Abstract:
Vehicular Internet of Things (V-IoT) applications have seen a rise in recent times, increasing the need for robust machine learning (ML) models trained on diverse and high-quality datasets. Curating high-quality data for real-world vehicular environments is often challenging due to high costs, privacy concerns, and data imbalance issues. Therefore, data augmentation becomes a crucial strategy. However, conventional data augmentation methods primarily focus on pixel-level transformations. They often fail to introduce meaningful semantic variations, limiting their effectiveness in real-world scenarios. To address this gap, we propose a novel Diffusion-guided Semantic Data Augmentation Method (DSDAM) that utilizes diffusion models to generate augmented data that still preserves key vehicular attributes. Our method incorporates a structured prompt generation mechanism to introduce controlled variations in environmental conditions such as background, lighting, road surface, and weather attributes. We evaluate our method using a convolutional neural network trained on a pedestrian dataset. The model achieves an accuracy of 93%, surpassing the conventional augmentation techniques. The results demonstrate that our method enhances model generalization, reduces overfitting, and improves the robustness of ML models in V-IoT applications. This work contributes to the development of data augmentation techniques that are useful for more reliable and efficient intelligent transportation systems.