Enhancing lane-keeping technologies with optimised convolutional neural networks for steering angle prediction

Lane departure incidents, often caused by human negligence, contribute significantly to road fatalities. Lane-keeping systems can mitigate these risks by accurately predicting steering angles under varying road conditions and external disturbances. Traditional approaches struggle with feature extraction due to environmental complexities. This study proposes an enhanced Convolutional Neural Network (CNN) architecture, optimised for real-time steering angle prediction. A hybrid metaheuristic method, combining the Bat Algorithm and Particle Swarm Optimisation (PSO), is employed to fine-tune CNN hyper-parameters and architecture. The two-stage optimisation first tunes learning rate, batch size, and activation function via the Bat Algorithm, followed by PSO for optimising layer configurations. The model, trained on the Udacity dataset and 25,000 additional samples, uses Adagrad (learning rate 0.0052) and ELU activation. Real-world testing across diverse conditions showed improved generalisation and reduced Mean Squared Error (MSE) from 0.00329 to 0.00119, confirming the effectiveness of metaheuristic optimisation in enhancing CNN-based lane-keeping performance.