obstacle detection and car control

[Benjaminmar8]

Hi, All, How does Autoware detects an obstacle (car, pedestrian, buildings, etc) and slows down or stops the car if it is necessary? Is there something similar to costmap as the ROS navigation stack? Thanks a lot for answering my question.

[mslavescu]

I'm interested in this also, and how easy is to plug new algorithms in Autoware. How can I use the 3D depth map from Autoware mono camera (not LIDAR) ORB_SLAM output to (assist) detect cars, pedestrians, and other objects?

I looking to integrate and extend into OSSDC.org as much as possible of Autoware and ROS existing functionality.

[yukkysaito]

@Benjaminmar8 Autoware has two areas(slows down, stops). If an obstacle exist these areas, it controls to slow down or stop.

Today I add the costmap to Autoware develop branch. However, cooperation with the local planning has not been completed yet. @mslavescu I think ORB SLAM can output pointcloud2 of feature points, this pointcloud2 topic can be substituted for points_raw published from Lidar

[yukkysaito]

https://www.youtube.com/watch?v=tW7S74EUH3Y

[Benjaminmar8]

@yukkysaito Thanks for the reply. Let's say a recorded path is given and the car is supposed to follow this path, will the car slow down or stop without using the costmap?

[mslavescu]

Thanks @yukkysaito!

Do you have a dataset/instructions to reproduce the behavior in the video?

Have you looked into this method? Multi-Level Mapping: Real-time Dense Monocular SLAM https://m.youtube.com/watch?v=qk2ViPVxmq0

How it compares with ORB_SLAM2 (Monocular) used in Autoware?

[yukkysaito]

@Benjaminmar8 In the current implementation, it will. However, we are also developing a method to use a cost map. like this (https://www.youtube.com/watch?v=TA0_0Zb7cjE). this video is not Autoware.

[yukkysaito]

@mslavescu I am interested in this video. But it cannot use to detect objects because it seems there is a big delay. I think it can use create a 3D map.

@sujiwo develop ORB SLAM in Autoware. Therefore he may be able to give good advice. What do you think?

[mslavescu]

Thanks @yukkysaito! They mentioned realtime 3D reconstruction at 30Hz without GPU, would not be that enough to input to detectors (not sure how dense is the point cloud, if it is enough or not)?

@sujiwo would be great if you can give me some advice how to plug some monocular videos in Autoware ORB_SLAM2 implementation, and if it is possible to do 3D reconstruction in Autoware, and maybe plug the generated point clouds into the LIDAR based detectors implemented in Autoware.

I would like to integrate this functionality in the simulator shown here, which I also plan to use in raceai.net to attract gamers into testing OSSDC software:

https://mobile.twitter.com/GTARobotics/status/827250102989422592

[yukkysaito]

@mslavescu for your reference. The following video is ORB SLAM developed by @sujiwo https://www.youtube.com/watch?v=2x_KvXACAgc

[mslavescu]

@yukkysaito thanks! That video looks amazing! @sujiwo would you be able to share some instructions on how to do what you did in the video with your dataset, but also on my own videos, including those from PS3/PS4 driving simulator games?

[gbiggs]

Thanks for using Autoware and for your question. However the issue tracker is for posting confirmed bugs and feature requests.

We ask that you please ask questions on the Autoware Discourse category following our support guidelines: https://github.com/CPFL/Autoware/wiki/Support-guidelines. Please pay particular attention to the information we ask you to provide.