A Pi Straight Flush – Comparing the Pi 2

Extract

The Pi 2 Model B was release on the 2nd February 2015 promising a 1.5 to 6 times performance boost over the earlier generation one hardware.  If that boost occurred across the entire pipeline then even a minimal 1.5 improvement would reduce the latency from 120 to 80 milliseconds (a 33% reduction) and make a very useable FPV system.  Not enough for racing, but enough for most everything else.  Anyway, a 250mm racing quad would be too small to carry a Raspberry Pi.

In reality the processor is only one small part of the video streaming pipeline and while the Pi 2 does reduce the latency, the improvement is not nearly as much as you might think.  In fact most of the improvement over previous tests has come from a change in test procedure.  The final conclusion though is that the Pi 2 does at least allow us to break the 100ms barrier.  If I can just get a 5 GHz wifi system running on the Pi 2, then the next step is to actually get it airborne.

In addition to the Pi 2, I also have a Pi 1 A+ that had not been tested.  The A+ is smaller and lighter than the previous boards which is an advantage if you’re trying to fit it in a tight fuselage.  As I was using a new build of Raspbian, I thought I would run the test on all versions of the Pi: the A, B, B+, A+ and Pi 2B; a straight flush of Pi.

PiStraightFlush

Setup

Nothing new here.  I used the build of Raspbian from the previous test.  As I haven’t managed to get 5 GHz WiFi running on the Pi 2 yet, the test was done using the Edimax EW-7612UAn V2 2.4GHz USB WiFi stick as this was compatible with all versions of the Pi.  Following the previous tests where the Dual Band AC600 Wifi Adapter had similar performance to the on board ethernet port, I also tested the B and 2B over LAN.  The idea being to see what the Pi 2B could achieve with a good WiFi adapter.

Procedure

These tests were a straight comparison between the five Pi devices using the same settings and the same testing methodology as previously.  The resolution was 1280 x 720 pixels with a 6Mbps bit rate.

Results

My first impression of the Pi 2B was that it was faster than the series one models, however when I started analysing the results, I could see no improvement after seven iterations.  So I went back and calculated the latency at each step.  This revealed a general trend for the latency to increase with each step.

With the Pi sampling at 48 fps, that’s 21ms for each frame.  The screen is refreshing at 60 fps or once every 17ms.  Depending on when the image appears on the screen bright enough for the camera to detect there could be a significant increase in latency towards the end of the iteration loop.  In practice these things tend to average out and the general trend seems to be a 20ms increase between iterations 1 and 7.

I started using the multiple iteration technique right back at the start of this testing when we were looking at a latency in the 200 to 300ms range. Back then there was also a lot of variation at each step and so this increasing trend was not immediately apparent.  Now that the latency is around 100ms and more consistent it has become so.

Straight Flush Iterations

Because of this, I have changed my result reporting to only use the average of the first iterations and I always collect at least 10.  From the above graph you can see that over WiFi the Pi 2B has the lowest latency at 103.3 ms, but that’s only  2.3 ms faster than the Pi 1B.

Using the ethernet port the Pi 2B does better with an average latency of 97.2 ms whereas the Pi 1B could only manage 107.4 ms.

Straight Flush Iterations Eth

Here is a summary of the final results.

Straight Flush Latency Summary

From fastest to slowest the final results are Pi2B, Pi1B, Pi1A+, Pi1A, Pi1B+.

Discussion

So where is the 1.5 to 6 times performance improvement, the 33% reduction?  The best result achieved is a 10% reduction.  The answer is that the processor represents only a part of the complete video capture and transmission pipeline and each stage involves converting the stream into different forms through multiple processors.  Remember each device (camera, bus controller, Pi, USB controller, USB hub, WiFi Adapters, PC CPU, Video Card and display) has their own processor and most are not as powerful as the Pi.  Unfortunately, there is no easy way to measure the latency at each stage to determine where the main bottlenecks are.

Video Pipeline

It must also be remembered that gstreamer is a complete pipeline in itself, with multiple conversions between the raw video and the packetized stream sent to the WiFi adapter.

It is interesting to note that the A models were faster than the Pi1B+ even though they have half its memory.  The A models only have a single USB interface.  The B models don’t have multiple USB interfaces, they have the same single USB interface and that feeds a built-in two or  four port hub.  This adds an extra stage to the transmission pipeline making the Pi1B+ slower than the A models even with twice the memory.   I assume the two port hub in the Pi1B is faster than the four port hub in the Pi1B+ which combined with the extra memory makes it faster than the A models.

