Abstract

The study presents an investigation on the integration of image and point cloud data towards optimized depth reliability, dynamic perception, fusion feature enhancement, robustness against sensor failures, and effectiveness across diverse 3D object detection datasets in autonomous driving systems. A critical review of existing methods is presented with an emphasis on the Lift-Splat (LS) framework and novel solutions for addressing current limitations. Through a number of hypotheses, the research aims at improving the accuracy of depth estimates, dynamic perception of scenes, and robustness in 3D detection systems. The proposed method advances fusion techniques such as spatiotemporal deformable attention mechanisms and depth estimation ranges within optimized parameters; this achieves some significant improvements for detection performance. Extensive experiments on various datasets have confirmed the improvements of all these innovations, which poses the proposed method as one of the most important advancements in autonomous driving perception.