An Improved Enhanced 3D Object Detection Algorithm

Abstract

With the rapid evolution of autonomous driving and intelligent robotics, point-cloud-based 3D object detection has emerged as a critical research frontier. However, existing voxel-based methods often suffer from semantic-geometric misalignment during multi-level feature fusion and spatial quantization bias inherent in local feature pooling. To address these challenges, this paper proposes an enhanced 3D object detection framework termed AC-Voxel R-CNN. First, an Adaptive Gated Fusion Module (AGFM) is designed to dynamically allocate weights between shallow geometric details and deep semantic abstractions through a lightweight channel attention mechanism, facilitating context-aware “on-demand” feature integration. Second, a Voxel Point Centroid Aggregation (VPCA) strategy is introduced, which substitutes fixed voxel centers with physical point centroids and utilizes neighboring relative position residuals for weighted modulation to faithfully capture local geometric topologies. Experimental results on the large-scale KITTI and Waymo Open datasets demonstrate that the proposed method significantly enhances detection precision, particularly for occluded and sparse targets, achieving a Moderate 3D AP of 85.63% on KITTI. The findings conclude that the synergistic integration of adaptive feature fusion and centroid-based encoding substantially improves the robustness and localization fidelity of 3D object detection in complex driving environments.