Conclusion

If the Pi 2B can be made to work with one of the dual band WiFi adapters and achieve this sub 100ms average latency then it has promise to finally deliver a low cost and useable high definition FPV system.

At the moment Raspbian is still optimised for the arm6 architecture of the first generation hardware.  With an operating system optimised for the arm7 in the Pi 2B further latency improvements should be possible.

There are no immediate plans for a Pi 2A.  If this was released with the 512MB of memory of the current B models, a single USB interface and a small form factor, it could be the idea platform for HD FPV.

Perhaps the way to improve the latency further is to switch from a raspivid/gstreamer RTSP (Real Time Streaming Protocol) solution to some other protocol.  Perhaps MpegStreamer holds the answer.  Only more testing will tell.

Advertisements

Raspivid v Gst-rpicamsrc (Updated)

Introduction

User bocorps pointed me in the direction of gst-rpicamsrc.  This is “… a GStreamer wrapper around the raspivid/raspistill functionality of the RaspberryPi, providing a GStreamer source element capturing from the Rpi camera.”

What this means is that instead of piping the output of raspivid into gstreamer, gstreamer has a source element to read the camera directly.  This is similar to using the video4linux (v4l) source element, but negates the need for a v4l driver.

My hope was that by integrating the camera functionality into a gstreamer source element the latency would be reduced.  Unfortunately, I actually saw an 18% increase in latency.

Installation

Before I could use the gst-rpicamsrc element, I needed to download the source and build it.  As I was working with a minimal install of Raspbian Jessie, I needed to install the git package before I could do anything else

sudo apt-get install git

With git installed I could download the latest sources for gst-rpicamsrc.

git clone https://github.com/thaytan/gst-rpicamsrc.git

With that done a look in the REQUIREMENTS file indicated what other packages were needed in order to accomplish the build.

sudo apt-get install autoconf automake libtool libgstreamer1.0-dev libgstreamer-plugins-base1.0-dev libraspberrypi-dev

Finally, I was able to complete the build and install.

./autogen --prefix=/usr --libdir=/usr/lib/arm-linux-gnueabihf/
make
sudo make install

The command `gst-inspect-1.0 rpicamsrc’ produces a list of the available parameters.

Factory Details:
  Rank                     none (0)
  Long-name                Raspberry Pi Camera Source
  Klass                    Source/Video
  Description              Raspberry Pi camera module source
  Author                   Jan Schmidt <jan@centricular.com>

Plugin Details:
  Name                     rpicamsrc
  Description              Raspberry Pi Camera Source
  Filename                 /usr/lib/arm-linux-gnueabihf/gstreamer-1.0/libgstrpicamsrc.so
  Version                  1.0.0
  License                  LGPL
  Source module            gstrpicamsrc
  Binary package           GStreamer
  Origin URL               http://gstreamer.net/

GObject
 +----GInitiallyUnowned
       +----GstObject
             +----GstElement
                   +----GstBaseSrc
                         +----GstPushSrc
                               +----GstRpiCamSrc

Pad Templates:
  SRC template: 'src'
    Availability: Always
    Capabilities:
      video/x-h264
                  width: [ 1, 2147483647 ]
                 height: [ 1, 2147483647 ]
              framerate: [ 0/1, 2147483647/1 ]
          stream-format: byte-stream
              alignment: au
                profile: { baseline, main, high }


Element Flags:
  no flags set

Element Implementation:
  Has change_state() function: gst_base_src_change_state

Element has no clocking capabilities.
Element has no URI handling capabilities.

Pads:
  SRC: 'src'
    Implementation:
      Has getrangefunc(): gst_base_src_getrange
      Has custom eventfunc(): gst_base_src_event
      Has custom queryfunc(): gst_base_src_query
      Has custom iterintlinkfunc(): gst_pad_iterate_internal_links_default
    Pad Template: 'src'

Element Properties:
  name                : The name of the object
                        flags: readable, writable
                        String. Default: "rpicamsrc0"
  parent              : The parent of the object
                        flags: readable, writable
                        Object of type "GstObject"
  blocksize           : Size in bytes to read per buffer (-1 = default)
                        flags: readable, writable
                        Unsigned Integer. Range: 0 - 4294967295 Default: 4096 
  num-buffers         : Number of buffers to output before sending EOS (-1 = unlimited)
                        flags: readable, writable
                        Integer. Range: -1 - 2147483647 Default: -1 
  typefind            : Run typefind before negotiating
                        flags: readable, writable
                        Boolean. Default: false
  do-timestamp        : Apply current stream time to buffers
                        flags: readable, writable
                        Boolean. Default: true
  bitrate             : Bitrate for encoding
                        flags: readable, writable
                        Integer. Range: 1 - 25000000 Default: 17000000 
  preview             : Display preview window overlay
                        flags: readable, writable
                        Boolean. Default: true
  preview-encoded     : Display encoder output in the preview
                        flags: readable, writable
                        Boolean. Default: true
  preview-opacity     : Opacity to use for the preview window
                        flags: readable, writable
                        Integer. Range: 0 - 255 Default: 255 
  fullscreen          : Display preview window full screen
                        flags: readable, writable
                        Boolean. Default: true
  sharpness           : Image capture sharpness
                        flags: readable, writable
                        Integer. Range: -100 - 100 Default: 0 
  contrast            : Image capture contrast
                        flags: readable, writable
                        Integer. Range: -100 - 100 Default: 0 
  brightness          : Image capture brightness
                        flags: readable, writable
                        Integer. Range: 0 - 100 Default: 50 
  saturation          : Image capture saturation
                        flags: readable, writable
                        Integer. Range: -100 - 100 Default: 0 
  iso                 : ISO value to use (0 = Auto)
                        flags: readable, writable
                        Integer. Range: 0 - 3200 Default: 0 
  video-stabilisation : Enable or disable video stabilisation
                        flags: readable, writable
                        Boolean. Default: false
  exposure-compensation: Exposure Value compensation
                        flags: readable, writable
                        Integer. Range: -10 - 10 Default: 0 
  exposure-mode       : Camera exposure mode to use
                        flags: readable, writable
                        Enum "GstRpiCamSrcExposureMode" Default: 1, "auto"
                           (0): off              - GST_RPI_CAM_SRC_EXPOSURE_MODE_OFF
                           (1): auto             - GST_RPI_CAM_SRC_EXPOSURE_MODE_AUTO
                           (2): night            - GST_RPI_CAM_SRC_EXPOSURE_MODE_NIGHT
                           (3): nightpreview     - GST_RPI_CAM_SRC_EXPOSURE_MODE_NIGHTPREVIEW
                           (4): backlight        - GST_RPI_CAM_SRC_EXPOSURE_MODE_BACKLIGHT
                           (5): spotlight        - GST_RPI_CAM_SRC_EXPOSURE_MODE_SPOTLIGHT
                           (6): sports           - GST_RPI_CAM_SRC_EXPOSURE_MODE_SPORTS
                           (7): snow             - GST_RPI_CAM_SRC_EXPOSURE_MODE_SNOW
                           (8): beach            - GST_RPI_CAM_SRC_EXPOSURE_MODE_BEACH
                           (9): verylong         - GST_RPI_CAM_SRC_EXPOSURE_MODE_VERYLONG
                           (10): fixedfps         - GST_RPI_CAM_SRC_EXPOSURE_MODE_FIXEDFPS
                           (11): antishake        - GST_RPI_CAM_SRC_EXPOSURE_MODE_ANTISHAKE
                           (12): fireworks        - GST_RPI_CAM_SRC_EXPOSURE_MODE_FIREWORKS
  metering-mode       : Camera exposure metering mode to use
                        flags: readable, writable
                        Enum "GstRpiCamSrcExposureMeteringMode" Default: 0, "average"
                           (0): average          - GST_RPI_CAM_SRC_EXPOSURE_METERING_MODE_AVERAGE
                           (1): spot             - GST_RPI_CAM_SRC_EXPOSURE_METERING_MODE_SPOT
