yolov3-hart

Code implementation of the work "Real-time adaptive object detection and tracking for autonomous vehicles"

Adaptive Stage Switch Model (YOLOv3 + HART) [2D/3D] | Keras/Tensorflow v2.0

If you consider using this code, consult our paper on the references tab for additional information on our model and proposed adaptive system.

Quick start

1. Clone this file

   $ git clone https://github.com/Hffmann/yolov3-hart.git
   

2. You are supposed to install some dependencies before getting out hands with these codes.

   $ cd yolov3-hart
   $ pip install -r ./data/requirements.txt
   

3. The yolo.h5 file can be generated using the YAD2K repository here: https://github.com/allanzelener/YAD2K

   Steps how to do it on windows:

   • Clone the above repository to your computer

   • Download the yolo.weights file from here: http://pjreddie.com/media/files/yolo.weights

   • Download the yolo.cfg file form here: https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolo.cfg

   • Copy the downloaded weights and cfg files to the YAD2K master directory

   • Run the following command on the shell and the h5 file will be generated.

   $ python yad2k.py yolo.cfg yolo.weights model_data/yolo.h5 
   

   • The YOLOv3 model training can be performed as described on https://github.com/YunYang1994/tensorflow-yolov3

4. Download the HART weights from http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/bvlc_alexnet.npy and put the file in the tracker/checkpoints folder.

   • The HART approach training can be performed as described on https://github.com/akosiorek/hart

5. [3D-Deepbox] Download the weights file from https://drive.google.com/file/d/1yAFCmdSEz2nbYgU5LJNXExtsS0Gvt66U/view?usp=sharing or follow the training process from https://github.com/smallcorgi/3D-Deepbox and put it in the model_data folder.

6. After downloading or training all weights on your own dataset, run the testing script :

   $ python main.py
   

   • Parse the arguments from the file main.py upon the main command execution to adjust the model to your use.

   • Opt. Download the video input from https://drive.google.com/file/d/1enfe_tcHLkAJGL0Y-PPNw4Hb6FYg8_om/view?usp=sharing and put it in the video_input folder

References

- Real-time adaptive object detection and tracking for autonomous vehicles

-tensorflow-yolov3

-hart

- 3D-Deepbox