Jan Schmidt

This is another in my series of updates on developing positional tracking for the Oculus Rift CV1 in OpenHMD

In the last post I ended with a TODO list. Since then I’ve crossed off a few things from that, and fixed a handful of very important bugs that were messing things up. I took last week off work, which gave me some extra hacking hours and enthusiasm too, and really helped push things forward.

Here’s the updated list:

• The full model search for re-acquiring lock when we start, or when we lose tracking takes a long time. More work will mean avoiding that expensive path as much as possible.
• Multiple cameras interfere with each other.
  • Capturing frames from all cameras and analysing them happens on a single thread, and any delay in processing causes USB packets to be missed.
  • I plan to split this into 1 thread per camera doing capture and analysis of the ‘quick’ case with good tracking lock, and a 2nd thread that does the more expensive analysis when it’s needed. Partially Fixed
• At the moment the full model search also happens on the video capture thread, stalling all video input for hundreds of milliseconds – by which time any fast motion means the devices are no longer where we expect them to be.
  • This means that by the next frame, it has often lost tracking again, requiring a new full search… making it late for the next frame, etc.
  • The latency of position observations after a full model search is not accounted for at all in the current fusion algorithm, leading to incorrect reporting. Partially Fixed
• More validation is needed on the camera pose transformations. For the controllers, the results are definitely wrong – I suspect because the controller LED models are supplied (in the firmware) in a different orientation to the HMD and I used the HMD as the primary test. Much Improved
• Need to take the position and orientation of the IMU within each device into account. This information is in the firmware information but ignored right now. Fixed
• Filtering! This is a big ticket item. The quality of the tracking depends on many pieces – how well the pose of devices is extracted from the computer vision and how quickly, and then very much on how well the information from the device IMU is combined with those observations. I have read so many papers on this topic, and started work on a complex Kalman filter for it.
• Improve the model to LED matching. I’ve done quite a bit of work on refining the model matching algorithm, and it works very well for the HMD. It struggles more with the controllers, where there are fewer LEDs and the 2 controllers are harder to disambiguate. I have some things to try out for improving that – using the IMU orientation information to disambiguate controllers, and using better models for what size/brightness we expect an LED to be for a given pose.
• Initial calibration / setup. Rather than assuming the position of the headset when it is first sighted, I’d like to have a room calibration step and a calibration file that remembers the position of the cameras.
• Detecting when cameras have been moved. When cameras observe the same device simultaneously (or nearly so), it should be possible to detect if cameras are giving inconsistent information and do some correction.
• hot-plug detection of cameras and re-starting them when they go offline or encounter spurious USB protocol errors. The latter happens often enough to be annoying during testing.
• Other things I can’t think of right now.

As you can see, a few of the top-level items have been fixed, or mostly so. I split the computer vision for the tracking into several threads:

• 1 thread shared between all sensors to capture USB packets and assemble them into frames
• 1 thread per sensor to analyse the frame and update poses

The goal with that initial split was to prevent the processing of multiple sensors from interfering with each other, but I found that it also has a strong benefit even with a single sensor. I realised something in the last week that I probably should have noted earlier: The Rift sensors capture a video frame every 19.2ms, but that frame then takes a full 17ms to deliver across the USB – the means that when everything was in one thread, even with 1 sensor there was only about 2.2ms for the full analysis to take place or else we’d miss a packet of the next frame and have to throw it away. With the analysis now happening in a separate thread and a ping-pong double buffer in place, the analysis can take quite a bit longer without losing any video frames.

I plan to add a 2nd per-sensor thread that will divide the analysis further. The current thread will do only fast pass validation of any existing tracking lock, and will defer any longer term analysis to the other thread. That means that if we have a good lock on the HMD, but can’t (for example) find one of the controllers, searching for the controller will be deferred and the fast pass thread will move onto the next frame and keep tracking lock on the headset.

I fixed some bugs in the calculations that move between frames of reference – converting to/from the global position and orientation in the world to the position and orientation relative to each camera sensor when predicting what the appearance of the LEDs should be. I also added in the IMU offset and orientation of the LED models from the firmware, to make the predictions more accurate when devices move in the time between camera exposures.

Yaw Correction: when a device is observed by a sensor, the orientation is incorporated into what the IMU is measuring. The IMU can sense gravity and knows which way is up or down, but not which way is forward. The observation from the camera now corrects for that yaw drift, to keep things pointing the way you expect them to.

Some other bits:

• Fixing numerical overflow issues in the OpenHMD maths routines
• Capturing the IMU orientation and prediction that most closely corresponds to the moment each camera image is recorded, instead of when the camera image finishes transferring to the PC (which is 17ms later)
• Improving the annotated debug view, to help understand what’s happening in the tracking computer vision steps
• A 1st order estimate of device velocity to help improve the next predicted position

I posted a longer form video walkthrough of the current code in action, and discussing some of the remaining shortcomings.