                           (2): backlist         - GST_RPI_CAM_SRC_EXPOSURE_METERING_MODE_BACKLIST
                           (3): matrix           - GST_RPI_CAM_SRC_EXPOSURE_METERING_MODE_MATRIX
  awb-mode            : White Balance mode
                        flags: readable, writable
                        Enum "GstRpiCamSrcAWBMode" Default: 1, "auto"
                           (0): off              - GST_RPI_CAM_SRC_AWB_MODE_OFF
                           (1): auto             - GST_RPI_CAM_SRC_AWB_MODE_AUTO
                           (2): sunlight         - GST_RPI_CAM_SRC_AWB_MODE_SUNLIGHT
                           (3): cloudy           - GST_RPI_CAM_SRC_AWB_MODE_CLOUDY
                           (4): shade            - GST_RPI_CAM_SRC_AWB_MODE_SHADE
                           (5): tungsten         - GST_RPI_CAM_SRC_AWB_MODE_TUNGSTEN
                           (6): fluorescent      - GST_RPI_CAM_SRC_AWB_MODE_FLUORESCENT
                           (7): incandescent     - GST_RPI_CAM_SRC_AWB_MODE_INCANDESCENT
                           (8): flash            - GST_RPI_CAM_SRC_AWB_MODE_FLASH
                           (9): horizon          - GST_RPI_CAM_SRC_AWB_MODE_HORIZON
  image-effect        : Visual FX to apply to the image
                        flags: readable, writable
                        Enum "GstRpiCamSrcImageEffect" Default: 0, "none"
                           (0): none             - GST_RPI_CAM_SRC_IMAGEFX_NONE
                           (1): negative         - GST_RPI_CAM_SRC_IMAGEFX_NEGATIVE
                           (2): solarize         - GST_RPI_CAM_SRC_IMAGEFX_SOLARIZE
                           (3): posterize        - GST_RPI_CAM_SRC_IMAGEFX_POSTERIZE
                           (4): whiteboard       - GST_RPI_CAM_SRC_IMAGEFX_WHITEBOARD
                           (5): blackboard       - GST_RPI_CAM_SRC_IMAGEFX_BLACKBOARD
                           (6): sketch           - GST_RPI_CAM_SRC_IMAGEFX_SKETCH
                           (7): denoise          - GST_RPI_CAM_SRC_IMAGEFX_DENOISE
                           (8): emboss           - GST_RPI_CAM_SRC_IMAGEFX_EMBOSS
                           (9): oilpaint         - GST_RPI_CAM_SRC_IMAGEFX_OILPAINT
                           (10): hatch            - GST_RPI_CAM_SRC_IMAGEFX_HATCH
                           (11): gpen             - GST_RPI_CAM_SRC_IMAGEFX_GPEN
                           (12): pastel           - GST_RPI_CAM_SRC_IMAGEFX_PASTEL
                           (13): watercolour      - GST_RPI_CAM_SRC_IMAGEFX_WATERCOLOUR
                           (14): film             - GST_RPI_CAM_SRC_IMAGEFX_FILM
                           (15): blur             - GST_RPI_CAM_SRC_IMAGEFX_BLUR
                           (16): saturation       - GST_RPI_CAM_SRC_IMAGEFX_SATURATION
                           (17): colourswap       - GST_RPI_CAM_SRC_IMAGEFX_COLOURSWAP
                           (18): washedout        - GST_RPI_CAM_SRC_IMAGEFX_WASHEDOUT
                           (19): posterise        - GST_RPI_CAM_SRC_IMAGEFX_POSTERISE
                           (20): colourpoint      - GST_RPI_CAM_SRC_IMAGEFX_COLOURPOINT
                           (21): colourbalance    - GST_RPI_CAM_SRC_IMAGEFX_COLOURBALANCE
                           (22): cartoon          - GST_RPI_CAM_SRC_IMAGEFX_CARTOON
  rotation            : Rotate captured image (0, 90, 180, 270 degrees)
                        flags: readable, writable
                        Integer. Range: 0 - 270 Default: 0 
  hflip               : Flip capture horizontally
                        flags: readable, writable
                        Boolean. Default: false
  vflip               : Flip capture vertically
                        flags: readable, writable
                        Boolean. Default: false
  roi-x               : Normalised region-of-interest X coord
                        flags: readable, writable
                        Float. Range:               0 -               1 Default:               0 
  roi-y               : Normalised region-of-interest Y coord
                        flags: readable, writable
                        Float. Range:               0 -               1 Default:               0 
  roi-w               : Normalised region-of-interest W coord
                        flags: readable, writable
                        Float. Range:               0 -               1 Default:               1 
  roi-h               : Normalised region-of-interest H coord
                        flags: readable, writable
                        Float. Range:               0 -               1 Default:               1

