基于Tv-SECOND的自动驾驶场景下的3D目标检测
3D Object Detection in Autonomous Driving Scenarios Based on Tv-SECOND
魏海跃 1杨奎河 1毕江峰1
作者信息
- 1. 河北科技大学信息科学与工程学院,河北石家庄 050018
- 折叠
摘要
针对自动驾驶场景中复杂环境下的3D目标检测任务,特别是远距离和遮挡条件下,为提高模型的检测准确率,在SECOND的基础上提出了Tv-SECOND两阶段算法.该算法提出一种基于Transformer架构的提案框特征提取模块,并在传统体素特征编码基础上提出可变形的体素特征编码模块.在KITT1数据集上进行测试,结果显示,所提出的算法相比SECOND在远距离和遮挡严重的情况下分别提高了 7.49％、9.72％.同时与其他先进的两阶段方法相比,检测精度有不同程度的提升,证明了 Tv-SECOND算法的有效性.新算法能够建立特征之间的依赖关系,聚合周边广域的上下文信息,增强模型的学习推理能力,有效提升了模型在远距离和遮挡的情况下的检测性能.
Abstract
In response to the issue of 3D object detection tasks in complex environments of autonomous driving scenarios,particu-larly under long-distance and occlusion conditions,a two-stage Tv-SECOND algorithm is proposed based on SECOND to enhance detec-tion accuracy.This algorithm introduces a proposal feature extraction module with a Transformer architecture,and additionally,proposes a deformable voxel feature encoding module based on traditional voxel feature encoding.Tested on the KITTI dataset,results show that compared to SECOND,our proposed algorithm improves the detection performance by 7.49％and 9.72％respectively in long-distance and severe occlusion situations.Moreover,it exhibits varying degrees of improvement in detection accuracy compared to other advanced two-stage methods,demonstrating the effectiveness of the Tv-SECOND algorithm.This new algorithm can establish dependencies among features,aggregate wide-ranging contextual information from surrounding areas,and enhance learning and inference capabilities.It effec-tively improves the detection performance of the model in long-distance and occluded scenarios.
关键词
自动驾驶/3D目标检测/Transformer/SECONDKey words
autonomous driving/3D object detection/Transformer/SECOND引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据