An improved method called CS-Voxel-RCNN is proposed to address the issue of insufficient detection accuracy of Voxel-RCNN algorithm in detecting small distant targets and occluded targets. Firstly, by introducing three data augmentation methods: random order, random dropout, and random noise, the diversity of training samples is enriched, thereby enhancing the robustness of the model. Secondly, by integrating CBAM in the 2D backbone network and utilizing channel attention mechanism and spatial attention mechanism, multi-scale features are processed in more detail, optimizing the feature fusion effect. Finally, by adding a DIoU loss branch, the original loss function is improved, emphasizing the distance information between the target bounding boxes, thereby improving the accuracy of the target bounding box regression task. Comparative experiments with some classic 3D object detection algorithms on the KITTI dataset are conducted. The results show that the newly proposed algorithm has significantly improved performance, compared with the original Voxel RCNN algorithm, with improvements of 2.91 percentage and 0.87 percentage for pedestrians and cyclists, respectively. The effectiveness of each improvement module is verified through ablation experiments. This series of improvement methods achieve positive results in improving the practicality and accuracy of 3D object detection in real scenes.
Voxel-RCNN[24]基于此问题引入了创新性的Voxel RoI Pooling模块,作为一种高效的区域池化方法.该模块采用体素聚合(voxel query)方法,允许网络更有效地检索感兴趣区域(region of interest, RoI)周围的体素信息.同时,算法结合了加速的PointNet++网络,旨在提升目标识别优化阶段中的局部体素特征提取效率.这种双重的优化策略不仅有效地提高了局部特征的表达能力,同时也显著缩短了整体网络的运行时间,因此Voxel-RCNN在车辆识别任务中表现出色.然而,由于其缺乏对不完整点云信息的处理方法,因此在检测远处目标以及受到遮挡的目标时算法性能有待提升.为解决这一问题,本文从原模型的数据增强模块、2D骨干网络以及损失函数这三个方面进行了系统性的改进,以期在更广泛的场景中提高检测性能.在改进的过程中,充分考虑了点云数据的特殊性,力求优化模型对不同目标的鲁棒性和泛化能力,从而使得新模型能够更好地适应复杂环境下的目标检测任务.
如图1所示,Voxel-RCNN由三部分组成,分别是3D骨干网络、2D骨干网络+RPN以及Voxel RoI Pooling(体素感兴趣区域池化)+检测头.其特别设计了一个名为voxel RoI pooling的模块,该模块能够直接从体素特征中提取RoI特征,以便进行进一步处理.实验结果表明,在KITTI数据集和Waymo数据集上,Voxel-RCNN相较于现有的基于体素的方法,不仅能保持实时帧处理速率(即在NVIDIA RTX 2080Ti GPU上达到25 FPS的速率),而且提供了更高的检测精度.这一研究成果为基于体素的目标检测方法的发展提供了有力的支持,具有显著的实际应用潜力.
Voxel RoI Pooling是算法的第二阶段,旨在对第一阶段提出的候选区域进行进一步的优化.一般的优化手段是采用RoI Pooling等技术,将所有候选区域分割成固定维度,以方便后续的并行计算.接着,3D体素特征被映射到候选区域内,过滤掉非候选区域的特征,重新生成新的候选区域3D体素.每个候选区域被看作一个独立的体素空间,进行特征提取,最终微调第一阶段的初步候选区域.
Voxel-RCNN算法提出了Voxel RoI Pooling方法,其运用了voxel query方法和加速的PointNet++方法,以提高体素特征映射和优化的效率.其具体流程如下:
为了提高算法对于小目标和受阻挡目标的检测效果,本文在原基础上引入了一个新分支,对预测边界框的位置回归进行进一步细化.采用了DIoU(distance intersection over union,距离交并比)损失,该损失函数考虑了边界框之间的距离信息,通过提高3D预测边界框和真实边界框的交并比来提高回归任务的效果.
交并比(intersection over union,IoU)是目标检测中常用的一种性能评价指标.它衡量了预测边界框(bounding box, bbox)与真实边界框(ground truth box,gt box)之间的重叠程度,通常用于衡量检测算法的准确性.公式如下:
实验采用平均精度(mean average precision, mAP)来衡量模型效果.mAP是目标检测任务中常用的性能评估指标之一,综合考虑了模型在不同类别上的精度,并通过计算各类别的平均值来提供对整体性能的度量.通过在不同阈值下计算P和R,并绘制P-R曲线,可以得到每个类别的精度信息.mAP就是对这个曲线下面积的平均值,即平均精度.计算公式为
WANGY D, TIANY L, LIG Q,et al. 3D object detection based on convolutional neural networks:a survey[J]. Pattern Recognition and Artificial Intelligence,2021,34(12):1103-1119.(in Chinese)
XIEZ X, LIM H .A survey on machine learning in recognition of 3D object based on point cloud[J]. Periodical of Ocean University of China, 2021, 51(6): 125-130.(in Chinese)
[5]
FERNANDESD, SILVAA, NÉVOAR,et al .Point-cloud based 3D object detection and classification methods for self-driving applications:a survey and taxonomy[J]. Information Fusion,2021,68: 161-191.