Usage

Because of the way gstreamer works, the parameters for the feed needed to be split and re-arranged in the streamer pipeline. Previously all the parameters are specified as part of raspivid.

/opt/vc/bin/raspivid -t $DURATION -w $WIDTH -h $HEIGHT -fps $FRAMERATE -b $BITRATE -n -pf high -o - | gst-launch-1.0 -v fdsrc ! 

GStreamer parameters like width, height and frame rate are configured through capabilities (caps) negotiation with the next element. Other parameters like the bit rate and preview screen are controlled as part of the source element.

gst-launch-1.0 rpicamsrc bitrate=$BITRATE preview=0 ! video/x-h264,width=$WIDTH,height=$HEIGHT,framerate=$FRAMERATE/1 !

The new stream script is

#!/bin/bash

source remote.conf

if [ "$1" != "" ]
then
  export FRAMERATE=$1
fi

NOW=`date +%Y%m%d%H%M%S`
FILENAME=$NOW-Tx.h264

gst-launch-1.0 rpicamsrc bitrate=$BITRATE preview=0 ! video/x-h264,width=$WIDTH,height=$HEIGHT,framereate=$FRAMERATE/1,profile=high ! h264parse ! rtph264pay config-interval=1 pt=96 ! udpsink host=$RX_IP port=$UDPPORT

Tests

This test was a straight comparison between the old and new scripts using the same settings and the same testing methodology as previously.  The resolution was 1280 x 720 pixels with a 6Mbps bitrate.

Update : Since the original article was published use Jan Schmitt spotted that I had misspelled “framerate” as “framereate” in the gst-rpicamsrc script.  He also suggested I should try using the baseline profile and a queue element to decouple the video capture from the UDP transmission.  With this in mind I have re-run the tests.

Results

Almost immediately I had the feeling that gst-rpicamsrc has slower.  Analysis of the video showed I was correct.  The latency using gst-rpicamsrc was 18% higher than using raspivid.

Update:

Running the original erroneous script with debugging on showed that the capture was running at 30fps instead of the intended 48 fps. Here are the new results averaged from 10 cycles.

Script gst-rpicamsrc @ 48 fps
raspivid @ 48 fps
Profile No queue With queue No queue With queue
baseline 184.2 175.4 153.9 156.9
high 186.2 185 154.1 159.7

gst-rpicamsrc @ 30 fps, high profile, no queue = 198.2 ms

Analysis

The first thing to note is that the raspivid latency (no queue, high profile) has risen from the 126ms found in the last tests to 154ms.  The only difference was that I cloned the Sandisk microSDHC card onto a Transcend 8GB.  I’ll set up some more tests to compare the cards.  As these tests were run from the same card and from the same boot, they are still valid for comparison.

It is immediately obvious that the gst-rpicamsrc latency is about 20% higher than the raspivid script, so the conclusion from the first publish of this article still stands.

What can be added is that using the baseline profile, does reduce the latency a little: 1 to 3ms in most cases.

Adding a queue element does provide a benefit for the gst-rpicamsrc script, especially with the baseline profile where a 9ms reduction in latency was observed.  For the raspivid script adding a queue element actually increased the latency by 3 to 4ms.  I suspect this is because the video stream is already decoupled from gstreamer by being piped in from an external process.

Conclusion

Using gst-rpicamsrc provides no benefit for reducing latency over raspivid.  That is not to say gst-rpicamsrc provides no other benefits.  For any use other than FPV, I would definitely use gst-rpicamsrc instead of having to pipe the video in through stdin.  It provides plenty of options for setting up the video stream as the command `gst-inspect-1.0 rpicamsrc’ above showed.

The problem here is that I am targeting this development for FPV use where low latency is the driving factor. At the moment my lowest latency for a adequate quality HD stream is 125ms and I really need to get this under 100ms to compete with current analog standard definition systems.  Whether it is possible to shave of another 25ms remains to be seen.

Update: Following the additional tests I would add that it is better to use the baseline profile over the high profile.

Effect of Frame & Bit rates on Latency

Extract

Back in March, an update to the Raspberry Pi camera software introduced some new video modes with higher frame rates for resolutions below full 1080p HD.  For 720p HD, frame rates up to 49 fps are now available and for VGA (640×480) there are new modes for 60 and 90 fps.  There were reports that, at these higher frame rates, the latency was reduced.  This didn’t come as a surprise as it has been known for a while that the GoPro video out has lower latency at 48 fps than at 30 fps.  The reasons for this are not totally clear, but is thought to be due to the differences between frames being less at the higher frame rates allowing for more efficient compression.

