Efficient Deep Learning for Driver-Road Interactions: The Role of Convolutional Layer Specificity in Reducing Data Requirements

This study investigates the role of convolutional layers in deep convolutional neural networks for scenarios involving interactions between the users and road-interaction environment. The user-driving behaviour and road environment exhibit unique features, and applications often require training-specific models for each user-environment pair. This leads to significant data collection and computational demands, including model training, thus necessitating efficient solutions to minimise these requirements. The aim is to determine whether the convolutional layers of deep convolutional autoencoders (DCAEs) are more specific to the user or the environment. The distinction might lead to optimising training strategies that reduce resource requirements. Case studies and data collection are performed on multiple bicyclists navigating various road segments. We evaluated the specificity of convolutional layers using metrics such as training epochs, execution time, model and parameter loading times, and total training loss. The results confirmed that the patterns learned by the outer convolutional layers are predominantly user-specific, emphasising individual behaviour over road-environment factors. This user-specific pattern recognition enhances model efficiency, reduces data requirements, and improves accuracy in predicting user behaviour across varying environments. Furthermore, after analysing training strategies, we found that complete refinement provided higher accuracy and stability at the cost of longer training and loading times. By contrast, freezing layers allowed faster initialisation but might necessitate extended training in complex cases. Finally, we examined the implications of these findings for improving safety and performance in driver-road interactions.