The field of autonomous driving perception is witnessing significant advancements, driven by the development of innovative models and techniques that enhance the accuracy and robustness of driving scene understanding. One of the key directions in this field is the integration of multi-modal information, including 3D representations, semantic occupancy prediction, and camera-radar fusion, to improve the complexity and fidelity of driving scene reconstruction. Researchers are also exploring new approaches to preserve fine-grained geometry details and improve the generation of tiny objects. Furthermore, the use of diffusion models and attention mechanisms is becoming increasingly popular in this field, allowing for more effective integration of cross-modal information and improved performance in downstream tasks such as BEV segmentation and 3D object detection. Noteworthy papers in this area include DualDiff, which proposes a dual-branch conditional diffusion model for multi-view driving scene generation, and OccCylindrical, which presents a multi-modal fusion approach with cylindrical representation for 3D semantic occupancy prediction. CaRaFFusion is also notable for its novel framework that enhances 2D semantic segmentation with camera-radar point cloud fusion and zero-shot image inpainting.