Micro bounty: visiond PC version

[nanamiwang]

A visiond version for linux, tested on ubuntu 16.04.

Features

camera_eon_stream.cc:

• Connect to Eon streaming server (loggerd --stream), decode the hevc YUV420 frames using libavcodec API.
• Emulate a rear camera, dispatch YUV420 frames to the visiond_pc.cc processing thread.

visiond_pc.cc

• A clone of visiond.cc
• Launch the emulated camera, run driving model and posnet model, so that all the rear camera vision pipeline can work as normal.
• Removed dependencies on android ion, visionbuf and visionimg.
• Removed streaming server support.
• Removed front camera and drive monitoring features.

Other changes

• Add some missing header inclusions in model.cc, snpemodel.cc and monitor.cc.
• Use CPU mode for x86_64 in snpemodel.cc.
• Fixed a memory leak in model.cc.
  Have tested the code on Eon, so that all the changes will not affect the android version of visiond.

Build Prerequisites

• Install SNPE and set SNPE_ROOT env.
• Build and install zmq & czmq.
• Install libavcodec-dev.

How to use

On Eon
Launch loggerd and visiond

loggerd --stream
visiond

On ubuntu
Launch visiond, set the target Eon IP address using env EON_IP:

EON_IP=192.168.1.105 ./visiond

Launch calibration daemon to handle posnet inference result and do live calibration:

python calibrationd.py

Launch openpilot-tools replay UI to see the decode frames and model inference result:

python openpilot-tools/replay/.py

How to verify the model inference result

• Play a recorded drive video on PC, like this:
  https://github.com/commaai/comma2k19/blob/master/Example_1/b0c9d2329ad1606b%7C2018-08-02--08-34-47/40/video.hevc
• Point the Eon at the screen, make sure the lanes is within the blue box on openpilot-tools replay UI.
• Check the model output window, will see that the lanes is detected correctly by driving model.