object-detector

Introduction

• Detect objects in images, videos and video streaming from a webcam.
• A YOLOv3 predictor writen in python with keras.
• Product of a personal investigation to internalize myself in the matter.
• Strongly based :) on YOLOv3 and keras-yolo3.

Sample Video

Webcam Screenshot

Sample Image

Requeriments

• anaconda
• 7z
• A respectable video card. i.e. GeForce GTX 1060 or higher (Recommended).
• A respectable webcam. i.e Logitech c922 (Recommended).

Setup

Step 1: Create project environment.

conda env create --file environment.yml

Step 2: Extract network weights.

7z x  model_data/yolo.h5.7z.001 -o./model_data 

Step 3: Before all you need activate object-detector environment every time you use it with this:

source activate object-detector

or forget use it creating a bash/zsh alias:

• bash

  echo -e "export PATH=.:\$PATH:$PATH" >> ~/.bashrc
  echo "alias object-detector='source activate object-detector;python object-detector.py'" >> ~/.bashrc
  source ~/.bashrc

• zsh

  echo -e "export PATH=.:\$PATH:$PATH" >> ~/.zshrc
  echo "alias object-detector='source activate object-detector;python object-detector.py'" >> ~/.zshrc
  source ~/.zshrc

with this you can use object detector as a regular command in the following way:

object-detector params

instead of:

source activate object-detector
python object-detector.py params

Use

• Detect objects in an image:

  object-detector \
      --input-image input/image.jpg \
      --output output/processed_image.jpg \
      --predict-bounding-boxes

• Detect objects on video:

  object-detector \
      --input-video input/video.mp4 \
      --output output/processed_video.mp4 \
      --predict-bounding-boxes \
      --show-preview

• Detect objects in a video streaming from webcam:

  object-detector \
      --input-webcam 0 \
      --output output/processed_video.mp4 \
      --output-fps 10 \
      --predict-bounding-boxes \
      --show-preview

  Note: --input-webcam: 0 is the integrated laptop webcam video port number, but also you can use another ports for external webcams or external video input devices. Values: From 0 to N.

• Press ESC key to end process. When input is a video this end process before process all video.

• To view all options:

  $ object-detector --help
  usage: object-detector [-h] [--input-image INPUT_IMAGE]
                         [--input-video INPUT_VIDEO] [--output OUTPUT]
                         [--input-webcam INPUT_WEBCAM] [--output-fps OUTPUT_FPS]
                         [--show-preview] [--preview-width PREVIEW_WIDTH]
                         [--predict-bounding-boxes]
  
  YOLO object detector :)
  
  optional arguments:
    -h, --help            show this help message and exit
    --input-image INPUT_IMAGE
                          Path of an image file.
    --input-video INPUT_VIDEO
                          Path of a video file.
    --output OUTPUT       Path of an output file.
    --input-webcam INPUT_WEBCAM
                          Input video port. Available ports: [0, 2]
                          (/dev/videoX).
    --output-fps OUTPUT_FPS
                          Output video FPS. Often used with webcam input videos
    --show-preview        Show preview window.
    --preview-width PREVIEW_WIDTH
                          Preview window width.
    --predict-bounding-boxes
                          Predict & plot bounding boxes

  Note:You can see that --input-webcam description shows video ports of currently active devices. In this case:

  • 0: Integrated laptop webcam.
  • 2: External USB webcam.

Known issues

• ...window.cpp:698: error: (-2:Unspecified error) The function is not implemented.

  pip install --upgrade pip
  pip install opencv-contrib-python

• macOS: Webcam input only works under classic terminal but not in another like iterm.

Future tasks

• Jupiter notebook to make transfer learning to predict another classes.
• Use Nvidia NVENC encoder/decoder (Process video using GPU cuda cores) to increase video frame rate.
• Rewrite yolo model to improve my understanding of the model.