In order to investigate this I ran some back-to-back tests at various frame rates.  I was expecting that with twice as many frames being squeezed into the same bit rates, the video quality would suffer, so the tests where repeated at three different bit rates.  The result of this testing was a 25% reduction in latency, with a new minimum of 121 milliseconds at 720p and 48 fps.  As expected the video quality at the old bit rate of 2.5Mbps suffered for fast moving backgrounds, but even when increased to 6.0Mbps, the latency was still only 126 Mbps; a 23% reduction.

With VGA resolution at 60fps and 2.5 Mbps the latency dropped to 87 ms and while this is useful, in terms of hardware size, convenience and even latency, it is still bettered by dedicated FPV hardware.  If you include the cost of the laptop, it looses out on price too.  Where the Pi has the advantage is at the 1280 x 720p HD resolution where there is currently no affordable competition.

Setup

The test environment was set up the same as previously, with the camera pointing at the laptop screen.  The camera was positioned so that at least ten iterations of the image were visible at once. I used a single LED torch as the measurement indicator.  The tests were run with no overclocking.

I used a UDP stream with some improved scripts to simplify running all the variations. The scripts can be found on the Current Scripts Page.

A GoPro 3 Black was used to record the screen at 120 fps.

Tests were run at frame rates of 25, 30, 36, 42, 45, 48 and 49 fps and at bit rates of 2.5, 4.5 and 6.0 Mbps.   Up to 5 LED on/led off cycles were recorded on the GoPro for later analysis.  With 10 iteration for each of the 5 on/offs, the calculated latency represents the average of 100 tests.

Results 1 – Latency

Across all the bit rates, the results show a definite reduction in latency with increasing frame rate right up to 48 frames per second. At 49 fps the latency increased slightly. The lowest latency achieved was 121ms at 48 fps for the 2.5 Mbps stream.  A reduction of 25% over the 25 fps result.  The 6.0 Mps stream still managed a reduced latency of 126 ms at 48 fps and the 4.5 Mbps stream achieved 125 ms.

The VGA test showed a much lower latency at 87ms, even though it was using the same bit rate as the smallest 720p HD stream.

The graph shows some odd behavior between the 4.5 and 6.0 Mbps curves.  Below about 40 fps the 4.5 Mbps stream has higher latency than the 6.0 Mbps stream.  Beyond 40 fps the curves switch over to what you would expect.

fps

In addition to the average latency, the graph below shows the minimum and maximums for each bit rate.  What is apparent is that there is a lot of overlap between the results and that the latency can vary +/- 20 fps, particularly at the lower bit rates.

fps2

Results 2 – Video Quality

The frame grabs below show a clear difference between the three bit rates at 48 fps. The 2.5 Mbps stream shows extreme pixelation and a general blurring of the image.

Quality25

Frame grab from 2.5 Mbps stream @ 48 fps

The  4.5 Mbps shows a reduction in the pixelation.  This is apparent in the sky and on the tiled roof.

Quality45

Frame grab from 4.5 Mbps Stream @ 48 fps

The 6.0 Mbps is the clear winner, with a minimum of pixelation and the sharpest image.

Quality60

Frame grab from 6.0 Mbps Stream @ 48 fps

Conclusion

The total latency is made up of capture, compression, transmission and display components.  The capture and  display components should be the same for the three bit rates as the resolution is the same.  The transmission component should increase in relation to the higher bit rates.  That just leaves the compression component.  The compressor has to work hard to squeeze the stream into a smaller bit rate whilst maintaining the best video quality possible.  For this reason decreasing the bit rate can increase the latency, although there is probably a point where the extreme pixelation starts reducing the latency again.

These opposing affects between compression and transmission are likely what has caused the odd behavior between the 4.5 Mbps and 6.0 Mbps streams.  At the lower frame rates, the compression delay for the 4.5 Mbps stream was more significant.  At the higher frame rates the transmission delay in the 6.0 Mbps stream became more significant.

While the VGA test used the same 2.5 Mbps bit rate as the HD test, with only a third of the pixels, the capture, compression and display components resulted in a much lower latency of 87 ms.

While the lowest latency was achieved for the 2.5 Mbps stream, the extreme pixelation of the image renders it pretty much unusable.  The 4.5 Mbps stream could be used but it only has a 1 ms advantage over the superior image quality of the 6.0 Mbps stream.  For this reason I plan to adopt the 48 fps, 6.0 Mbps, 126 ms stream as my new baseline when flying.

