Part 2: How to build a High-Definition FPV UAV using a Rasperry PI with HD camera, using a high speed WiFi link

Part Two: Here is the original picture of the finished product:

This is the second part of a 2-part series on 'How to build a High-Definition FPV UAV using a Raspberry PI with HD camera, using a high speed WiFi link.

In my first post on the subject (located here), I discussed the parts I used, and how to install them into a Hobby King Go-Discover FPV model. 

In this post, I will discuss installing the Raspberry PI and the PI camera in the Go-Discover gimbals, and the software configuration for both the Raspberry PI and the ground station PC.

From the previous post, step 3 was completed by installing the Ubiquity Rocket M5 in the model.  Now onto step 4:

Step 4: Install the Raspberry PI and PI Camera

Here is a photo of the position of the PI in the Go-Discover model:

The PI fits nicely just behind the camera gimbals, with the USB and HDMI ports on top. In the right side you can see the Cat5 network cable attached. This cable connects to the ethernet switch, which is also connected to the Rocket M5 input port.  

The two cables shown on top are the servo control wires for the gimbals, which I have directly connected to channel 4 and 5 on my radio.  I am using channel 4 (normally the rudder stick on my radio. Since there is no rudder on a flying wing, this is a convenient channel to use to move left and right with the camera. I have not (yet) moved to a head tracker, but if you already have that setup, just assign the channels accordingly.

To install the PI camera, remove the stock plate from the gimbals (for a GoPro), and mount the PI camera as shown in this photo:

The PI camera case fits very nicely into the slot, and again I used a small piece of velcro to hold it down. You could use a couple of small screws instead if you want a more secure hold.  The two gimbals servos are also shown here. They are simple to install, just follow the Go-Discover instructions.

Here is a front view of the PI camera installed:

Here is the block diagram describing all the connections:

Some comments on my previous post suggested that it is possible to eliminate the ethernet switch and serial-to-ethernet converter using the Raspberry PI and a serial port on the PI. I believe this post describes how to talk to the PI via the NavLink, but in this case, I want to use the PI to bridge the connection from the ground station to the APM/PixHawk. Somebody please comment on this if you know more about it.   I believe it would require a TCP/IP to serial link from the PI to the telemetry port on the APM, and some software on the PI to act as the bridge.  The main connection to the ground station is via the Rocket M5 and TCP/IP, not through a telemetry link (900 Mhz or Zigbee like I used on my other models).

Step 5: Getting it all to work with software configuration (the really fun part starts now).

Check out this post on what others have done with streaming and the PI.  My experiments showed that using GStreamer on both the PI and on Windows gives really good results with very low latency, if you use the right parameters. 

Get GStreamer on the PI by following this blog.   This is the same version of GStreamer that I am using on my setup. 

Make sure your PI camera works ok by plugging in the PI to a standard monitor using the HDMI port and follow the instructions on the Raspberry PI website on how to get the camera up and running (without GStreamer).  Once you have a working PI and camera, you can then proceed to stream things over the network.  

Note: It is suggested that you first get the PI streaming video by plugging it directly into your local network where you can also connect your ground station PC with the correct IP addresses (without the Rocket M5).   For my PI, I picked,  and for the ground station,    Make sure you can ping the PI from your PC and the PC from the PI.  

For streaming, you will also have to make sure all the ports you intent to use are open on the firewall (described later).

For the ground station PC,  you can download GStreamer here.  Make sure when you install, select to install everything , or full installation (not the default). 

Here is the command I use for the PI to pipe the camera output to GStreamer:

raspivid -t 0 -w 1280 -h 720 -fps 30 -b 1700000 -o - | gst-launch1.0 -v fdsrc ! h264parse config-interval=1 ! rtph264pay ! udpsink host = port= 9000

The command is explained as follows:

raspivid is the command to start the camera capture on the PI.  The -w switch is for the width in pixels, and the -h switch is for the height.  In this case, I am using 1280 X 720, but you can try any combination that fits your needs. 

The -b switch is the bit rate for the sampling. In this case I chose 1.7mbs to send over the stream. Again you can experiment with higher or lower values. This settings seems to work good for me, and the latency is almost unnoticeable.  

the "-o - |" is piping the output to gstreamer.  Make sure you include the dash before the pipe "|" symbol. 

For the GStreamer command, all the filters are separated with an exclamation point "!", as these are individual drivers that are part of GStreamer.  Since the PI has hardware accelerated video, the output is in a format called "H264", which is a highly-compressed stream. The GStreamer filters are configured to transport the output via a UDP socket connection to the target PC. Notice the 'udpsink' element which specifies the host - in this case your ground station, and the UDP port.  I am using port 9000, but you can use any open port on your system, but be sure to open the firewall or it won't work!  You can also use TCP instead of UDP, but for such a data stream, I chose to use UDP since dropouts are certainly possible, and with UDP this is ok, but with TCP, you could have socket problems and higher latency. 

Note: to get the PI to execute this command on boot, make a shell script with the above command and add it to your local.rc boot sequence. That way when the PI boots, you get the stream without having to log into the PI remotely. 

