UQRacing / lidar_processing

LiDAR processing (depth map generation) for cone detection on the AV
Mozilla Public License 2.0
2 stars 0 forks source link

Online Segmentation Implementation #11

Closed RileyBowyer closed 2 years ago

RileyBowyer commented 4 years ago

Overview

Many different approaches can be taken to identify cones, this issue will detail progress into the development of online segmentation, following the implementation as described in Efficient Online Segmentation for Sparse 3D Laser Scans.


Working Branch: feature/online_segment

Main Tasks

There are four main tasks to complete in order to successfully be able to use this algorithm to identify cones.

Conversion into Spherical Depth Image

As this algorithm relies on the use of spherical co-ordinates, the pointcloud must be cast to this format. In addition, a depth image must be constructed to simplify the comparison of neighboring ranges.

Ground Removal

Following the process of the aforementioned paper, the detection and removal of the ground is the initial segmentation step. This involves creating a matrix of angle increments constructed from the columns of the depth image, smoothing said angles and conducting a search to identify points that belong to the ground.

Segmentation

Following the removal of the ground, the remaining points can be segmented into individual objects by computing the angle of a the plane connecting the points in consecutive columns.

Classification

Once a list of segmented objects are found, this list must be classified to identify what is and isn't a cone. This will consist of a range of methods to be determined once the segmented clouds are computed.

RileyBowyer commented 4 years ago

Placed on pause, pending decision on sensor.

RileyBowyer commented 4 years ago

With the move to ring based sensors, this method has been revived.

mattyoung101 commented 2 years ago

Archiving as legacy