This is a demonstration for low latency streaming of the Pi's camera module to any reasonably modern web browser, utilizing Dominic Szablewski's excellent JSMPEG project. Other dependencies are the Python ws4py library, my picamera library (specifically version 1.7 or above), and FFmpeg.
Firstly make sure you've got a functioning Pi camera module (test it with
raspistill to be certain). Then make sure you've got the following packages
installed:
$ sudo apt-get install ffmpeg git python3-picamera python3-ws4py
Next, clone this repository:
$ git clone https://github.com/waveform80/pistreaming.git
Run the Python server script which should print out a load of stuff to the console as it starts up:
$ cd pistreaming
$ python3 server.py
Initializing websockets server on port 8084
Initializing HTTP server on port 8082
Initializing camera
Initializing broadcast thread
Spawning background conversion process
Starting websockets thread
Starting HTTP server thread
Starting broadcast thread
Now fire up your favourite web-browser and visit the address
http://pi-address:8082/ - it should fairly quickly start displaying the feed
from the camera. You should be able to visit the URL from multiple browsers
simultaneously (although obviously you'll saturate the Pi's bandwidth sooner or
later).
If you find the video stutters or the latency is particularly bad (more than a second), please check you have a decent network connection between the Pi and the clients. I've found ethernet works perfectly (even with things like powerline boxes in between) but a poor wifi connection doesn't provide enough bandwidth, and dropped packets are not handled terribly well.
To shut down the server press Ctrl+C - you may find it'll take a while to shut down unless you close the client web browsers (Chrome in particular tends to keep connections open which will prevent the server from shutting down until the socket closes).
The server script is fairly simple but may look a bit daunting to Python newbies. There are several major components which are detailed in the following sections.
This is implemented in the StreamingHttpServer and StreamingHttpHandler
classes, and is quite simple:
- In response to an HTTP GET request for "/" it will redirect the client to "/index.html".
- In response to an HTTP GET request for "/index.html" it will serve up the contents of index.html, replacing @ADDRESS@ with the Pi's IP address and the websocket port.
- In response to an HTTP GET request for "/jsmpg.js" it will serve up the contents of jsmpg.js verbatim.
- In response to an HTTP GET request for anything else, it will return 404.
- In response to an HTTP HEAD request for any of the above, it will simply do the same as for GET but will omit the content.
- In response to any other HTTP method it will return an error.
This is implemented in the StreamingWebSocket class and is ridiculously
simple. In response to a new connection it will immediately send a header
consisting of the four characters "jsmp" and the width and height of the video
stream encoded as 16-bit unsigned integers in big-endian format. This header is
expected by the jsmpg implementation. Other than that, the websocket server
doesn't do much. The actual broadcasting of video data is handled by the
broadcast thread object below.
The BroadcastOutput class is an implementation of a picamera custom
output.
On initialization it starts a background FFmpeg process (avconv) which is
configured to expect raw video data in YUV420 format, and will encode it as
MPEG1. As unencoded video data is fed to the output via the write method, the
class feeds the data to the background FFmpeg process.
The BroadcastThread class implements a background thread which continually
reads encoded MPEG1 data from the background FFmpeg process started by the
BroadcastOutput class and broadcasts it to all connected websockets. In the
event that no websockets are currently connected the broadcast method simply
discards the data. In the event that no more data is available from the FFmpeg
process, the thread checks that the FFmpeg process hasn't finished (with
poll) and terminates if it has.
Finally, the main method may look long and complicated but it's mostly
boiler-plate code which constructs all the necessary objects, wraps several of
them in background threads (the HTTP server gets one, the main websockets
server gets another, etc.), configures the camera and starts it recording to
the BroadcastOutput object. After that it simply sits around calling
wait_recording until someone presses Ctrl+C, at which point it shuts
everything down in an orderly fashion and exits.
Since authoring the picamera library, a frequent (almost constant!) request has been "how can I stream video to a web page with little/no latency?" I finally had cause to look into this while implementing a security camera system using the Pi.
My initial findings were that streaming video over a network is pretty easy: open a network socket, shove video over it, done! Low latency isn't much of an issue either; you just need a player that's happy to use a small buffer (e.g. mplayer). Better still there's plenty of applications which will happily decode and play the H.264 encoded video streams which the Pi's camera produces ... unfortunately none of them are web browsers.
When it comes to streaming video to web browsers, the situation at the time of writing is pretty dire. There's a fair minority of browsers that don't support H.264 at all. Even those that do have rather variable support for streaming including weird not-really-standards like Apple's HLS (which usually involves lots of latency). Then there's the issue that the Pi's camera outputs raw H.264, and what most browsers want is a nice MPEG transport stream (TS). FFmpeg seemed like the answer to that, but the version that ships with Raspbian doesn't seem to like outputting valid PTS (Presentation Time Stamps) with the Pi's output. Perhaps later versions work better, but I was looking for a solution that wouldn't involve users having to jump through hoops to create a custom FFmpeg build (mostly because I could just imagine the amount of extra support questions I'd get from going that route)!
So, what about other formats? Transcoding to almost anything else (WebM, Ogg, etc.) is basically out of the question because the Pi's CPU just isn't fast enough, not to mention none of those really solve the "universal client" problem as there's plenty of browsers that don't support these formats either. MJPEG looked an intruiging (if thoroughly backward) possibility but I found it rather astonishing that we'd have to resort to something as primitive as that. Surely in this day and age we could at least manage a proper video format?!
Then, out of the blue, and by sheer coincidence a group in Canada got in contact to ask whether the Pi could produce raw (i.e. unencoded) video output. This wasn't something I'd ever been asked before but it turned out to be fairly simple, so I added it to the list of tickets for 1.7 and finished the code for it about a week later. I confess I pretty much skimmed the rest of their e-mail the first time I read it, but with the implementation done I went back and read it properly. They wanted to know whether they could use the picamera library with Dominic Szablewski's Javascript-based MPEG1 decoder.
This was an interesting idea! Javascript implementations are near universal in browsers these days, and Dominic's decoder was fast enough that it would run happily even on relatively small platforms (for example it runs on iPhones and reasonably modern Androids like a Nexus). Furthermore, the Pi is just about fast enough to handle MPEG1 transcoding with FFmpeg (at least at low resolutions).
Okay, it's not a modern codec like the excellent H.264. It's not using "proper" HTML5 video tags. All round, it's still basically a hack, and yes it's pretty appalling that we have to resort to hacks like this just to come up with a universally accessible video streaming solution. But hey ... it works, and it's not (quite) as primitive as MJPEG so I'm happy to declare victory. I spent an evening bashing together a Python version of the server side. It turned out a bit too complex to include as a recipe in the docs, hence why it's here, but I think it provides a reasonable basis for others to work from and extend.
Enjoy!
Dave.