Abstract

This paper examines how to combine image and point cloud data in optimizing depth reliability, dynamic perception, enhancement of fusion features, robustness against sensor failures, and efficiency in various 3D object detection datasets on autonomous driving. This paper generally critically examines existing approaches, especially the Lift-Splat framework, and designs some new solutions to overcome current limitations. The research work focuses on the improvement of the depth accuracy, dynamic scene perception, and robustness of 3D detection systems by a series of hypotheses. It incorporates advanced fusion techniques, spatiotemporal deformable attention mechanisms, and optimized depth estimation ranges for achieving significant improvements in detection performance. Results from comprehensive experiments validate the effectiveness of these innovations across multiple datasets, thereby positioning the proposed method as a key advancement in the field of autonomous driving perception.