Using Keras to build a deep learning model for lane lines tracking
- This model use the basic encoder-decoder and LSTM architecture to track lines frame-by-frame.
pip install -r requirements.txt
- The training data comes from CULane.
- Using culane2tusimple to transform data format to TuSimple's format.
- 17622 sequence frames for training, 8811 for validation and 2938 for testing.
- Currently, the model is trained to track the right line of driving lane.
- Training for 250 epochs and batch size set to 64.
- Below is the architecture:
- Training loss:
- The model need the previous 5 frames as input, and it can track the lines in the 6th frame as output.
6 ground truth frames
Tracking result in 6th frames
