Efficient Multimodal 3D Object Detection via Dynamic Feature Fusion of LiDAR and Camera Data

Current 3D detection methods, whether single-modal or multimodal, face notable limitations. Single-modal detectors, using either camera or LiDAR, struggle with spatial accuracy and object differentiation due to insufficient depth information or difficulty distinguishing semantically similar objects. Existing multimodal fusion techniques, while improving performance, often suffer from high computational costs, false positives, and complex architectures, especially when utilizing anchor-based pipelines. To address these challenges, we propose an efficient pointwise fusion method that directly extracts point features from enhanced RGB images and fuses them with corresponding point cloud features, preserving essential spatial and semantic information. This fused data is then processed through a three-dimensional neural network, significantly improving inference speed and detection performance. Our framework is designed for multi-class 3D object detection, leveraging the complementary strengths of LiDAR and camera data without the need for multiple backbones or complex synchronization steps. Extensive experiments on the KITTI benchmark demonstrate that the proposed method outperforms state-of-the-art LiDAR-camera fusion techniques, achieving 92.5% AP for 3D detection and 95.41% AP for BEV detection, making it particularly suitable for autonomous driving systems. These results highlight the effectiveness of the proposed fusion strategy in balancing accuracy, computational efficiency, and robustness in complex 3D environments.