This project is aimed to detect any specific type of object in real time by passing a dataset of the object. There are three folders that show the different algorithms in this project.

• Has support to use Zynq UltraScale+ as a processor for optical flow for Faster R-CNN

Use the obj-detection-frcnn for best results.

All of the algorithms are tested on the Town Centre data set. Results for each algorithm are shown below:

• Ubuntu 18.04
• Python 3
• Keras
• Tensorflow
• OpenCV
• Imutils
• Numpy

Note: Make sure to use a computer with a powerful GPU(s) or Google Cloud/Amazon instance with a GPU(s) for best efficiency

• Download TownCenterXVID.avi from the Town Centre data set and rename the video to test.avi
• Copy and paste the text from TownCentre-groundtruth.top from the data set website into a file called data.csv
• Create an empty folder called data in the corresponding algorithm folder(i.e. obj-detection-frcnn) where there is a script called create_data.py.
• Place the test.avi file and data.csv file into the data folder
• Run the create_data.py script with the following command

  python3 create_data.py

• If using Faster R-CNN, once create_data.py is finished, rename the created train.csv in the data folder to train.txt
• Run the following command to train the model/weights:
  • If using Faster R-CNN:

    python3 train_frcnn.py -o simple -p data/train.txt

    Note: for optional training arguments, read inside train_frcnn.py to find other arguments
  • If using U-Net:

    python3 train.py

• In order for create_data.py to be used on a custom data set, the data set should be layed out as followed:
  • video of file called test.avi inside of the data folder
  • csv file called data.csv inside of the data folder where each line is configured as shown:

    frame_number, top_left_point_x, top_left_point_y, bottom_right_point_x, bottom_right_point_y

• When using a custom data set, create_data.py is ran as shown below:

  python3 create_data.py -d custom

• If using Faster R-CNN, once create_data.py is finished, rename the created train.csv to train.txt
• Run the following command to train the model/weights:
  • If using Faster R-CNN:

    python3 train_frcnn.py -o simple -p data/train.txt

    Note: for optional training arguments, read inside train_frcnn.py to find other arguments
  • If using U-Net:

    python3 train.py

Note: models have to be trained before running object detection.

• To run the object detection on each algorithm, locate the start.py script in the corresponding algorithm folder and run the following command:

  python3 start.py

  • This script has multiple options:
  • set --wr to True if wanting to save a video
  • set --wf to the the number of frames wanting to save to video (otherwise it saves the whole video)
  • --width and --height can be set for a custom image size

• If the display is having issues running due to D-Bus, run the following command:

  sudo killall ibus-daemon

• Retrain Faster R-CNN model to be more accurate
  • Train with a faster network
• Implement a faster algorithm able to run object detection at a feasible frame rate
• Link pretrained models

Faster R-CNN Implementation: https://github.com/you359/Keras-FasterRCNN