they are classified only as "hazard". Is there a way to obtain a correct object classification, considering also the other possible classes (pedestrians, car, truck ..)?

david-alejo commented 3 months ago

Hi Angel,

Could you give us more details on the experimental setup that you have used in the experiments? So that we can replicate it in our lab.

Best regards,

David

angeladifazio commented 3 months ago

Hi David,

We used the ARS548 radar with the Jetson AGX Orin and we placed it at one meter high, on a desk, in front of some cars (as depicted in the photos). The radar was able to detect cars and to show them as point clouds, using the proposed node. However when we printed the messages published on the topic, we obtained only "hazard" and not "car", as classification.

cars

ars548

arsa548_side

ObjectList

david-alejo commented 3 months ago

Hi again Angela,

Thank you very much for your quick response and your feedback

I have looked throughout the driver to see if there is some bug inside but I could not find any. In this case, what we do is just to translate the fields of the information that is received to a custom ROS2 message with the same fields.

The classification is done internally by the ARS548 radar with a non-documented method. We have recorded some bags for traffic monitoring keeping the sensor in a fixed position and it is truth that the only field that varies is the hazard one, the others remained at 0 during the whole experiment.

I cannot say whether it is a problem on the classification method or our driver. I guess that the training is done with a RADAR sensor onboard a car, so probably when used this way the other classes should appear.

In your case, it seems that the objects are relatively close (7 meters) and probably the classification outputs hazard under these conditions.

We will conduct some more experiments to further check it, I will give you some feedback ASAP. In addition, any ideas are very welcome so that we can fix this issue or at least figure out why this output is happening.

Thank you again. Best,

David

david-alejo commented 3 months ago

Edit: I have checked in some bags and for example the pedestrian probability raises from 0 to small values. Therefore, it seems that the problem then is the dataset.

However, if you don't mind, please keep us in the loop of the results obtained and if you successfully get some labels right.