Sunday, 23 June 2019

Jetson Nano - Developing a Pi v1.3 camera driver Part 1

I liked to thank for sponsoring the hardware and development time for this article. 

The jetson nano is fairly capable device considering the appealing price point of the device. In fact its one of the few ARM devices which out of the box provides a decent (and usable) X11 graphics stack (even though the drivers are closed source).
Although the jetson nano supports the same 15 pin CSI connector as the RPI camera support is currently limited to Pi V2 cameras which is host the imx219. The older Pi v1.3 cameras are appealing partly because there are numerous low cost clones available and partly because there are numerous add ons such as lenses and night mode options.

The v1.3 cameras uses the OV5647 which apparently is discontinued by OmniVision furthermore the full datasheet isn't freely available (only under NDA). There is a preliminary datasheet on the internet but it seems to be incomplete or worse inconsistent in places. This does hinder the process some what as debugging errors can be very time consuming and at time frustrating.

One noticeable different is that the v1.3 camera hosts a 25Mhz crystal where most non rpi OV5647 boards use a standard 24Mhz. This can make the tuning more difficult as some of the default setting need adjustments.

The first step in bringing up the camera was ensuring the board was powered on so that it could be detected for through its i2c interface (address 0x36). After numerous attempts the OV5647 finally appeared:

Warning: Can't use SMBus Quick Write command, will skip some addresses
WARNING! This program can confuse your I2C bus, cause data loss and worse!
I will probe file /dev/i2c-6.
I will probe address range 0x03-0x77.
Continue? [Y/n] Y
     0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f
30: -- -- -- -- -- -- UU --                        
50: -- -- -- -- -- -- -- -- -- -- -- -- -- -- -- --

The second step was to develop enough of a skeleton kernel driver to initialise the OV5647 and enable it as v4l2 device. Although this sounds may easy it turned out be extremely time consuming for two reasons. Firstly due to the lack of documentations for OV5647 and secondly the NVIVIDA camera driver documentation is also poor and a in number of cases the documentation doesn't match the code. Finally after a few weeks a v4l2 device appeared:

jetson-nano@jetsonnano-desktop:~$ v4l2-ctl -d /dev/video0 -D
Driver Info (not using libv4l2):
    Driver name   : tegra-video
    Card type     : vi-output, ov5647 6-0036
    Bus info      :
    Driver version: 4.9.140
    Capabilities  : 0x84200001
        Video Capture
        Extended Pix Format
        Device Capabilities
    Device Caps   : 0x04200001
        Video Capture
        Extended Pix Format

Next step was to put the camera in test pattern mode and capture a raw image. The OV5647 outputs raw bayer format in our case 10 bit so the captured raw data file needs to be converted to a displayable format. Conversion can be done using a utility like bayer2rgb. Finally I arrived a valid test pattern.

Next stage was to configure the OV5647 to a valid resolution for image capture again which has been extremely challenging for the reasons stated above. Some of the images from numerous attempts are shown on the left and right.

Current progress is that the camera is outputting 1920x1080@30fps however this is work in progress as the driver is in a primitive state and the output format requires further improvements. On the plus side to it is now possible to stream with the nvarguscamerasrc gstreamer plugin. Below is a 1080 recording from the OV5647 with a pipeline based on nvarguscamerasrc and nvv4l2h264enc.

Update: In my 2nd post we have a driver that you can test with.