Exploring Udacity’s Open-Source Lidar Dataset

Knowing how to work with lidar sensors and the algorithms that utilize them can be a highly sought after skill in robotics, but is also hard to get started with. Lidar sensors (and depth mapping cameras) can be prohibitively expensive and a pain to set up, delaying and distracting developers from what they are actually interested in — coding something cool that makes use of the data. Compare that to computer vision, where all you have to do to get started is connect to a cheap webcam, or load up a video from the internet, and you are ready to get coding. For people studying computer vision, you can go from having nothing to making an face tracking program in a couple of hours or less.

There is nothing like that for lidar. No easy way to get some data and start working with it out of the box.

With the goal of making it simple and easy for developers to start working with lidar data, I went looking for datasets to work with and develop into some future guide. After a quick Twitter conversation with David Silver (head of the Udacity SDC program) and Oliver Cameron (CEO of Voyage), I decided to begin by investigating Udacity’s 3.5 hour driving dataset.

In this post I’ll be walking you through my exploration of the Udacity dataset and some of the troubleshooting needed to get it running. Ultimately, I was not able to get the point cloud data playing back with full accuracy, but I hope that with this work log and a bit of help from the community, we can get the data clean enough to start running algorithms and use as a teaching tool.


Setup

Dataset Download

Udacity’s dataset library includes a variety of sensor logs primarily used for their Self-Driving Car Challenge, which you can read all about from Oliver Cameron’s writeup.

For my purposes, there are two datasets of interest which each contain camera, GPS, lidar, and steering/throttle/brake data. The driving data is stored in ROS bags and downloaded via the torrent links found below. I recommend you start downloading these now; they are pretty large!

  1. The first dataset, CHX_001, is a small (1.5 GB) recording of a lap around the block at the Udacity office, which is great for getting started since it won’t take long to download.
  2. The second dataset, CH03_002, contains a large (60.1 GB) continuous recording of a trip on the El Camino Real highway from the Udacity office to San Francisco.

As noted, the lidar data for both of these logs come from a Velodyne HDL-32E.

After downloading, you can extract these files to anywhere. Personally, I put them in a new folder in my home directory: ~/udacity-driving-data/.

Software Installation

ROS. To view the data, you will need ROS Indigo (though Kinetic may work). I already had this installed, but if you are following along, you can find the ROS Indigo install guide on their website, and should follow-up that guide by creating your catkin workspace.

Playback Script. You will also need to download the files used for playing back the data in ROS. The ROS package containing these files are found in the datasets/udacity_launch/ folder in Udacity’s self-driving-car repository. To properly install this, use the following steps:

  1. Download Udacity’s self-driving-car repository and copy the udacity_launch folder to the src folder of your ROS catkin workspace. (In hindsight, you could also use a symbolic link in order to preserve git tracking.)
  2. Register udacity_launch with ROS via catkin_make.
  3. (Optional) If you don’t want to run setup.sh in every new terminal where you use a ROS command, you should also add setup.bash to your ~/.bashrc. Some ROS guides may have already had you do this.

Exploration

Viewing the Dashboard Video

Once you have one of the datasets downloaded, ROS installed, and udacity_launch set up in your catkin workspace, you can easily view the videos stored in each log with the following steps.

  1. Open the terminal and start ROS.
  2. In a second terminal, navigate to your dataset and start publishing it over ROS.
  3. In a third terminal, start playback via udacity_launch.
  4. In a fourth terminal, start rviz to visualize the data.

If everything works, rviz will open just as below, showing the videos from the car’s three front facing cameras. If you get any errors, make sure you have run source ~/catkin_ws/devel/setup.bash in every terminal by hand or via your ~/.bashrc.

Viewing the Lidar Data

Troubleshooting

Visualizing the lidar point cloud ended up being a bit more difficult than the videos, since Udacity does not provide much support for it. In fact, I still don’t have it working entirely!

