结合改进YOLOX与改进Second的道路车辆融合检测

刘凯; 罗素云; 魏丹

doi:10.12299/jsues.24-0068

结合改进YOLOX与改进Second的道路车辆融合检测

doi: 10.12299/jsues.24-0068

上海工程技术大学机械与汽车工程学院, 上海 201620

基金项目: 国家自然科学基金(62101314)

详细信息

作者简介:
刘凯：刘　凯（1997 − ），男，硕士生，研究方向为多传感器融合感知。E-mail：l15837070577@126.com

通讯作者:
罗素云（1975 − ），女，副教授，硕士，研究方向为无人驾驶汽车环境感知与控制。E-mail：luosuyun196@sues.edu.cn

中图分类号: TP391
计量
- 文章访问数: 14
- HTML全文浏览量: 7
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2024-03-13
- 网络出版日期: 2025-09-30
- 刊出日期: 2025-06-30

Combining improved YOLOX and improved second for road vehicle fusion detection

School of Mechanical and Automotive Engineering, Shanghai University of Engineering Science, Shanghai 201620, China

摘要

摘要: 针对主流检测算法对于道路车辆易出现漏检、误检、方向角度预测不准情况，设计一种结合改进YOLOX与改进Second的车辆融合检测算法。利用图像和点云的互补优势，采用两个子网络分别进行基于图像与基于点云的车辆检测。在图像检测中，采用基于卷积块的注意力模块、焦点损失和高效交并比损失来提升YOLOX的检测性能；在点云检测中，设计残差式稀疏卷积中间层来增强Second算法的特征表达和上下文信息关联，有效降低了车辆漏检率。构建角度多区域优化车辆角度预测，采用CLOCs的融合框架得到最终的融合检测结果。在KITTI数据集上进行算法实验，结果显示，相比原算法，新算法在简单、中等、困难目标的3D平均精度分别提升1.00%、1.38%、2.66%，检测目标旋转角度准确率得到极大提高。
- 车辆检测 /
- 残差式稀疏卷积中间层 /
- 角度多区域 /
- 融合检测
Abstract: Common detection algorithms often struggled with missed detections, false positives, and large deviations in predicted orientation angles in road vehicle detection. A fusion detection algorithm combining improved YOLOX and Second was designed. By leveraging images and point clouds, two sub-networks were employed for vehicle detection. For image detection, convolutional block attention module, focal loss, and efficient intersection over union loss function were used to improve the detection performance of existing YOLOX. For point cloud detection, a residual sparse convolutional middle layer was designed to enhance the feature expression and context information association of Second algorithm, effectively reducing the missed detection rate of vehicles. The predictive directional angles was optimized by constructing a multi-bins strategy. Experimental conducted on KITTI dataset show that the algorithm surpassed the original, with improvements of 1.00%, 1.38%, and 2.66% in 3D average precision for easy, moderate, and hard targets, respectively. The accuracy of detecting target rotation angles is also greatly improved.
- vehicle inspection /
- residual sparse convolutional middle layer /
- multi-bins /
- fusion detection

HTML全文

图 1 YOLOX网络结构图

Figure 1. YOLOX network structure diagram

下载: 全尺寸图片幻灯片

图 2 CBAM注意力机制流程图

Figure 2. CBAM attention mechanism flowchart

下载: 全尺寸图片幻灯片

图 3 改进YOLOX网络结构图

Figure 3. Improved YOLOX network structure diagram

下载: 全尺寸图片幻灯片

图 4 Second网络结构图

Figure 4. Second network structure diagram

下载: 全尺寸图片幻灯片

图 5 残差式稀疏卷积中间层

Figure 5. Residual sparse convolutional middle layer

下载: 全尺寸图片幻灯片

图 6 Multi-bins分区示意图

Figure 6. Multi-bins partition diagram

下载: 全尺寸图片幻灯片

图 7 2D与3D检测框及投影

Figure 7. 2D and 3D detection boxes and projection

下载: 全尺寸图片幻灯片

图 8 感知融合网络完整结构

Figure 8. Complete structure diagram of perception fusion network

下载: 全尺寸图片幻灯片

图 9 预测角度对比

Figure 9. Comparison of prediction angles

下载: 全尺寸图片幻灯片

图 10 融合网络train Loss曲线图

Figure 10. Fusion network train Loss curve

下载: 全尺寸图片幻灯片

图 11 融合检测结果可视化

Figure 11. Visualization of fusion detection results

下载: 全尺寸图片幻灯片

表 1 改进YOLOX消融实验

Table 1. Improved YOLOX ablation experiment

方法	CBAM	focal loss	EIoU	mAP₅₀	mAP_50:95
YOLOX	×	×	×	0.862	0.753
改进1	√	×	×	0.886	0.771
改进2	√	√	×	0.889	0.775
改进3	√	√	√	0.907	0.793

下载: 导出CSV

表 2 图像检测模型对比实验

Table 2. Comparison experiment of image detection models

方法	mAP₅₀	mAP_50:95	FPS/(帧·s⁻¹)
Faster R-CNN	0.786	0.699	4.9
YOLOV3	0.843	0.731	26.3
YOLOV5	0.854	0.745	36.7
YOLOX	0.862	0.753	36.2
改进YOLOX	0.907	0.793	35.6

下载: 导出CSV

表 3 改进Second算法消融实验表

Table 3. Improved Second ablation experiment

方法	RSCML	multi-bins	简单/%		中等/%		困难/%
方法	RSCML	multi-bins	3D mAP	aos	3D mAP	aos	3D mAP	aos
Second	×	×	88.02	94.16	78.19	88.71	77.03	84.92
改进1	√	×	89.22	93.87	80.18	88.67	77.93	84.99
改进2	√	√	89.57	97.42	80.40	90.17	78.08	89.72

