Traffic Light Detection

Objectives

• Detect traffict lights from traffic iamges (videos)
• Classify the lights into 3 colors
• This is used for self driving car

Approach

• Use transfer learning of pre-trained model
• Atchitecture
  • Use pretrained Mobile Net (CNN) for quick and memory saving detection, since speed is reauired. Mobile Net is trained for modification for this task
  • Use SSD (Single Shot Detection) to detect the location and the classification at the same time
  • SSD_Mobilenet 11.6.17 version from Tensorflow model zoo
• Training Data: Used Bosch Small Traffic Light dataset (already annoted the class and positions) https://github.com/bosch-ros-pkg/bstld

Result

• See video above
• Detected red lights well, but still some missing lights
• Detecting other color lights not tested

Setup (for Windows 10)

• Download Bosch dataset and convert that into tfrecords format

• Setup the transfer learning (Following comes from comes from repo of Alex Lechner's repo . Appreciate the describing the complex Tensorflow (of an older version of TF that has limited info.) setup!)

• Install Tensorflor 1.4: pip install tensorflow==1.4 (newer version is not compatible!)

  • (for Linux machine with GPU, pip install tensorflow-gpu==1.4)

• Install relevant packages: pip install pillow lxml matplotlib

• Download protoc-3.4.0-win32.zip to extract to C:\Program Files\protoc-3.4.0-win32

  • (for Linux: sudo apt-get install protobuf-compiler python-pil python-lxml python-tk)

• Create /TensorFlow directory

• Clone Models repository into TensorFlow directory: git clone https://github.com/tensorflow/models.git

• Navigate to the /models directory and change the repo version to: git checkout f7e99c0

• Navigate to research folder, and creating .py files by executing: "C:\Program Files\protoc-3.4.0-win32\bin\protoc.exe" object_detection/protos/*.proto --python_out=.

• Add following paths to environment variable PYTHONPATH (you may need to restart the terminal to activate the env variables):

  • models
  • models/research
  • models/research/slim
  • models/research/object_detection

• run: python builders/model_builder_test.py at /object_detection directory

• configure a tensorflow config file and place in /config folder.

• train the model: python train.py --logtostderr --train_dif=./models/train --pipeline_config_path=./config/<TFMODELCONFIG.config>

• freeze the model: python export_inference_graph.py --input_type image_tensor --pipeline_config_path./config/ssd_mobilenet_c1-B.config --trained_checkpoint_prefix ./models/train/model.ckpt-2000 --output_directory models

Reference

• Tensorflow Object Detection install https://github.com/tensorflow/models.git
• Bosch Small Traffic Light dataset

Note

• This repository is not fully updated, there maybe some parts that are not updated