Deep learning has revolutionized multi-sensor fusion, enabling safe and efficient autonomous vehicle operations in IoT-enabled smart cities. This study introduces a novel framework employing a Transformer-based architecture, specifically designed to process heterogeneous sensor data from vehicles, infrastructure, and IoT devices. The framework incorporates Temporal And Spatial Attention Mechanisms, allowing for accurate real-time decision-making. Utilizing TensorFlow as the primary development tool, the model integrates data from LiDAR, radar, cameras, and V2X communication to enhance situational awareness and navigation precision. Experimental results demonstrate significant improvements in obstacle detection, path planning, and collision avoidance, achieving a 20% increase in operational efficiency compared to existing methods. The proposed approach not only ensures the safety and reliability of autonomous vehicles but also lays a foundation for scalable deployments in smart cities.