As previously, the code is available at https://github.com/thaytan/OpenHMD/tree/rift-correspondence-search

In my last post here, I gave some background on how Oculus Rift devices work, and promised to talk more about Rift S internals. I’ll do that another day – today I want to provide an update on implementing positional tracking for the Rift CV1.

I was working on CV1 support quite a lot earlier in the year, and then I took a detour to get basic Rift S support in place. Now that the Rift S works as a 3DOF device, I’ve gone back to plugging away at getting full positional support on the older CV1 headset.

So, what have I been doing on that front? Back in March, I posted this video of a new tracking algorithm I was working on to match LED constellations to object models to get the pose of the headset and controllers:

The core of this matching is a brute-force search that (somewhat cleverly) takes permutations of observed LEDs in the video footage and tests them against permutations of LEDs from the device models. It then uses an implementation of the Lambda Twist P3P algorithm (courtesy of pH5) to compute the possible poses for each combination of LEDs. Next, it projects the points of the candidate pose to count how many other LED blobs will match LEDs in the 3D model and how closely. Finally, it computes a fitness metric and tracks the ‘best’ pose match for each tracked object.

For the video above, I had the algorithm implemented as a GStreamer plugin that ran offline to process a recording of the device movements. I’ve now merged it back into OpenHMD, so it runs against the live camera data. When it runs in OpenHMD, it also has access to the IMU motion stream, which lets it predict motion between frames – which can help with retaining the tracking lock when the devices move around.

This weekend, I used that code to close the loop between the camera and the IMU. It’s a little simple for now, but is starting to work. What’s in place at the moment is:

• At startup time, the devices track their movement only as 3DOF devices with default orientation and position.
• When a camera view gets the first “good” tracking lock on the HMD, it calls that the zero position for the headset, and uses it to compute the pose of the camera relative to the playing space.
• Camera observations of the position and orientation are now fed back into the IMU fusion to update the position and correct for yaw drift on the IMU (vertical orientation is still taken directly from the IMU detection of gravity)
• Between camera frames, the IMU fusion interpolates the orientation and position.
• When a new camera frame arrives, the current interpolated pose is transformed back into the camera’s reference frame and used to test if we still have a visual lock on the device’s LEDs, and to label any newly appearing LEDs if they match the tracked pose
• The device’s pose is refined using all visible LEDs and fed back to the IMU fusion.

With this simple loop in place, OpenHMD can now track multiple devices, and can do it using multiple cameras – somewhat. The first time the tracking block associated to a camera thinks it has a good lock on the HMD, it uses that to compute the pose of that camera. As long as the lock is genuinely good at that point, and the pose the IMU fusion is tracking is good – then the relative pose between all the cameras is consistent and the tracking is OK. However, it’s easy for that to go wrong and end up with an inconsistency between different camera views that leads to jittery or jumpy tracking….

In the best case, it looks like this:

Which I am pretty happy with 🙂

In that test, I was using a single tracking camera, and had the controller sitting on desk where the camera couldn’t see it, which is why it was just floating there. Despite the fact that SteamVR draws it with a Vive controller model, the various controller buttons and triggers work, but there’s still something weird going on with the controller tracking.

What next? I have a list of known problems and TODO items to tackle:

• The full model search for re-acquiring lock when we start, or when we lose tracking takes a long time. More work will mean avoiding that expensive path as much as possible.
• Multiple cameras interfere with each other.
  • Capturing frames from all cameras and analysing them happens on a single thread, and any delay in processing causes USB packets to be missed.
  • I plan to split this into 1 thread per camera doing capture and analysis of the ‘quick’ case with good tracking lock, and a 2nd thread that does the more expensive analysis when it’s needed.
• At the moment the full model search also happens on the video capture thread, stalling all video input for hundreds of milliseconds – by which time any fast motion means the devices are no longer where we expect them to be.
  • This means that by the next frame, it has often lost tracking again, requiring a new full search… making it late for the next frame, etc.
  • The latency of position observations after a full model search is not accounted for at all in the current fusion algorithm, leading to incorrect reporting.
• More validation is needed on the camera pose transformations. For the controllers, the results are definitely wrong – I suspect because the controller LED models are supplied (in the firmware) in a different orientation to the HMD and I used the HMD as the primary test.
• Need to take the position and orientation of the IMU within each device into account. This information is in the firmware information but ignored right now.
• Filtering! This is a big ticket item. The quality of the tracking depends on many pieces – how well the pose of devices is extracted from the computer vision and how quickly, and then very much on how well the information from the device IMU is combined with those observations. I have read so many papers on this topic, and started work on a complex Kalman filter for it.
• Improve the model to LED matching. I’ve done quite a bit of work on refining the model matching algorithm, and it works very well for the HMD. It struggles more with the controllers, where there are fewer LEDs and the 2 controllers are harder to disambiguate. I have some things to try out for improving that – using the IMU orientation information to disambiguate controllers, and using better models for what size/brightness we expect an LED to be for a given pose.
• Initial calibration / setup. Rather than assuming the position of the headset when it is first sighted, I’d like to have a room calibration step and a calibration file that remembers the position of the cameras.
• Detecting when cameras have been moved. When cameras observe the same device simultaneously (or nearly so), it should be possible to detect if cameras are giving inconsistent information and do some correction.
• hot-plug detection of cameras and re-starting them when they go offline or encounter spurious USB protocol errors. The latter happens often enough to be annoying during testing.
• Other things I can’t think of right now.

