Create hybrid mode of acquisition, such as:

• Stops between points for some image capture;
• Passes all the laserscan for the laser pipeline.

There was a gap showing in each pause in hybrid mode, as shown here:

Initially it was expected to be something related to the code, but finally it was discovered that it is due to the flir_ptu_d46 action server default error parameters, that were too big for the application.

class PTUControl:
    ...
    def _init_(self):
        ...
        self.pstep = rospy.get_param('/ptu/pan_step', 0.00089759763795882463)
        self.tstep = rospy.get_param('/ptu/tilt_step', 0.00089759763795882463)
        ...

Changing each one to a 10x smaller value leads to the following results:

Start from here

Related to #6

This is related to #15.

Just run an acquisition and record a bag file.

There is no need for this segmentation of the laser pipeline, so better if wait_for_done node is merged into the accumulator node.

Here are the results of the first real acquisition.

download

Implement launch file to start joint state publisher.

Should remap to /ptu/cmd topic

The new calibration method for the ptu-laser calibration is on development.

Methodology

1. A pointcloud generated by a initial estimate is segmented into several clusters of points belonging to a plane. This clusters are extracted every time a new pointcloud is generated by the calibration algorithm.
2. For each plane cluster of points, a plane fitting algorithm is done to calculate the RMS of the distance to the plane and the normal of the plane.
3. A set of relations between pairs of planes are given. Right now, the relation is the dot product of the normals of the planes. This is useful for establishing the angle between the planes.
4. A cost function is calculated, using the the relations and the RMS of the planes, dependent of the ptu_laser_calibration. This is expressed by 6 values: the position vector (x, y, z) and the rotation in axis-angle notation (rx, ry, rz) (the angle is the norm of the vector and the axis is the normalized vector).
5. The cost function is minimized. This is currently done by the Nelder-Mead method, which is available in scipy.optimize.

Results

The following figures show a side by side comparison of the the new method (in white) vs the old method (in red):

Conclusion

This method is superior to the old method (Hand2Eye + Radlocc) by the following reasons:

1. It shows better results.
2. It decouples the camera of the system, making it possible to focus uniquely on the geometry, with is now only dependent of this calibrations, instead of 2.

However, there are some cons too:

1. It need a good initial estimate, which can be done by the previous method or by a manual estimate.
2. It requires some manual work in the segmentation of the pointcloud in the different clusters, as well as creating the relations between the planes.
3. The ground-truth need to be very good (for example, the walls need to be planar and need to be perpendicular).

Create a node that accumulates point cloud messages

http://wiki.ros.org/wstool

The final test is to install all that is needed for the lemonbot system with single command in a blank os

The milestones for the radlocc calibration is:

• Change the TF tree (put the laser as a child of the camera)
• Find a manual calibration
• Create a automatic calibration for camera <-> laser with radlocc
• Test calibrations (radlocc and hand2eye)
• Create filter nodes (for point clouds and lasers)

Could also show the cad models ...

The idea is to paint each vertex in the accumulated point cloud (created with the laser) with color given from the image of the camera.

The node will be called point_cloud_colorizer

Hi @bernardomig ,

for your thesis it would be interesting to add a table with 4 or 5 columns, each corresponding to a scan speed. Could show that, the fastest the scan speed, the sparsest the point cloud. This may sound rather obvious but we don't know about other effects, e.g. will the registration fail more?

You can actually make this experiment, prepare the thesis document and fill the table.

Also we could analyse distance between point clouds, for example using a slow speed scan as ground truth. I can give you some ideas on that.

Goal

• create xacro with <unless> calibrated
• get the calibration of both the camera and laser from xacro in calibration/hand_in_eye.urdf.xacro

Acquisition node

is configured using a yaml file and sends the ptu to several positions and acquires an image / point cloud at each position.

Should also have a continuous acquisition mode.

• test the following systems and integrate into all.launch:
  • camera
  • laser
  • ptu

This is a task for Tiago Pereira. Tiago, please accept.