Traffic Light Detection Node - Ridgebeck/CarND-Capstone-Team-Herbie GitHub Wiki

Example of traffic light classification:

https://github.com/alex-lechner/Traffic-Light-Classification

Possible data sets:

  • Bosch Small Traffic Lights Dataset
  • LaRA Traffic Lights Recognition Dataset
  • LISA Traffic Light Dataset
  • Udacity's ROSbag file from Carla

Configuration

Check ros/src/tl_detector/sim_traffic_light_config.yaml or site_traffic_light_config.yaml

Model

Path: ros/src/tl_detector/light_classification/model.
Which model to use is configured in the yaml files.

Train with AWS:

  • please make sure it has enough graphic memory for training. you can use p2.xlarge, g3 4x large has better GPU performance, but only have 8G graphic memory.
  • when use udacity-carnd-advanced-deep-learning image, need reinstall python 3.5. Run the following command firstly, then install other module.
  1. conda create -n py35 python=3.5
  2. source activate py35
  • install more package. pip install matplotlib pip install Pillow
  • if you see the following error, please make sure clean up your train_dir

/ExponentialMovingAverage not found in checkpoint 2017-07-02 22:59:48.794763: W tensorflow/core/framework/op_kernel.cc:1165] Not found: Key BoxPredictor_1/ClassPredictor/biases not fo und in checkpoint 2017-07-02 22:59:48.794821: W tensorflow/core/framework/op_kernel.cc:1165] Not found: Key BoxPredictor_1/BoxEncodingPredictor/biases/ RMSProp_1 not found in checkpoint

Troubleshootings

  • AttributeError: 'module' object has no attribute 'mark_flag_as_required'
    If you get this message when trying to export the model, you need to upgrade your TL to 1.4.
    pip install --upgrade tensorflow-gpu==1.4