Advancing Spatiotemporal Fusion for Enhanced Perception in Dynamic Environments

The increasing demand for robust perception systems in autonomous driving has highlighted the limitations of traditional single-modal approaches, particularly in noisy and dynamic urban environments. This paper introduces the DFBEV framework, which utilizes innovative spatiotemporal sampling and fusion strategies to enhance 3D perception. By integrating multi-frame data, DFBEV effectively captures dynamic changes and maintains target consistency. Key components, such as Pillar queries and an image denoising module, significantly improve detection accuracy and computational efficiency. Despite relying on image data, DFBEV provides a strong foundation for future enhancements, including multimodal sensor integration and vehicle-to-infrastructure (V2I) collaborations. Through comprehensive ablation studies, we demonstrate the framework's potential in transforming next-generation intelligent transportation systems.