基于教师-学生模型的点云目标检测算法
Point Cloud Object Detection Algorithm Based on Teacher-student Model
文峰 1石明泽 1刘思萌 1殷向阳1
作者信息
- 1. 沈阳理工大学 信息科学与工程学院,沈阳 110159
- 折叠
摘要
针对现有点云目标检测算法对室外自动驾驶场景中遮挡物体检测精度低的问题,提出一种基于知识蒸馏和注意力增强的点云目标检测算法.以CIA-SSD模型为基础,设计了一种基于教师-学生模型的密集特征生成模块,提出基于交并比匹配策略的密集数据生成方法,将稀疏特征转换为密集特征.密集特征生成模块位于学生模型中,学生模型在教师模型生成的软目标监督训练下,推断出完整的密集目标特征,实现目标特征的补全;在教师模型中设计空间注意力和通道注意力机制,增强密集目标点云,提升特征图的质量.在KITTI数据集上的验证实验结果表明:与SE-SSD模型和CIA-SSD模型相比,本文提出的算法保持了单阶段目标检测速度的优势,同时明显提升了检测精度.
Abstract
A point cloud object detection algorithm based on knowledge distillation and attention enhancement is proposed to address the issue of low detection accuracy for occluded objects in out-door autonomous driving scenarios.Building upon the CIA-SSD algorithm model,a dense feature generation module based on a teacher-student model is designed.A dense data generation method u-sing intersection over union matching strategy is proposed to convert sparse features into dense fea-tures.The dense feature generation module is located within the student model,which infers com-plete dense object features under the soft target supervision generated by the teacher model,achie-ving the completion of object features.Spatial attention and channel attention mechanisms are de-signed in the teacher model to enhance the dense point cloud of the objects and improve the quality of the feature maps.Validation results on the KITTI dataset demonstrate that the proposed method significantly improves detection accuracy while maintaining the speed advantage of single-stage ob-ject detection compared to the SE-SSD model and CIA-SSD model.
关键词
点云目标检测/CIA-SSD/教师-学生模型/密集特征模块/注意力机制/交并比匹配策略Key words
point cloud object detection/CIA-SSD/teacher-student model/dense feature generation module/attention mechanism/intersection over union matching strategy引用本文复制引用
基金项目
国家重点研发计划"社会治理与智慧社会科技支撑"重点专项(2022YFC3302502)
出版年
2024
NETL
NSTL
万方数据