Skip to content

Latest commit

 

History

History

dl-with-carla

README.md

Key concepts

1. Make data for model

To obtain driving data, we used Carla, an open-source simulator. We employed two main methods to acquire driving data from Carla:

  • Automatic Control

    Using the autonomous driving code implemented by Carla, we save five types of data while the vehicle is in motion: speed, steering, throttle, brake, and lane invasion. These data types are stored according to the VSS standard. Depending on the need, this data can be formatted into CSV or input into a virtual CAN.

  • Manual Control

    You can manually drive according to Carla's physics engine, although it is quite challenging. Like in automatic control, the same five data types are saved.

2. LSTM Learning Model

lstmmodel.png

We utilized an LSTM model to calculate driving scores. We chose this model because of its strength in handling time-series data, which is crucial since our CAN data is time-dependent. For more details on the model, please refer to lstm-training.py.

When the five types of data form a sequence of 100, a corresponding score is assigned. With 300,000 CAN data points, a total of 300 data sequences have been trained. Although it's not a vast amount of data, it should be sufficient to represent your driving score!"