If you look at udacity_launch/launch/bag_play.launch, you can see that there are two optional arguments that have to do with lidar, velodyne_packet2pointcloud and velodyne_packets2pointcloud, which run specific ROS nodes from the velodyne_decoder or velodyne_pointcloud ROS packages, respectively.

However, if you try to run bag_play.launch with either of these arguments true, it will probably not work because ROS does not come with the velodyne_decoder or velodyne_pointcloud packages. I couldn’t find any documentation about this from Udacity, so after a bit of digging around on the internet I was able to find the velodyne_pointcloud package, which is just part of the larger velodyne ROS package. I was not able to figure out what velodyne_decoder was, so I went ahead using bag_play.launch with only the velodyne_packet2pointcloud argument.

Hoping that everything would now work, I installed the velodyne ROS package, ran bag_play.launch with velodyne_packets2pointcloud:=True, and re-configured rviz to display the point cloud data. This didn’t work too well, so don’t install the velodyne package just yet. As you can see below, while the point cloud seems to generally be correct, it also curves upwards the further away from the car the points are, forming something like a giant funnel.

This seemed like a calibration problem. Looking at bag_play.launch again, I noticed the following line, where we could change the calibration file.

Aha! That’s not the right Velodyne device. The datasets said they were recorded on the Velodyne HDL-32E, the Velodyne VLP-16 is a different lidar sensor, so maybe we just need to change this to the correct one. But, Udacity does not provide any calibration file, and if you navigate to the params folder by running roscd velodyne_pointcloud/params/, you will see there is no HDL-32E calibration file. Further investigation online showed that the reason for this is that HDL-32E calibration file is not included in the current public release of the velodyne package. Instead, if you need that file you should build it from source. I’ll discuss that next.

The Best Results, so far

Fixing the calibration file. With some troubleshooting out of the way, we now know that in order to properly visualize the lidar point cloud, we need to install the velodyne ROS package from source and use the HDL-32E calibration file. You can do this by:

Then open ~/catkin_workspace/src/udacity_launch/launch/bag_play.launch and change the line

to

PS: There are actually two HDL-32E calibration files. Th file64e_utexas.yaml is by The University of Texas. Another one, called 64e_s2.1-sztaki.yaml, can be found in the same folder, but did not seem to work as well.

Changing rviz settings. When you run rviz.launch it opens rviz to show the camera video only, as defined by udacity_launch/rviz/disp.rviz, so even if the point cloud data was properly published you would not see it until you added the point cloud to the list of things to visualize as well as tune a few other settings. To save you some time, you can download this new copy of disp.rivz and replace the old one to easily visualize the point cloud data.

Displaying the Point Cloud. Now we are ready to visualize everything. We need to run all four terminals from earlier again, but this time with velodyne_packets2pointcloud:=True for bag_play.launch, as below.

Now with a proper calibration file, we the results below, which look a lot better. As you can see, the scan lines no longer curve upwards in the distance.

While things are better, they are not perfect. For example objects seem to warp as they pass by and some of the scan lines are floating out in the middle of space (particularly noticeable in the video above). Flat surfaces, such as the side of a car, start to bend and curve, or even diverge into different directions. This especially significant when objects are close to the car and happens a few times in the video below.


Conclusion

Seeing these results, it seems like further calibration or some form of processing is needed for the lidar.

For now, though, Udacity’s lidar data does not seem reliable or clean enough to confidently work with, especially for learning algorithms. Perhaps there is a missing configuration file, or additional pre-processing is needed, or perhaps there is something I am doing wrong, but since this is the beginning of my own exploration with lidar, I can’t be sure. If you have any thoughts or useful experience, please share! I don’t want to give up on this data set entirely and I’d still love to get it working.

Next, I plan to look into The KITTI Vision Benchmark Suite and getting it working in ROS. Since it is a more commonly used dataset, I hope it will be better suited for learning and teaching how to work with lidar. Stay posted!

Leave a Reply