I will probably also switch to a 24 fps pipeline for youtube videos to get the best image quality from the 48 fps video, dropping every other frame.  I have seen this done where people have used the 48 fps recording on GoPro cameras.

Raspberry Pi Camera Latency Testing – Part 2

Netcat v Gstreamer

Extract

After a number of tests I have established that the Raspberry Pi is able to provide a low latency HD stream over a wifi connection.  There are limitation, but the results have been encouraging enough for me to push from desk to airborne testing.

Introduction

Back in May 2013 when the RPi camera appeared I did some latency testing to see if the platform might be suitable for FPV duties.  Using the default NetCat based solution the minimum latency I could achieve was 300ms and the stream wasn’t particularly smooth, so the conclusion was no.

Then a few couple of months later I was asked to try running the tests again using gstreamer instead of netcat.  It’s taken longer than I had hoped, but I scrapped my first set of results and re-ran them with a more structured methodology.  As the results were encouraging, I expanded the tests to see how far it would go.

In fact the tests were so encouraging I kept extending and extending and never got around to posting the results.  Now with the UK locked in perpetual rain and wind, I though I’s take the time to share the results.

Hardware Components.

For the tests I used:

  • Raspberry Pi model B
  • Raspberry Pi Camera.
  • Edimax EW-7612UAn V2 USB WiFi Adapter (with high gain antenna).
  • MSI Wind U100 Single Core 1.6 GHz Atom Netbook, Circa 2008
  • DELL XPS 17, Quad Core i7-2630QM 2.0GHz Laptop, Circa 2012

Raspberry Pi System Setup.

Initial setup was done with the Pi connected via LAN using the then current build of Raspbian.  The first tasks were to expand the storage to fill the SD card, enable the Camera, enable ssh and update the system using apt-get.

I followed Gbaman’s instructions on setting up gstreamer on the Pi, although I used Lubuntu as the client instead of mac os.  Lubuntu was run from a USB stick so I could move it between laptops to compare performance. This was the system I used for the first set of tests which was later discarded as I felt the system could be improved.

I wanted the Pi to act as a WiFi access point.  This removes the need for a WiFi base station without having to use an ad-hoc network. I followed Adafruit’s instructions to setup hostapd, but not isc-dhcp-server or NAT translation. For dhcp I setup dnsmasq based on the first half of Adafruit’s Ad blocking access point setup,  stop at “dig adafruit.com”.  With this setup I could connect the laptop to the Pi’s WiFi and login over ssh.  Apart from plugging in the power, no other connections to the Pi are required.

I created a script on the Pi to start the video stream and another on the laptop to receive it.  The Pi camera was pointed at the laptop screen so multiple iterations could be recorded and the whole test was recorded on a GoPro at 120 frames per second.  For some of the early tests a screen was connected to the Pi’s hdmi output to show the local preview.

Test 1

The first tests started where the first video finished at 640 x 480 pixel resolution and 0.7Mbit/s bitrate run from the RPi desktop with local preview.  The gstreamer feed was about three times faster than netcat.  The gstreamer feed was also more consistent.

GS2

Netcat : 505 ms ±50%, Gstreamer : 167 ms ± 17%

This sub 200ms latency for gstreamer is probably a usable value for FPV, so initial results were promising.

Test 2

For the second test gstreamer was started from the Pi’s command prompt instead of the desktop

NetCat : 514 ms 60%,    Gstreamer 174ms 30%

NetCat : 514 ms ±64%, Gstreamer 174ms ±30%

The results were actually slightly slower for both systems but gstreamer was still under the 200ms mark.

Netcat testing was dropped at this point as it was never going to compete with gstreamer.

Tests 3 & 4

 The next tests were run with the Pi preview removed as this would never be needed in a FPV system. The tests were run for 640 x 480 x 0.7 Mbit/s, 720 x 480 x 1 Mbit/s and 1280 x 720 x 2.5 Mbit/s streams. I also ran another test at the native screen size of the Netbook 1024 x 600 x 1.6 Mbit/s, but this is not shown on the video or graph.

