LaneTracker

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.

Installation

pip install -r requirements.txt

Data

• The training data comes from CULane.
• Using culane2tusimple to transform data format to TuSimple's format.

Training

• 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:

Test model

• 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