xrosliang/vehicle-speed-estimation

Vehicle Speed Estimation from Video using Deep Learning and Optical Flow in PyTorch.

vehicle-speed-estimation

If you want to know more about this project checkout my medium post.

Requirements

pip3 install -r requirements.txt

How to use

Pre-trained model

You can simply use the model that I trained before. It is under models. Use the inference.ipynb to load the model and run an inference.

Train yourself

You can train the network (EfficientNet) to predict the speed of a vehicle using optical flow. If you want to train yourself, you will need to create the optical flow images first and save them as .npy files in a directory of your choice. You can do this here: SharifElfouly/opical-flow-estimation-with-RAFT.

Results

If you are interested on how well the model performs, watch this validation video on YouTube.

Another approach

You can also just stack 2 frames together so you have 6 channels for each input and feed that to a conv net. This is what I did in train_cnn_2frames.ipynb.