Using depth->scan for Lidar based algorithms in addition to RGB-D VSLAM

[alexk1976]

Hi Mathieu, Our setup has 4 RGB-D (~6 meters range) cameras covering almost 360 FOV. I explore ways to improve loop-detection accuracy that is not perfect now (sometimes there are wrong loop-closures/sometimes no loop-closure where expected). Whether make sense to convert depth to scan data and use Lidar rtabmap capabilities on top of VSLAM/VO rtabmap algos. Particularly proximity detection in loop-closures? Any other recommendations ? thank you

[alexk1976]

Forgot to mention that currently we run classical pipeline for 3 rgb-dd cameras : rgbd-sync -> F2M-> rtabmap.

[matlabbe]

I don't think simulating a fake lidar from the camera will increase loop closure accuracy/precision.

wrong loop-closures/sometimes

There are some ways to decrease wrong loop closures to be accepted, like increasing Vis/MinInliers or decreasing RGBD/OptimizeMaxError.

[dStanhill]

Hi Mathieu, Further pursuing Alex's direction - if we do want to check if fake lidar using depthimage_to_laserscan can help us in cases we have multiple cameras covering close to 360deg. Is multi-camera (=> multi scans) supported ? How do you suggest going about it, merging the multiple scanlines to one? Thanks, David

[matlabbe]

Yes, if you have multiple cameras to increase FOV, a resulting fake >180 deg FOV laser scan could help with loop closures.

You may use a node to convert the laser scan to point cloud, then assemble the points clouds using point_cloud_aggregator with a fixed frame like wheel odometry to adjust frames taken a different time. Feed the combined cloud to rtabmap or convert it back to a laser scan so that rtabmap knows it is a 2D laser scan, not arbitrary 3d point cloud. For loop closure it wont change anything, but for 2D occupancy grid ray tracing, it would be faster in 2D.