下载: 导出CSV

表 4 点云检测模型3D检测精度实验对比

Table 4. Experimental comparison of 3D detection accuracy for point cloud detection models

方法	模态	阶段	简单/%		中等/%		困难/%
方法	模态	阶段	3D mAP	aos	3D mAP	aos	3D mAP	aos
VoxelNet	Lidar	Two	81.97	91.94	65.46	80.67	62.85	70.64
TANet^[15]	Lidar	Two	84.39	93.52	75.94	90.11	68.68	84.86
PointPillars	Lidar	One	85.86	95.13	75.75	91.79	70.65	86.51
Second	Lidar	One	88.02	94.16	78.19	88.71	77.03	84.92
ACDet^[16]	Lidar	One	88.47	96.07	78.85	92.36	73.86	89.18
改进Second	Lidar	One	89.57	97.42	80.40	90.17	78.08	89.72

下载: 导出CSV

表 5 融合算法精度对比实验

Table 5. Comparison experiment of fusion algorithm accuracy

方法	模态	简单/%		中等/%		困难/%
方法	模态	3D mAP	aos	3D mAP	aos	3D mAP	aos
Second	Lidar	88.02	94.16	78.19	88.71	77.03	84.92
MV3D	Lidar + Camera	86.55	91.32	78.10	86.53	76.67	84.71
F-PointNet	Lidar + Camera	88.02	95.85	81.16	92.17	75.33	85.42
AVOD-FPN	Lidar + Camera	86.80	94.98	81.79	89.22	77.70	82.14
PFF3D	Lidar + Camera	83.11	94.86	80.26	91.06	75.43	86.28
CLOCs	Lidar + Camera	91.07	96.77	81.65	93.66	77.97	87.31
Ours(One)	Lidar + Camera	91.56	97.00	81.67	93.67	78.65	87.36
Ours(Two)	Lidar + Camera	92.07	99.21	83.03	93.53	80.63	92.97

下载: 导出CSV

参考文献(18)

[1]	BALTRUŠAITIS T, AHUJA C, MORENCY L P. Multimodal machine learning: a survey and taxonomy[J] . IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(2): 423 − 443. doi: 10.1109/TPAMI.2018.2798607
[2]	刘伟. 基于激光雷达和机器视觉的智能车前方障碍物检测研究[D] . 哈尔滨: 哈尔滨理工大学, 2019.
[3]	郑少武, 李巍华, 胡坚耀. 基于激光点云与图像信息融合的交通环境车辆检测[J] . 仪器仪表学报, 2022, 40(12): 143 − 151.
[4]	陆峰, 徐友春, 李永乐, 等. 基于信息融合的智能车障碍物检测方法[J] . 计算机应用, 2017, 37(S2): 115 − 119.
[5]	CHEN X Z, MA H M, WAN J, et al. Multi-view 3D object detection network for autonomous driving[C] //Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu: IEEE, 2017: 6526−6534.
[6]	KU J, MOZIFIAN M, LEE J, et al. Joint 3D proposal generation and object detection from view aggre-gation[C] //Proceedings of 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems. Madrid: IEEE, 2018: 1−8.
[7]	YIN T W, ZHOU X Y, KRÄHENBÜHL P. Multimodal virtual point 3D detection[C] //Proceedings of the 35th International Conference on Neural Information Processing Systems. [S. l. ] : Curran Associates Inc. , 2021: 1261.
[8]	杨飞, 朱株, 龚小谨, 等. 基于三维激光雷达的动态障碍实时检测与跟踪[J] . 浙江大学学报(工学版), 2012, 46(9): 1565 − 1571. doi: 10.3785/j.issn.1008-973X.2012.09.003
[9]	GE Z, LIU S T, WANG F, et al. YOLOX: exceeding Yolo series in 2021[EB/OL] . (2021-07-18)[2024-01-22] . https://doi.org/10.48550/arXiv.2107.08430.
[10]	YAN Y, MAO Y X, LI B. SECOND: sparsely embedded convolutional detection[J] . Sensors, 2018, 18(10): 3337. doi: 10.3390/s18103337
[11]	PANG S, MORRIS D, RADHA H. CLOCs: camera-LiDAR object candidates fusion for 3D object detection[C] //Proceedings of 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems. Las Vegas: IEEE, 2020: 10386−10393.
[12]	REDMON J, FARHADI A. YOLOv3: An incremental improvement[EB/OL] . (2018-04-08)[2024-01-22] . https://doi.org/10.48550/arXiv.1804.02767.
[13]	刘凯, 罗素云. 基于改进YOLOX-S的交通标志识别[J] . 电子测量技术, 2023, 46(1): 112 − 119.
[14]	ZHOU Y, TUZEL O. VoxelNet: end-to-end learning for point cloud based 3D object detection[C] //Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 4490−4499.
[15]	LIU Z, ZHAO X, HUANG T T, et al. TANet: robust 3D object detection from point clouds with triple attention[C] //Proceedings of the 34th AAAI Conference on Artificial Intelligence. New York: AAAI, 2020: 11677−11684.
[16]	XU J L, WANG G J, ZHANG X, et al. ACDet: attentive cross-view fusion for LiDAR-based 3D object detection[C] //Proceedings of 2022 International Conference on 3D Vision. Prague: IEEE, 2022: 74−83.
[17]	QI C R, LIU W, WU C X, et al. Frustum pointnets for 3D object detection from RGB-D data[C] //Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 918−927.
[18]	WEN L H, JO K H. Fast and accurate 3D object detection for lidar-camera-based autonomous vehicles using one shared voxel-based backbone[J] . IEEE Access, 2021, 9: 22080 − 22089. doi: 10.1109/ACCESS.2021.3055491