Machine learning models for self-driving cars. This repo is a WIP and will be updated regularly.
Requires:
The specific task here is to predict driving speed from a dashboard-mounted camera; there's tons of this data open-sourced by the folks at comma.ai here. I experimented with several flavors of deep nets:
- Static CNN -- The model is based off of NVIDIA's "End-to-end Self-driving" model. It's static because each video frame is assumed independent by the model, ignoring temporal dependencies. The resulting model is in models/cnn_prediction.py.
- CNN + LSTM -- In order to incorporate temporal info into the static CNN, I experimented with several architectures discussed in "Beyond Short Snippets: Deep Networks for Video Classification". The resulting model is in models/cnn_lstm_prediction.py.
- Optical flow -- Tracking feature points with sparse optical flow implicitly encodes temporal dependencies between frames. (This model will be pushed soon!)
Check out speed_prediction.md for more.
![](training_screenshot.png?raw=true "Screenshot from training" =200x)
A few places to find open-source data to play with: