Optimizing CNN Architectures for Steering Angle Prediction for Self-Driving Vehicles in Unstructured Roads: A Comparative Study of Activation Functions and Model Complexity

This study investigates convolutional neural network (CNN) architectures for predicting steering angles in self-driving vehicles navigating unstructured roads, using road-facing image data. Two complementary experiments are conducted. First, the impact of three activation functions—Exponential Linear Unit (ELU), Rectified Linear Unit (ReLU), and Leaky ReLU—is evaluated on a baseline CNN model. Trained on 14,754 images and validated on 3,585 images, the model with ELU activation achieves the lowest validation mean squared error (MSE) compared to ReLU and Leaky ReLU, demonstrating superior convergence and generalization. Second, the effect of model complexity is examined using ELU activation across simple, moderate, and complex CNN variants. Results indicate that the moderately complex architecture yields the best performance, outperforming both simpler (underfitting) and more complex (overfitting) models in terms of validation MSE. These findings underscore the critical role of appropriate activation functions and balanced network depth in achieving robust, efficient steering prediction for autonomous driving in challenging, unstructured environments.