A new, agile and efficient collision checker in Control

[simonranjith]

A new collision checker, referred to as Spheric Collision Detector (SCD), is proposed as a replacement for Obstacle Collision Checker(OCC) and Autoware Collision Detector(ACD) in Autoware. The method approximates the ego-vehicle and its surrounding objects using a set of spheres, thereby simplifying collision detection to a mere distance evaluation.

The proposed algorithm is similar to OCC; However unlike OCC it does not rely on point cloud data for object representation. Instead it uses the length and width of objects to construct the spherical approximation. This approach has proven effective for detecting collisions, particularly in emergency situations.

Experimental results were simulated using CARLA (0.9.15), Autoware (2024.01) and ROS 2 bridges. They indicate that the proposed approach is at least 200 times faster than OCC.

To top all this, the algorithm was also tested in PixKit 2.0, an autonomous driving and development vehicle using a 5-foot inflated jumbo beach ball.

In the figure, an ego-vehicle is approximated by 3 3D spheres, shown using dashed lines. The stopping/safety distance (3 3D spheres in translucent orange) of the ego-vehicle is defined by the aggregrate, s_b + s_t, where s_b and s_t are the braking and thinking distances respectively. The rectanglar grey area is the footprint of the ego-vehicle. The rear axle center (base link) of the vehicle is marked with a red dot. ENU coordinate axes convention is used for position and movement of vehicles in space: East (X), North
(Y), Up (Z).

[maxime-clem]

Thank you for sharing your interesting work.
If I understand correctly, your approach offers a lowered runtime by approximating the footprint of the objects using circles.

it does not rely on point cloud data for object representation

This makes your approach vulnerable to object detection issues (e.g., incorrect object dimensions, delays, etc), whereas directly using the point cloud data is more reliable (mostly becomes sensitive to hardware/network issues).
Since the point in circle check is very cheap, I believe it would be interesting if your collision checker could also support point cloud.

[simonranjith]

Hi @maxime-clem, Thanks for your reply. It is appreciated. Still I would like to see how this approach would be vulnerable to object detection issues - via some tests. Do you have any suggestions?

Here I'm attaching a small clip of the current implementation.

The collision checker can support point cloud. But here is my concern: if the point cloud data has noise then that can lead to false collision warnings. Don't you think?

output.mp4

[simonranjith]

Hello again @maxime-clem, I'm using the length and width of the detected objects from the topic "autoware_perception_msgs/msg/DetectedObjects". I presume these to be the dimensions of the bounding boxes estimated via clustering + shape estimation algorithms. But the raw input would be the preprocessed point cloud data from the Sensing component. So I'm still not clear when you say there can be object detection issues. Am looking forward to your reply.

[maxime-clem]

You are right that noise in the input point cloud can cause issues, and in practice we would filter the pointcloud.
In the AEB, some filtering is performed in the node, but this could also be done by a preprocessing node.

---

Let me give a better motivation for supporting pointcloud.
On a real vehicle, we may have multiple computers running different parts of Autoware. For example:

• Computer 1: Sensing.
• Computer 2: Perception/Localization/Planning/Control.
• Computer 3: Emergency handling.

If computer 2 fails, then the perception (i.e., detected and predicted objects) becomes wrong.
In such scenario, we want your module to run on computer 3 and detect the possible collision only using the pointcloud data (generated by computer 1).

This is just a simple example.

---

In any case, your module looks good and the video is very nice. Pointcloud support can be added later.
Additionally, the sphere decomposition technique is very useful and it would be great if other modules can easily reuse the decomposition functions implemented in your module.

If you want to open a PR to add your code to autoware_universe, I am happy to review your code and provide help if needed.

[simonranjith]

Hello @maxime-clem, I have opened a PR. Please check. Thanks for your time and help. It is appreciated.