A Lightweight Convolutional Neural Network to Predict Steering Angle for Autonomous Driving Using CARLA Simulator

End-to-end learning for autonomous driving uses a convolutional neural network (CNN) to predict the steering angle from a raw image input. Most of the solutions available for end-to-end autonomous driving are computationally too expensive, which increases the inference of autonomous driving in real time. Therefore, in this paper, CNN architecture has been trained which is lightweight and achieves comparable results to Nvidia's PilotNet. The data used to train and evaluate the network is collected from the Car Learning to Act (CARLA) simulator. To evaluate the proposed architecture, the MSE (mean squared error) is used as the performance metric. Results of the experiment shows that the proposed model is 4x lighter than Nvidia's PilotNet in term of parameters but still attains comparable results to PilotNet. The proposed model has achieved 5.1x10(-4 )MSE on testing data while PilotNet MSE was 4.7x10(-4).

Automated summary

This paper introduces a lightweight CNN architecture for end-to-end learning in autonomous driving, which is 4 times lighter in parameters compared to Nvidia's PilotNet while achieving comparable results. Trained and evaluated using data from the CARLA simulator, the proposed model achieved a lower MSE than PilotNet.