关键词: 三维目标检测;多编码特征融合;并行训练, Transformer;自动驾驶

Abstract: To address the core challenges in 3D object detection under complex scenarios, such as high false negative , false positive rates, , and long training times for single-GPU models, this paper proposes a two-stage algorithm based on Multi-Encoding Feature Fusion (MFF). MFF utilizes a 3D sparse convolutional voxel encoder and a multi-scale pillar encoder to extract point cloud features hierarchically, while capturing global semantic dependencies through a Transformer attention mechanism. On the nuScenes dataset, MFF achieves 64.7% mAP and 69.7% NDS, with detection accuracy significantly outperforming comparable algorithms. To tackle the bottleneck of training efficiency, MFF adopts a distributed data parallel (DDP) framework, combined with a Ring-Allreduce communication strategy and a Warmup learning rate optimization strategy. In a 64-GPU parallel training scenario, it achieves a 6.89× speedup compared to 8-GPU training, greatly shortening the training time and effectively overcoming the resource constraints of large-scale model training. Experiments fully validate the dual breakthrough improvements of MFF in detection accuracy and training efficiency, providing a high-performance and efficient solution for environmental perception in autonomous driving.

Key words: 3D Object Detection, Multi-Encoder Feature Fusion, Parallel Training, Transformer, Autonomous Driving