For the ground station PC, once you have installed GStreamer and opened the correct ports, use this command (from the command prompt) to view the stream:

c:\gstreamer\1.0\x86_64\bin\gst-launch-1.0 udpsrc port=9000 ! application/x-rtp,encoding-name=H264,payload=96 ! rtph264depay ! avdec_h264 ! videoconvert ! autovideosink

If all goes well, you should see the PI camera output on your PC screen in a popup window.  For those of you what want to use FPV goggles, you can connect to the HDMI port on your PC to display the output if your goggles support HDMI. 

I have this command in a batch file (with a PAUSE) statement at the end to keep the window open.

WHEW!  If you got this far, you are amazing. 

The last step to complete the build is to connect to the APM from mission planner.  The method I used to connect was to install a utility that converts a TCP connection to a virtual serial port, but I also think that directly connecting the mission planner to the TCP port will also work, however I have not tried it. I will post back later after trying it.

Here is the link to setup the serial to ethernet device to have an IP address and port.

Here is the link to the configuration utility for installing the virtual serial port.   

Once you have a serial connection over TCP/IP working to the APM, you should be able to connect with Mission Planner. On the maiden flight, it worked perfectly, and I didn't see a single drop in the telemetry data or anything noticeable in the video transmission, however my first flight was limited to 2km.

The last step is to connect the Rocket M5 to the Nano M5 and test everything using the OTA (over the air) connection. If all is well, you are ready to fly!  But be careful on your maiden, you just spent $700. 

Finally, here is a photo of my Antenna Tracker with the Nano M5 attached. My next update will include a video of a longer flight.  

Happy Flying!

Views: 95715

Comment by John Arne Birkeland on October 1, 2014 at 6:46am

Here are some suggestions that might improve GStreamer latency and robustness in your receiver pipeline. I have also included some Windows specific improvements, since I see that is the platform you are using.


udpsrc port=9000 ! application/x-rtp,encoding-name=H264,payload=96 ! rtph264depay ! avdec_h264 ! videoconvert ! autovideosink


udpsrc port=9000 do-timestamps=false buffer-size=60000 ! application/x-rtp,encoding-name=H264,payload=96 ! rtph264depay ! h264parse ! queue ! avdec_h264 ! video/x-raw, format=I420 ! d3dvideosink sync=false

Reasons for changes.


Since your primary goal is low latency video, there is no need to use timestamps in the pipeline. Instead you want it to process as quickly as possible.


I have found that working with small data chunks sometimes degrade performance and make the h264 processing more likely to break apart. Waiting until you have received a reasonable amount of data helps with this, and does not affect latency as long as the data buffer size is kept withing reason. 60K is a guesstimate where I simply did bit-rate / frame-rate and rounded up.


I added a h264parse to help ensure proper h264 for the decoder. Together with buffer-size this helps making the stream more robust against errors.


It is always a good idea to add a queue before the h264 decoder to prevent the pipeline from stalling while the decoder is busy. This will not increase latency since the decoder will be much faster then the actual video feed on any modern computer.

video/x-raw, format=I420 ! d3dvideosink

This is a Windows performance optimization. Using the I420 (h264 decoder native) color-space format ensures the quickest decoding possible, and also enables direct-rendering and hardware accelerated (GPU) color-space conversion in the D3D video sink.


Since low latency is the primary goal, dump the image out as quickly as possible without trying to time them.

Comment by Tommy Larsen on October 1, 2014 at 7:00am

Very nice info, John Arne :)

Comment by John Arne Birkeland on October 1, 2014 at 7:22am

Correction. I forgot to convert from bit to byte when calculating the buffer-size. The 60K buffer-size should be 7-8K instead.

Comment by Patrick Duffy on October 1, 2014 at 7:50am

Hello John,  I'll certainly give that pipeline a try. I will also try to post the actual latency using the utility Tommy posted comparing different GStreamer configurations.

Comment by MD on October 1, 2014 at 8:55am

@Patrick - SpiroNet 5.8 GHz RHCP Patch Antenna, SpiroNet 5.8 GHz LHCP Patch Antenna, FPVLR Pinwheel LHCP and RHCP.

Spoke with Alex, Mashton, Circular Wireless and a few others but all involved waiting for various reasons so will try the above first.

Comment by Nils Högberg on October 1, 2014 at 1:27pm

Really interesting project. I'll be following this closely.

Been looking for a solution to stream HD video for my own project where I plan to stream (preferably stereo) a HD stream to my Oculus Rift using a intel NUC and two webcams. This might also improve latencys with head tracking over mavlink.

Comment by Patrick Duffy on October 1, 2014 at 6:06pm

Here's a shot showing the latency using Tommy's tool. It's about 140ms for the 1280x720 pipeline. So my original claim of ~100ms was not that far off.

Comment by Patrick Duffy on October 1, 2014 at 6:08pm

Comment by MD on October 1, 2014 at 6:13pm


Comment by Artem on October 1, 2014 at 6:16pm

this is actually decent, I was worrying it would be in the 300ms+ range. with 140-150ms delay one can fly rather safely at altitude, of course not for proximity flying, but is decent for flying slower than 40mph , however people routinely fly 60-70mph, and for this a regular analog camera would be the best.  It would be really good if we could bring this in the 80ms range. 


You need to be a member of DIY Drones to add comments!

Join DIY Drones

© 2019   Created by Chris Anderson.   Powered by

Badges  |  Report an Issue  |  Terms of Service