Speed-Estimation-Using-Homography-Transformation-and-Regression-Network

This work includes homography transformation which is change monocular view into bird eye view and calculate each of vehicel speed in the videos using regression network.

Requirements

To run the training or testing step, you must installed library as follow:

1. Tensorflow-gpu = 1.14.0
2. keras = 2.2.4
3. opencv-contrib-python = 3.2.0.8

How to Training

1. Download and generate the datasets The datasets based on the carla dataset and to generate datasets directly using "generate_raw_carla_van_dataset.py" if you want to use directly without do preprocessing first can be downloaded in here
2. Convert dataset into generator If you already download the dataset, run convert_raw_carla_van_to_tfrecords.py to create one file contains parameter to compute homography matrix
3. Training process After we put dataset into generator, datasets ready to train. Run "train_carla_van_horizon_vpz.py" to start training process and set epoch, learning_rate, and batch_size as you desired.
4. Finish. You got checkpoint and choose the better checkpoint based on minimum loss will be used as pretrained model in new images.

How to Testing

1. To run the testing, you can choose single images or video files run "run_image.py" or "run_video.py", respectively. for the single image can be seen in "images" folder, while the videos files can be downloaded in here
2. Wait until testing finish, and you got the bird eye view results.