A nice side effect of all this work is that it can all feed in later to Rift S support. The Rift S uses inside-out tracking to determine the headset’s position in the world – but the filtering to combine those observations with the IMU data will be largely the same, and once you know where the headset is, finding and tracking the controller LED constellations still looks a lot like the CV1’s system.

If you want to try it out, or take a look at the code – it’s up on Github. I’m working in the rift-correspondence-search branch of my OpenHMD repository at https://github.com/thaytan/OpenHMD/tree/rift-correspondence-search

For some time now, I’ve been involved in the OpenHMD project, working on building an open driver for the Oculus Rift CV1, and more recently the newer Rift S VR headsets.

This post is a bit of an overview of how the 2 devices work from a high level for people who might have used them or seen them, but not know much about the implementation. I also want to talk about OpenHMD and how it fits into the evolving Linux VR/AR API stack.

OpenHMD

http://www.openhmd.net/

In short, OpenHMD is a project providing open drivers for various VR headsets through a single simple API. I don’t know of any other project that provides support for as many different headsets as OpenHMD, so it’s the logical place to contribute for largest effect.

OpenHMD is supported as a backend in Monado, and in SteamVR via the SteamVR-OpenHMD plugin. Working drivers in OpenHMD opens up a range of VR games – as well as non-gaming applications like Blender. I think it’s important that Linux and friends not get left behind – in what is basically a Windows-only activity right now.

One downside is that does come with the usual disadvantages of an abstraction API, in that it doesn’t fully expose the varied capabilities of each device, but instead the common denominator. I hope we can fix that in time by extending the OpenHMD API, without losing its simplicity.

Oculus Rift S

I bought an Oculus Rift S in April, to supplement my original consumer Oculus Rift (the CV1) from 2017. At that point, the only way to use it was in Windows via the official Oculus driver as there was no open source driver yet. Since then, I’ve largely reverse engineered the USB protocol for it, and have implemented a basic driver that’s upstream in OpenHMD now.

I find the Rift S a somewhat interesting device. It’s not entirely an upgrade over the older CV1. The build quality, and some of the specifications are actually worse than the original device – but one area that it is a clear improvement is in the tracking system.

CV1 Tracking

The Rift CV1 uses what is called an outside-in tracking system, which has 2 major components. The first is input from Inertial Measurement Units (IMU) on each device – the headset and the 2 hand controllers. The 2nd component is infrared cameras (Rift Sensors) that you space around the room and then run a calibration procedure that lets the driver software calculate their positions relative to the play area.

IMUs provide readings of linear acceleration and angular velocity, which can be used to determine the orientation of a device, but don’t provide absolute position information. You can derive relative motion from a starting point using an IMU, but only over a short time frame as the integration of the readings is quite noisy.

This is where the Rift Sensors get involved. The cameras observe constellations of infrared LEDs on the headset and hand controllers, and use those in concert with the IMU readings to position the devices within the playing space – so that as you move, the virtual world accurately reflects your movements. The cameras and LEDs synchronise to a radio pulse from the headset, and the camera exposure time is kept very short. That means the picture from the camera is completely black, except for very bright IR sources. Hopefully that means only the LEDs are visible, although light bulbs and open windows can inject noise and make the tracking harder.

If you have both IMU and camera data, you can build what we call a 6 Degree of Freedom (6DOF) driver. With only IMUs, a driver is limited to providing 3 DOF – allowing you to stand in one place and look around, but not to move.

OpenHMD provides a 3DOF driver for the CV1 at this point, with experimental 6DOF work in a branch in my fork. Getting to a working 6DOF driver is a real challenge. The official drivers from Oculus still receive regular updates to tweak the tracking algorithms.

I have given several presentations about the progress on implementing positional tracking for the CV1. Most recently at Linux.conf.au 2020 in January. There’s a recording at https://www.youtube.com/watch?v=PTHE-cdWN_s if you’re interested, and I plan to talk more about that in a future post.

