重庆理工大学学报2025，Vol.39Issue(17)：31-37,7.DOI:10.3969/j.issn.1674-8425(z).2025.09.004
融合注意力机制与残差网络的三维目标检测
Fusion of attention mechanisms and residual networks for 3D object detection
摘要
Abstract
To address high misdetection rate for sparse points and inability to fully extract object features in PointPillars algorithm,an improved PointPillars algorithm is proposed,which integrates the spatial attention mechanism and the improved backbone network.First,the spatial attention mechanism is incorporated at the feature network layer after the point cloud is encoded as a pseudo image that is processed by 2D convolution.Then,the original conventional 2D convolutional backbone network is replaced by residual network for feature extraction from 2D pseudo-images.Finally,the algorithm is validated on KITTI dataset and on a campus environment.Compared with the baseline PointPillars,the algorithm improves the average detection accuracy for vehicles,pedestrians,and cyclists by 2.32％,1.97％,and 3.27％respectively,at medium detection difficulty in KITTI.It achieves a single-frame inference time of approximately 102 ms on an NVIDIA Jetson Orin AGX device.Experimental results show the improved algorithm has the potential to be deployed on lightweight devices and maintains the speed of real-time detection,optimizing unmanned vehicle technologies.关键词
三维目标检测/自动驾驶/PointPillars/伪图像/KITTIKey words
3D object detection/autonomous driving/PointPillars/pseudo image/KITTI分类
信息技术与安全科学引用本文复制引用
但远宏,程东,季勇..融合注意力机制与残差网络的三维目标检测[J].重庆理工大学学报,2025,39(17):31-37,7.基金项目
重庆市科委重点攻关计划项目(2021CCB03) （2021CCB03）
