<p>Intelligent transportation systems (ITS) facilitate safer transport, improve mobility, save lives and optimize the performance of existing infrastructure. Road safety is a growing concern due to accidents occurs by different collision types. Traffic collisions play critical role in determining the impact of accident, Injuries and loss of life. This study introduces a smart hybrid framework TransEdge-GATNet: A Transformer-Edge Fusion Network with Graph Attention for ITS and metaheuristic feature engineering to enhance traffic safety in intelligent transportation systems (ITS). The proposed model integrates EdgeNet, Graph Attention Networks (GATs), and Transformer Encoders to capture fine-grained spatial patterns, dynamic inter-vehicle interactions, and long-range temporal dependencies, respectively. A metaheuristic feature engineering module, utilizing Grey Wolf Optimizer (GWO), is incorporated to select optimal input features such as speed, GPS, braking events, and timestamps, reducing overfitting. EdgeNet enhances object boundary detection from visual inputs, while GAT constructs interaction graphs based on proximity or communication data. The outputs from these modules are fused via an attention pooling layer and passed through dense layers for classification into four categories: Head-On collisions, other collision, pedestrian collisions, and hit objects. Optional components such as federated learning and pruning make the framework privacy-preserving and suitable for edge deployment.</p>