Deep Velocity Estimation

Background

Speed estimation from a single dashboard camera using Deep Convolutional Networks. This is a response to the challenge posed here: https://github.com/commaai/speedchallenge

There is an associated blog post here

See the post for more details on the experiments and the associated model architectures

Results

Experiment	MSE
Deep Velocity Estimation (Grayscale Input, Frame Delta: 1)	> 10
Deep Velocity Estimation (RGB Input, Frame Delta: 1)	> 10
Deep Convolutional Network with Farneback Flow (RGB Input)	> 10
DeepER Velocity Estimation (RGB Input, Depth: 20, Frame Delta: 1)	< 1**

** Looking for someone to independently verify the performance, if you verify, please submit an issue with your results

Usage

Installation

Processing the images requires FFMPEG. See the installation guidelines here for your platform

To install the python requirements, run:

pip install -r requirements.txt

Run model training

After installing the requirements, from the root directory of the repo, run

python3 src/train.py

Verification

As referenced above, the MSE on the model is less than 1, which could be close to SOTA. I'm looking for someone to verify the results I published on the blog.

If you do want to verify, please include the following details with your verification:

• Platform
• GPU Type
• Batch Size
• Any modifications to the parameters run

The results of the run are located here

Trained parameters for the best epoch are here

References

See the blog post for references to relevant papers.