Jetson | http : //jetson.unhcr.org | Machine Learning library

The rtsp server uses TCP because your client query asked for that using rtspt where last t queries for TCP transport. Just using rstp instead should use UDP. You may have a look to protocols property of rtspsrc for more details.

Full story is in the comments here and continued to solution here: Gstreamer Android HW accelerated H.264 encoding

You would use decodebin that would select the decoder according to a rank built into plugins. That should select the most efficient one available for decoding:

After doing some research I learned about the difference between joydev and evdev on linux. It seems my PS4 controller showed up as a joydev device on the Jetson, whereas on the Coral running Mendel it showed up as an evdev device. I found some example C++ code for evdev devices from Croepha / David Butler here: https://handmade.network/forums/t/3673-modern_way_to_read_gamepad_input_with_c_on_linux

I modified this code for the PS4 controller like this:

There is no linux-aarch64 version of pytorch on the default conda channel, see here

This is of course package specific. E.g. there is a linux-aarch64 version of beautifulsoup4 which is why you wre able to install it without an issue.

You can try to install from a different channel that claims to provide a pytorch for aarch64, e.g.

I am using Python 3.7 in the conda environment and Python 3.6 outside.

This is the issue. You have a cp36 whl file, so python 3.6. I am suspecting that when you run sudo pip, your systems pip is invoked, whereas when you run pip, then pip from your conda env is used, and cannot install a python 3.6 whl to a python 3.7 env.

Either you need to get the cp37 whl or create a conda env that has python 3.6 installed

You might want to have a look at the following article which shows how to do the connection with core Python Socket library https://blog.kevindoran.co/bluetooth-programming-with-python-3/.

The way BlueZ does this now is with the Profile API.

There is a Python example of using the Profile API at https://git.kernel.org/pub/scm/bluetooth/bluez.git/tree/test/test-profile

hciattach, hciconfig, hcitool, hcidump, rfcomm, sdptool, ciptool, and gatttool were deprecated by the BlueZ project in 2017. If you are following a tutorial that uses them, there is a chance that it might be out of date and that Linux systems will choose not to support them.

Your problem is that decodebin selects nvv4l2decoder that outputs into NVMM memory. videoconvert cannot read from NVMM memory. You would use nvvidconv instead that can read from NVMM and output into system memory.

However, it is not mandatory to decode h264 for reencoding into h264. This simple pipeline should do the job:

decodebin internally selects subplugins for each type of parsing/decoding. There is a ranking that allows to choose when several plugins are available for a type of decoding. For H264 decoding, on Jetsons the highest rank would be nvv4l2decoder, that outputs into NVMM memory (in NV12 format unless otherwise specified) that is suitable for autovideosink that will instantiate plugin nvoverlaysink reading from NVMM memory.

For reading from OpenCv application, you would need frames in system memory, and better in BGR format for color frames. So you would first convert NV12 in NVMM memory output of the decoder into BGRx format and copy into system memory with plugin nvvidconv. Then use CPU-based plugin videoconvert for conversion into BGR:

You would have to specify V4L capture backend for using V4L controls. Seems here that without specifying it, gstreamer backend is used, and it instanciates a pipeline with v4l2src, but you cannot change the resolution after creation.

So the simplest solution is just using V4L2 capture backend: