Research on Target Detection Algorithm Based on Improved BEVFusion

作者：huichung yu, Rui Zhang, Haijing Hou, Peng Xia · 发表于：2025 IEEE 5th International Conference on Power, Electronics and Computer Applications (ICPECA) · 年份：2025 · DOI：10.1109/ICPECA63937.2025.10928764

In fields such as autonomous driving and intelligent transportation, 3D target detection technology is crucial for environment perception and decision making. In this paper, we propose a target detection algorithm based on improved BEVFusion, which utilizes MobileNetV1 and PointN et to extract image and point cloud features, respectively, and combines with a dynamic feature fusion module to achieve deep multimodal fusion. In this paper, the knowledge distillation technique is introduced, and ResNet-101 is used as the teacher network to improve the feature extraction capability and detection accuracy of the student model by passing high-quality feature information to MobileNetV1. On the nuScenes dataset, the detection accuracy of this paper's method reaches 0.724 in the “Car” category, which is higher than 0.554 in PointPillars and 0.548 in BEVFormer; in the “Pedestrian” category, the accuracy is 0.730, which is close to the highest value of 0.732. In addition, the MobileNetV1 in this paper is optimized by knowledge distillation to achieve the nuScenes dataset. Optimized by knowledge distillation, it reaches NDS 0.581 and mAP 0.526, which demonstrates its robustness and practicality in complex scenarios.