KAIST traffic light control system using video image processing based on DNN and adversarial NN

KAIST safety-team still use human-power for handling traffic and crosswalk. Therefore, we suggest deep-learning based traffic handling system. We used YOLONET for object(human,car,ducks) detection and adversarial neural etwork for traffic handling. We also compare this adversarial NN with our handmade algorithm.

• python 3.5.2 environment
• anaconda
• yoloNet python library (download link : https://github.com/qqwweee/keras-yolo3)
• opencv-contrib-python (link : https://pypi.org/project/opencv-python/)
• PIL library
• required pip packages :

  tensorflow                    1.6.0
  keras                         2.1.5
  tensorflow-gpu                1.0.1
  matplotlib                    3.0.3       (link : https://pypi.org/project/matplotlib/)
  opencv-python                 4.1.0.25    (link : https://pypi.org/project/opencv-python/)
  Pillow                        6.0.0       (link : https://pypi.org/project/Pillow/2.2.1/)
  Cython                        0.29.7      (pip install cython)
  other packages needed for above

1. construct tensorflow conda environment

  conda create -n deep python=3.5.2 tensorflow=1.6.0 tensorflow-gpu=1.0.1 keras=2.1.5

2. download yoloNet-python library (link)
3. download required pip packages (opencv, link)

• /detection

  • setting_opencv.py : make calibration of skewed angle, crosswalk, central line and neighbor lane need for position detection using opencv-python library
  • setting_cnn.py : same function but increase accuracy using CNN
  • measure.py : using calibrated value, measure speed, position of detected object in YOLO net

• /application

  • server, for information share and decision making for traffic handling
  • client, for information calculation and send to server

• /data

  • dataset of collected and preprocessed image(frame) and
  • trained weight file included

• /decision

  • /algorithm : decision making based on measured value and algorithm
  • /rein_learn : decision making based on reinforcement learning using SUMO simulation https://github.com/jshan2017/QL-sumo

editor : JaeminBest