Incorrect Bounding Boxes using pcap file with 3D Perception Stack of Autoware.Auto

asked 2021-02-22 14:43:43 -0600

joshuakurien
1 ●1 ●2 ●2

updated 2021-02-23 09:18:43 -0600

Hi, I have been trying to run the 3D perception stack demo (https://autowarefoundation.gitlab.io/...), but seem to be getting incorrect bounding boxes (see video below). The demo ran in an ade environment using the .aderc-amd64-foxy file. There seems to be a similar issue here, but in my case I am not using the LGSVL simulator. To run the demo the following commands were done:

Terminal 1:

ade$ udpreplay -r -1 route_small_loop_rw.pcap

Terminal 2:

ade$ source /opt/AutowareAuto/setup.bash
ade$ ros2 run velodyne_nodes velodyne_cloud_node_exe --model vlp16 --ros-args --remap __ns:=/lidar_front --params-file /opt/AutowareAuto/share/velodyne_nodes/param/vlp16_test.param.yaml

Terminal 3:

source /opt/AutowareAuto/setup.bash

Terminal 4:

ade$ source/opt/AutowareAuto/setup.bash 
ade$ ros2 launch autoware_demos lidar_bounding_boxes_pcap.launch.py

This is my current rviz visualization: https://youtu.be/ANKaX4vJIFE. The bounding boxes seem to be incorrect and I am not too sure why. Any input would greatly be appreciated. Thank you.

edit retag flag offensive close merge delete

Comments

Can you be more specific? What do you mean by "seem to be incorrect?" Maybe a screenshot would help.

Josh Whitley ( 2021-02-23 08:36:16 -0600 )edit

Yes, sorry about that. Here is a video of my visualization currently https://youtu.be/ANKaX4vJIFE.

joshuakurien ( 2021-02-23 09:20:37 -0600 )edit

Playing back route_small_loop_rw.pcap does show the roads being detected as non-ground objects and thus explains the euclidean clustering algorithm creating a bounding box for it.

The question is, are the parameters for the 3D perception stack set according the appropriate location and orientation of the sensors when the route_small_loop_rw.pcap file was captured? If yes, then what is the explanation for the roads being detected as non-ground objects? If no, then is there another pcap file that could be used to test the perception stack with the current parameters values?

Also, I noticed the orientation of the rear lidar is reversed and it had to be rotated 180 degrees by changing the quaternion in /opt/AutowareAuto/share/point_cloud_filter_transform_nodes/param/vlp16_sim_lexus_filter_transform.param.yaml when testing with the provided file, route_small_loop_rw.pcap.

asobhy-qnx ( 2021-03-18 08:36:47 -0600 )edit

add a comment

answered 2021-02-23 17:27:30 -0600

shlokgoel
23 ●9 ●15 ●13

I think the problem has something to do with the ray ground classifier node. This node is responsible for segregating the filtered point cloud (output of Point Cloud Filter node) into ground and non-ground objects. From the visualization video you referenced, it seems that in your case, the ray-ground classifier node is classifying the entire plane as non-ground. Hence, the euclidean cluster node (which is responsible for clustering and generating bounding boxes) is creating bounding boxes that cover your entire plane.

In order to fix this, I would recommend changing the parameters of the config file for the Ray Ground Classifier Node which can be found here: /opt/AutowareAuto/share/ray_ground_classifier_nodes/param

The config. (param) file used in the demo is called vlp16_sim_lexus_ray_ground.param.yaml

The content of the param file are as follows:

image description

Here you would want to modify the sensor_height_m parameter to get the right results. One thing to keep in mind is that you might have to build the autoware.auto repository for the changes to take place.

Hope this resolves your isse!

Comments

Hi shlokgoel,

Thanks for the suggestion. This did help improve the visualization, but it still needs some more improvement.

joshuakurien ( 2021-02-25 08:32:23 -0600 )edit

You might want to modify the param (.yaml) file for the Point Cloud Filter node as well. That file contains the static transformer (x,y, and z) and that might help you make the visualization better. Also, if you think the answer was satisfactory, you can mark my answer as correct (by pressing the tick mark next to the answer)

shlokgoel ( 2021-02-25 08:40:06 -0600 )edit

I will try that as well, thanks. I am still seeing if there are other ways to improve the object detection.

joshuakurien ( 2021-03-02 08:29:34 -0600 )edit

Why would those parameters have to be changed if he is running the pcap file that was captured using the same settings?

asobhy-qnx ( 2021-03-16 11:30:45 -0600 )edit

add a comment

Incorrect Bounding Boxes using pcap file with 3D Perception Stack of Autoware.Auto

Comments

1 Answer

Comments

Question Tools

Stats

Related questions

Incorrect Bounding Boxes using pcap file with 3D Perception Stack of Autoware.Auto edit

Comments

1 Answer

Comments

Question Tools

Stats

Related questions

Incorrect Bounding Boxes using pcap file with 3D Perception Stack of Autoware.Auto