Test 3 & 4


640 x 480 x 0.7 Mbit/s 186ms ± 18%
720 x 480 x 1.0 Mbit/s 169ms ± 20% (Ignoring initial long iteration)
1024 x 600 x 1.6 Mbit/s 163ms ± 5%
1280 x 720 x 2.5 Mbit/s 194ms ± 24%

The averaged results are strange because for the first three results the latency comes down as the resolution and bit rate go up.  The difference is still withing the measurement accuracy so is probable that the latency for the lower resolutions are actually all in the same ballpark.  Only once we hit the lower HD resolution of 1280 x 720 x 2.5 Mbit/s does the latency go back up, suggesting that a saturation point has been reached somewhere in the pipeline.

Tests 5 & 6

These tests were run to compare results between having a Single core 1.6 Ghz Atom Netbook and a Quad Core i7-2630QM 2.0GHz Laptop at the receiving end. The tests were run with 640 x 480 x 0.7 Mbit/s and 1280 x 720 x 2.5 Mbit/s streams

640 x 480 x 0.7 Mbit/s    Netbook     186ms ± 18% Laptop       154ms ± 30% 1280 x 720 x 2.5M bit/s  Netbook     194ms ± 24% Laptop       171ms ± 17%

640 x 480 x 0.7M bit/s Netbook 186ms ± 18%
640 x 480 x 0.7 Mbit/s Laptop 154ms ± 30%
1280 x 720 x 2.5 Mbit/s Netbook 194ms ± 24%
1280 x 720 x 2.5 Mbit/s Laptop 171ms ± 17%

On the quad core laptop the 1280 x 720 stream has dropped back down to the 170ms range and the 640 x 480 stream has dropped to almost 150ms.

Test 7

I re-ran the test again at the full HD resolution of 1920 x 1080 x 6 Mbit/s.  This worked OK on the laptop although the latency rose to 233ms ± 27%. This may be usable on a big powered glider or attitude stabilised multicopter, but is pushing it for rate or acro mode flying.

The netbook took almost 9 seconds (8982ms ± 2%) for one iteration.  If you want to run full HD you are going to need a good laptop at the receiving end.

Test 8

Now I had a usable stream, the next requirement is to record it.  On the Raspberry Pi we are sending the camera output to gstreamer through a stdout/stdin pipe.  This makes it easy to to capture the stream to a file using the tee command.  The *nix tee command copies the stdin stream to stdout and saves a copy to a specified file. This allows a copy of the video stream to be saved to the SD card.

raspivid <options> | tee file.h264 | gstreamer <options>

tee

The results showed that the tee command added nothing to the latency.  This means the stream can always be recorded at the sending end without worrying about missed frames due to WiFi dropout.

Recording the stream at the receiving end is a much harder problem.  Without writing a dedicated gstreamer application I can’t see an easy way to save the stream to a valid mp4 file.  I’ve got a workaround that allows me to capture the stream using gst-launch-1.0 and mux it into a usable file, but at the moment, it requires multiple steps across Linux and Windows.   I’ll document that in a later post.

Summary

  • Best SD latency (640 x 480): 154ms
  • Best HD latency (1280 x 720): 171ms
  • Best Full HD latency (1920 x 1080): 233ms
  • Saving a copy of the stream on the RPi using the tee command does not affect the latency.

Conclusions

Based on these results I believe the Raspberry Pi with camera could be used as an HD FPV platform in conjunction with a reasonably good laptop, although I would limit it to 720p HD and not 1080p.  Because the system uses 2.4Ghz Wifi consideration needs to be given to interference from other WiFi networks and 2.4Ghz RC transmitters.  This could be solved using 5Ghz WiFi dongle provided a USB dongle can be found that’s compatible with the Raspberry Pi and features a high gain antenna.

A second consideration is range.  In a house WiFi range can be as little as 10m due to various walls.  However, in free air the range can be a lot more.  Testing is required to establish the actual usable range.

During the tests 1 & 2 I recorded the time to display the preview on the RPi.  The latency for this was typically 86 ms.  As this preview was shown before the stream was compressed it can be assumed that at least half of the latency comes in reading and encoding the stream in the Raspberry Pi.  Any optimization that could be made here would have a significant benefit.

The Next Step

I’ve got an old slope soarer that uses a 35Mz RC system which won’t interfere with the 2.4GHz Wifi.  I’ve made a power pod for it and will start some some line of sight range testing with the Raspberry Pi banded on top of the wing.  I’ve soldered up some cables to connect a Hobbywing 3A BEC to a USB Micro B plug to feed the required 5V to the Raspberry Pi.