From 940e837860e1f079c7fdebee4ea0e81fbee2e37c Mon Sep 17 00:00:00 2001 From: Ralf524 Date: Fri, 1 Nov 2024 09:53:09 +0100 Subject: [PATCH] Create Obstacle_detection.md --- .../paf24/perception/Obstacle_detection.md | 48 +++++++++++++++++++ 1 file changed, 48 insertions(+) create mode 100644 doc/research/paf24/perception/Obstacle_detection.md diff --git a/doc/research/paf24/perception/Obstacle_detection.md b/doc/research/paf24/perception/Obstacle_detection.md new file mode 100644 index 00000000..f55fa22e --- /dev/null +++ b/doc/research/paf24/perception/Obstacle_detection.md @@ -0,0 +1,48 @@ +# Research: Why does the car hit obstacles + +## Summary of current obstacle detection in perception: +- [Object detection](#Object-detection) +- [Distance calculation](#Distance-calculation) +- [Publishing of Outputs](#Publishing-of-Outputs) + +For more details, please refer to the current documentation. + +## Object detection: +Active sensors: +- Center camera +- LIDAR + +Inactive sensors: +- Cameras: right, left, back +- Radar + +## Distance calculation +- The vision_node receives depth-images from the lidar_distance_node for the specified camera angle + +- According to the distance-to-objects documentation, the LIDAR sensor's 3D values are projected onto a 2D image that matches the exact dimensions of the camera image. +Numpy formulas are used to reconstruct the depth image, with distances in meters represented as pixel values (grayscale image). +If no distance is found in the depth image, “infinity” is returned for this bounding box. + +- The LIDAR sensor may flicker; a higher spin rate of the LIDAR provieds fully reconstructed depth images but with lower resolution for the LIDAR points + +## Publishing of Outputs: +- Class_index +- Min_X +- Min_Abs_y + +## Dependencies of detected and hit obstacles: + +- Are all obstacles detected? No +- Does successful object detection depend on: + - Speed of the car? No + - Size of obstacles? No + - Type of obstacle? No + - Dynamic/static obstacle? No + +## Important aspects found during research: + +- Sometimes objects are not detected (sometimes e.g.: cyclists, construction site signs, open car doors) +- Most object that are hit are detected +- Sometimes distance calculation is incorrect (distance is set to infinity just before the collision) +- Distance in x is alway > ~0.5 (is that a problem?), sometimes distance to a collided car is greater than the distance to a parked car on the side of the road +- Significant issues in dead-end situations: -> 1.2 seconds of reversing -> the car often immediately collides with the obstacle again