TL Detector: Raise classifier confidence threshold and add RED->UNKNOWN

team-fusionx / CarND-Capstone

Team FusionX - Capstone Project for Udacity Self-Driving Car Nanodegree

MIT License

8 stars 3 forks source link

TL Detector: Raise classifier confidence threshold and add RED->UNKNOWN #136

Closed edufford closed 6 years ago

edufford commented 6 years ago

Based on classifier accuracy data from FX-v2.2 Carla site test, raise classification confidence threshold so that <70% will become UNKNOWN light state.
Add RED->UNKNOWN transition to be allowed if car is already stopped for the red light and light state becomes unknown (possible green light). Velocity signal smallest non-zero value in data is ~0.2 m/s so threshold to judge stopped is set to 0.3 m/s to be slightly above this.

dinoboy197 commented 6 years ago

The code looks good. However, someone else should take a look at the performance with an Nvidia GPU (I have none and I've had trouble getting a locally compiled version of Tensorflow to run properly with optimized CPU instructions).

meenunat commented 6 years ago

I am currently travelling and may not be able to help in testing the code with GPU.

edufford commented 6 years ago

Thanks for the code checks. Hopefully @teeekay can test it with GPU later on this weekend or next week.

teeekay commented 6 years ago

Ran tests on Simulator Track 1, and using rosbag images from the submission 2.2 site test. The Simulator ran perfectly. When using the rosbag, performed as expected when close to traffic light with detection of UNKNOWN about 1 second after light switches from Red to Green. Detection of Yellow is also poor at this distance, but this will hopefully not be an issue for this test run.