Differences

This shows you the differences between two versions of the page.

--- stereo_cameras [2019/02/28 17:44]
paul [Stereo Cameras]
+++ stereo_cameras [2019/03/31 14:49] (current)
@@ Line 1: / Line 1: @@
 ====== Stereo Cameras ======
-Here are notes of various Stereo Cameras I've used and built. I used a number of them on [[primo_robot|Primo]] to recreate the outside world.
+Here are notes of various Stereo Cameras I've used and built. I used lots of them on [[primo_robot|Primo]] to recreate the outside world.
 {{::cleaned_up.gif?direct&600|}}
-===== ZED Camera =====
-==== Specs ====
-**Name:**       ZED
-**Baseline:**   120mm
-==== Setup ====
-I used the ZED in Primo running Ubuntu 16.04. I used the ROS driver package
-provided.
-  * OpenCV was installed already
-  * Installed cuda using the debian. Ran into an X server issue and ''%%glxgears%%'' didn't work. Restarted and it fixed everything
-  * Downloaded the ZED sdk run file. Made it executable and ran it.
-  * It installs in ''%%/usr/local/zed%%''
-  * Ran the ZED Explorer in the tools section of the install dir
-  * Updated the firmware from 922 to 1142. After I updated i did notice some frame glitching every now and then
-  * Went to the samples and built them using CMake
-==== ROS Install ====
-  * Installed the source from github as a package in the catkin_ws workspace.
-  * I changed the FPS in the launch file to 15.
-  * Also can play with the confidence level.
-==== General Notes ====
-Quality of the point cloud doesn't look great, and lots of "fall offs" on edges.
-With that said I didn't see any speckles and CPU usage is VERY low especially
-with the lower framerates. It is a very fast little device.
-==== Software ====
-Found the resolution definitions here:
-<code c++>
-RESOLUTION_HD2K,   /**< 2208*1242, available framerates: 15 fps.*/
-RESOLUTION_HD1080, /**< 1920*1080, available framerates: 15, 30 fps.*/
-RESOLUTION_HD720,  /**< 1280*720, available framerates: 15, 30, 60 fps.*/
-RESOLUTION_VGA,    /**< 672*376, available framerates: 15, 30, 60, 100 fps.*/
-RESOLUTION_LAST
-</code>
-So with 15 FPS and VGA resolution, and a level 25 confidence (which is now and
-low) I'm still getting 120 mb/s throughput on the point cloud, which means i
-have to fucking voxel grid filter it. I did find a ''%%pcl_cuda%%'' based voxel grid
-filter though.
-Ok I think i just spent a half hour trying to get PCL with Cuda to build and I
-am gonna go right ahead and give the fuck up. This will be a timesink like no
-other. Gonna figure out what the nodelet manager is and attach a CPU based voxel
-filter to and end of story.
 ===== E-Con Tara =====
-[[primo_robot|Primo]] has a Tara - USB 3.0 Stereo Vision Camera
+An machine vision hardware company called E-Conn Systems makes an inexpensive and really great stereo camera called Tara. It is a USB 3.0 Stereo Vision Camera
 [[https://www.e-consystems.com/3D-USB-stereo-camera.asp#key-features|link]]. It
-was a great inexpensive stereo camera with excellent accurate PCL creation from
+has excellent accurate PCL creation fromranges of 0.5 meters to 3 meters. Below are outputs from two Tara cameras used on [[primo_robot|Primo]]. An custom self-filter and a voxel-grid-filter were used with the raw PCL output from the cameras.
-ranges of 0.3 meters to 3 meters.
+{{::stereo_rig.png?direct&600|}}
 ==== Specs ====
@@ Line 217: / Line 164: @@
 deviceID = 1
 </code>
+===== Custom Stereo Camera =====
+I built a stereo camera from scratch using USB3 monocular cameras, and by
+building a variable hardware triggering circuit to sequence frame grabs. This
+allowed me to get a camera with a C-Mount lens holder, giving me access to
+quality and inexpensive optics, and also allowed me to vary the baseline to
+focus accuracy on where I needed it. I created my own driver from a See3Cam fork
+that allowed me to sync the two cameras and also dynamically reconfigure exposure
+and sensor brightness levels.
+{{::img_6559.jpg?direct&600|}}
+In the above image, the two mono cameras separated on a rail, along with a triggering circuit in the middle comprise the custom stereo camera. Two Taras are housed above looking down on the left and right corners.
+{{::fwd_cam_expo.gif?direct&800|}}
+===== ZED Camera =====
+==== Specs ====
+**Name:**       ZED
+**Baseline:**   120mm
+The following is a sample output dump from the ZED camera. It had RGB coloring which was nice, and was accurate to near ranges, but for points further than 3 meters wavy inaccuracies formed.
+{{::zed_1.png?600|}}
+==== Setup ====
+I used the ZED in Primo running Ubuntu 16.04. I used the ROS driver package
+provided.
+  * OpenCV was installed already
+  * Installed cuda using the debian. Ran into an X server issue and ''%%glxgears%%'' didn't work. Restarted and it fixed everything
+  * Downloaded the ZED sdk run file. Made it executable and ran it.
+  * It installs in ''%%/usr/local/zed%%''
+  * Ran the ZED Explorer in the tools section of the install dir
+  * Updated the firmware from 922 to 1142. After I updated i did notice some frame glitching every now and then
+  * Went to the samples and built them using CMake
+==== ROS Install ====
+  * Installed the source from github as a package in the catkin_ws workspace.
+  * I changed the FPS in the launch file to 15.
+  * Also can play with the confidence level.
+==== General Notes ====
+Quality of the point cloud doesn't look great, and lots of "fall offs" on edges.
+With that said I didn't see any speckles and CPU usage is VERY low especially
+with the lower framerates. It is a very fast little device.
+==== Software ====
+Found the resolution definitions here:
+<code c++>
+RESOLUTION_HD2K,   /**< 2208*1242, available framerates: 15 fps.*/
+RESOLUTION_HD1080, /**< 1920*1080, available framerates: 15, 30 fps.*/
+RESOLUTION_HD720,  /**< 1280*720, available framerates: 15, 30, 60 fps.*/
+RESOLUTION_VGA,    /**< 672*376, available framerates: 15, 30, 60, 100 fps.*/
+RESOLUTION_LAST
+</code>
+So with 15 FPS and VGA resolution, and a level 25 confidence (which is now and
+low) I'm still getting 120 mb/s throughput on the point cloud, which means i
+have to fucking voxel grid filter it. I did find a ''%%pcl_cuda%%'' based voxel grid
+filter though.
+Ok I think i just spent a half hour trying to get PCL with Cuda to build and I
+am gonna go right ahead and give the fuck up. This will be a timesink like no
+other. Gonna figure out what the nodelet manager is and attach a CPU based voxel
+filter to and end of story.
+===== Filtering =====
+Voxel grid filtering from the PCL library was used to significantly lower the data throughput on the PCLs generated by the stereo cameras. On the left image below is a raw PCL dump and on the right is a voxel grid filtered one.
+{{::voxel_filter.png?600|}}