[6]
CHARLESR Q, HAOS, MOK C,et al .PointNet:deep learning on point sets for 3D classification and segmentation[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu,HI,USA. IEEE,2017:77-85.
[7]
QIC R, YIL, SUH,et al .PointNet++[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach,California,USA. ACM,2017:5105-5114.
[8]
LIY Y, BUR, SUNM C, et al. PointCNN: Convolution on x-transformed points[C]//NIPS’18: Proceedings of the 32nd International Conference on Neural Information Processing Systems. Montréal, Canada. 2018.
[9]
VORAS, LANGA H, HELOUB,et al .PointPainting:sequential fusion for 3D object detection[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle,WA,USA. IEEE,2020:4604-4612.
[10]
YANGZ T, SUNY N, LIUS,et al .3DSSD:point-based 3D single stage object detector[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle, WA,USA. IEEE,2020:11040-11048.
[11]
ZHANGY, ZHOUZ X, DAVIDP,et al .PolarNet:an improved grid representation for online LiDAR point clouds semantic segmentation[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle,WA,USA. IEEE,2020:9598-9607.
[12]
THOMASH, QIC R, DESCHAUDJ E,et al .KPConv:flexible and deformable convolution for point clouds[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul,Korea (South). IEEE, 2019: 6411-6420.
[13]
LUON, YUH Q, HUOZ F,et al .KVGCN:a KNN searching and VLAD combined graph convolutional network for point cloud segmentation[J].Remote Sensing,2021,13(5):1003.
[14]
WANGL, HUANGY C, HOUY L,et al. Graph attention convolution for point cloud semantic segmentation[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA,USA. IEEE, 2019:10288-10297.
[15]
ZHOUY, TUZELO .VoxelNet:end-to-end learning for point cloud based 3D object detection[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City,UT,USA. IEEE,2018:4490-4499.
LANGA H, VORAS, CAESARH,et al .PointPillars:fast encoders for object detection from point clouds[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach,CA,USA. IEEE,2019:12689-12697.
[18]
LIJ Y, LUOC X, YANGX D .PillarNeXt:rethinking network designs for 3D object detection in LiDAR point clouds[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Vancouver,BC,Canada. IEEE,2023:17567-17576.
[19]
SSINGHA R. PointPillars++: An encoder for 3-D object detection and classification from point clouds[D]. North Carolina State University, 2021.
[20]
YANGZ T, SUNY N, LIUS,et al .STD:sparse-to-dense 3D object detector for point cloud[C]//2019 IEEE/CVF International Conference on Computer Vision (ICCV). Seoul,Korea (South). IEEE, 2019: 1951-1960.
[21]
WUF Z, BAOL C, CHENY J,et al .MVF-Net:multi-view 3D face morphable model regression[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach,CA,USA. IEEE,2019:959-968.
[22]
KUANGH W, WANGB, ANJ P,et al. Voxel-FPN:multi-scale voxel feature aggregation for 3D object detection from LIDAR point clouds[J].Sensors, 2020, 20(3): 704.
[23]
WUZ R, SONGS R, KHOSLAA,et al. 3D ShapeNets:a deep representation for volumetric shapes[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston,MA,USA. IEEE,2015:1912-1920.
[24]
ZHUH Q, DENGJ J, ZHANGY,et al .VPFNet:improving 3D object detection with virtual point based LiDAR and stereo data fusion[J].IEEE Transactions on Multimedia,2022,25:5291-5304.
[25]
HEKIMOGLUA, SCHMIDTM, MARCOS-RAMIROA .Monocular 3D object detection with LiDAR guided semi supervised active learning[C]//2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Waikoloa, HI, USA. IEEE, 2024: 2335-2344.
[26]
DENGJ J, SHIS S, LIP W, et al. Voxel R-CNN:towards high performance voxel-based 3D object detection[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(2):1201-1209.
[27]
SHIS S, GUOC X, JIANGL, et al. PV-RCNN:point-voxel feature set abstraction for 3D object detection[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Seattle,WA,USA. IEEE,2020:10529-10538.
[28]
RENS Q, HEK M, GIRSHICKR,et al .Faster R-CNN:towards real-time object detection with region proposal networks[J].IEEE Transactions on Pattern Analysis and Machine Intelligence,2017, 39(6): 1137-1149.
[29]
WOO S, PARKJ, LEEJ Y,et al .CBAM:convolutional block attention module[C]// Computer Vision-ECCV 2018. Cham:Springer International Publishing, 2018: 3-19.
[30]
REZATOFIGHIH, TSOIN, GWAKJ,et al. Generalized intersection over union:a metric and a loss for bounding box regression[C]//2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA,USA.IEEE,2019: 658-666.
[31]
GEIGERA, LENZP, URTASUNR .Are we ready for autonomous driving?The KITTI vision benchmark suite[C]//2012 IEEE Conference on Computer Vision and Pattern Recognition. Providence,RI,USA. IEEE, 2012: 3354-3361.
基金资助
国家重点研发计划资助项目(2021YFB2501800)
National Key Research and Development Program of China(2021YFB2501800)