Rift S Tracking

The Rift S uses Inside Out tracking, which inverts the tracking process by putting the cameras on the headset instead of around the room. With the cameras in fixed positions on the headset, the cameras and their view of the world moves as the user’s head moves. For the Rift S, there are 5 individual cameras pointing outward in different directions to provide (overall) a very wide-angle view of the surroundings.

The role of the tracking algorithm in the driver in this scenario is to use the cameras to look for visual landmarks in the play area, and to combine that information with the IMU readings to find the position of the headset. This is called Visual Inertial Odometry.

There is then a 2nd part to the tracking – finding the position of the hand controllers. This part works the same as on the CV1 – looking for constellations of LED lights on the controllers and matching what you see to a model of the controllers.

This is where I think the tracking gets particularly interesting. The requirements for finding where the headset is in the room, and the goal of finding the controllers require 2 different types of camera view!

To find the landmarks in the room, the vision algorithm needs to be able to see everything clearly and you want a balanced exposure from the cameras. To identify the controllers, you want a very fast exposure synchronised with the bright flashes from the hand controller LEDs – the same as when doing CV1 tracking.

The Rift S satisfies both requirements by capturing alternating video frames with fast and normal exposures. Each time, it captures the 5 cameras simultaneously and stitches them together into 1 video frame to deliver over USB to the host computer. The driver then needs to split each frame according to whether it is a normal or fast exposure and dispatch it to the appropriate part of the tracking algorithm.

There are a bunch of interesting things to notice in these camera captures:

• Each camera view is inserted into the frame in some native orientation, and requires external information to make use of the information in them
• The cameras have a lot of fisheye distortion that will need correcting.
• In the fast exposure frame, the light bulbs on my ceiling are hard to tell apart from the hand controller LEDs – another challenge for the computer vision algorithm.
• The cameras are Infrared only, which is why the Rift S passthrough view (if you’ve ever seen it) is in grey-scale.
• The top 16-pixels of each frame contain some binary data to help with frame identification. I don’t know how to interpret the contents of that data yet.

Status

This blog post is already too long, so I’ll stop here. In part 2, I’ll talk more about deciphering the Rift S protocol.

Thanks for reading! If you have any questions, hit me up at mailto:thaytan@noraisin.net or @thaytan on Twitter

I’ve pushed some new changes to my Raspberry Pi camera GStreamer wrapper, at https://github.com/thaytan/gst-rpicamsrc/

These bring the GStreamer element up to date with new features added to raspivid since I first started the project, such as adding text annotations to the video, support for the 2nd camera on the compute module, intra-refresh and others.

Where possible, you can now dynamically update any of the properties – where the firmware supports it. So you can implement digital zoom by adjusting the region-of-interest (roi) properties on the fly, or update the annotation or change video effects and colour balance, for example.

The timestamps produced are now based on the internal STC of the Raspberry Pi, so the audio video sync is tighter. Although it was never terrible, it’s now more correct and slightly less jittery.

The one major feature I haven’t enabled as yet is stereoscopic handling. Stereoscopic capture requires 2 cameras attached to a Raspberry Pi Compute Module, so at the moment I have no way to test it works.

I’m also working on GStreamer stereoscopic handling in general (which is coming along). I look forward to releasing some of that code soon.

 

Way back in 2006, Andy Wingo wrote some small scripts for GStreamer 0.10 to demonstrate what was (back then) a fairly new feature in GStreamer – the ability to share a clock across the network and use it to synchronise playback of content across different machines.

Since GStreamer 1.x has been out for over 2 years, and we get a lot of questions about how to use the network clock functionality, it’s a good time for an update. I’ve ported the simple examples for API changes and to use the gobject-introspection based Python bindings and put them up on my server.

To give it a try, fetch play-master.py and play-slave.py onto 2 or more computers with GStreamer 1 installed. You need a media file accessible via some URI to all machines, so they have something to play.

Then, on one machine run play-master.py, passing a URI for it to play and a port to publish the clock on:

./play-master.py http://server/path/to/file 8554

The script will print out a command line like so:

Start slave as: python ./play-slave.py http://server/path/to/file [IP] 8554 1071152650838999

On another machine(s), run the printed command, substituting the IP address of the machine running the master script.

After a moment or two, the slaved machine should start playing the file in synch with the master:

If they’re not in sync, check that you have the port you chose open for UDP traffic so the clock synchronisation packets can be transferred.

This basic technique is the core of my Aurena home media player system, which builds on top of the network clock mechanism to provide file serving and a simple shuffle playlist.

For anyone still interested in GStreamer 0.10 – Andy’s old scripts can be found on his server: play-master.py and play-slave.py