基于自适应微调的露天矿山低空无人机旋转目标检测

高铭; 陈鑫; 蒋烁; 胡满江; 秦洪懋; 边有钢

doi:10.19818/j.cnki.1671-1637.2026.158

基于自适应微调的露天矿山低空无人机旋转目标检测

doi: 10.19818/j.cnki.1671-1637.2026.158

湖南大学机械与运载工程学院, 湖南长沙 410082

基金项目:

国家自然科学基金项目 52472429

详细信息

作者简介:
高铭（1991-），男，山东乳山人，副研究员，工学博士，博士后，E-mail：gaoming@hnu.edu.cn

通讯作者:
秦洪懋（1984-），男，江苏盐城人，副教授，博士生导师，工学博士，博士后，E-mail：qinhongmao@vip.sina.com

中图分类号: U495
计量
- 文章访问数: 13
- HTML全文浏览量: 10
- PDF下载量: 4
- 被引次数: 0
出版历程
- 收稿日期: 2025-10-10
- 录用日期: 2026-01-23
- 修回日期: 2025-12-22
- 刊出日期: 2026-03-28

Rotated object detection using low-altitude UAVs for open-pit mines with adaptive fine-tuning

College of Mechanical and Vehicle Engineering, Hunan University, Changsha 410082, Hunan, China

Funds:

National Natural Science Foundation of China 52472429

More Information

Corresponding author: QIN Hong-mao, associate professor, PhD, E-mail: qinhongmao@vip.sina.com

Article Text (Baidu Translation)

摘要

摘要: 为实现低空立体交通运输系统中露天矿山全场景实时视觉感知，提出基于自适应微调的无人机旋转目标检测方法（AFTDet）。针对矿用车辆从无人机视角观测时姿态变化显著的问题，设计自适应空间回归损失函数以优化角度学习并提高高纵横比目标的旋转边界框回归精度，提出微调非极大值抑制算法以利用重叠检测框的空间信息并通过定位参数差异的加权融合提升预测精度，构建包含乘用车辆、小型挖掘机、装载机和自卸卡车共4 540个旋转标注样本的露天矿山旋转目标检测数据集（MineR），最终在公开遥感数据集DOTAv1.0和自建MineR数据集上对AFTDet进行验证。结果表明：在公开遥感数据集DOTAv1.0上，AFTDet取得78.61%的平均计算精度AP₅₀和55.45%的平均计算精度AP₇₅，较基准模型RTMDet-R-m分别提升0.47%和1.80%；在自建MineR数据集上取得76.25%的AP₅₀和44.38%的AP₇₅，较基准模型分别提升1.06%和3.62%；消融试验中自适应标签分配策略使AP₅₀提升0.99%、AP₇₅提升2.50%，微调非极大值抑制使AP₇₅进一步提升1.09%，检测速度达50.5帧·s^-1，参数量维持2.467×10⁷不变。自适应微调检测方法显著提升了旋转目标的姿态估计性能，尤其改善了大纵横比矿用车辆的检测召回率，在保持实时检测能力的同时为低空立体交通运输系统的无人机视觉感知提供了有效技术支撑，促进了露天矿山智能监控与调度系统的发展。
- 低空立体交通运输系统 /
- 旋转目标检测 /
- 自适应微调 /
- 露天矿山 /
- 视觉感知 /
- 无人机检测
Abstract: To achieve full-scene real-time visual perception for open-pit mines in low-altitude three-dimensional transportation systems, an adaptive fine-tuning detection method (AFTDet) was proposed for unmanned aerial vehicle (UAV)-based rotated object detection. To address the significant pose variation of mining vehicles observed from UAV perspectives, an adaptive spatial regression loss function was designed to optimize angle learning and improve the rotated bounding box regression accuracy for high-aspect-ratio targets. A fine-tuned non-maximum suppression algorithm was proposed to leverage spatial information from overlapping detection boxes and enhance prediction accuracy through weighted fusion of localization parameter differences. The open-pit mine rotated object detection dataset (MineR) was constructed, comprising 4 540 rotated annotated samples covering passenger vehicles, small excavators, loaders, and dump trucks. AFTDet was validated on both the public remote sensing dataset DOTAv1.0 and the self-built MineR dataset. The results demonstrate that AFTDet achieves 78.61% AP₅₀ and 55.45% AP₇₅ on DOTAv1.0, representing improvements of 0.47% and 1.80% respectively over the baseline model RTMDet-R-m. On the MineR dataset, it achieves 76.25% AP₅₀ and 44.38% AP₇₅, with improvements of 1.06% and 3.62% over the baseline. Ablation studies indicate that the adaptive label assignment strategy improves AP₅₀ by 0.99% and AP₇₅ by 2.50%, while the fine-tuned non-maximum suppression further improves AP₇₅ by 1.09%. The detection speed reaches 50.5 frames·s^-1 with parameters maintained at 2.467×10⁷. The adaptive fine-tuning detection method significantly enhances pose estimation performance for rotated objects, particularly improving the detection recall of large-aspect-ratio mining vehicles, providing effective technical support for UAV visual perception in low-altitude three-dimensional transportation systems while maintaining real-time detection capabilities. In addition, it promotes the development of intelligent monitoring and scheduling systems for open-pit mines.
- low-altitude three-dimensional transportation system /
- rotated object detection /
- adaptive fine-tuning /
- open-pit mine /
- visual perception /
- UAV detection

HTML全文

图 1 AFTDet的模型结构

Figure 1. Model structure of AFTDet

下载: 全尺寸图片幻灯片

图 2 标签分配流程示例

Figure 2. Example of the label assignment process

下载: 全尺寸图片幻灯片

图 3 $ \boldsymbol{f}\left(\boldsymbol{\theta }\right) $与$ {\boldsymbol{\theta }}_{\mathrm{g}\mathrm{t}}-{\boldsymbol{\theta }}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{d}} $的关系曲线

Figure 3. Relationship curve between $ \boldsymbol{f}\left(\boldsymbol{\theta }\right) $ and $ {\boldsymbol{\theta }}_{\mathbf{g}\mathbf{t}}-{\boldsymbol{\theta }}_{\mathbf{p}\mathbf{r}\mathbf{e}\mathbf{d}} $

下载: 全尺寸图片幻灯片

图 4 角度不连续性问题和类正方形问题

Figure 4. Angular discontinuity problems and square-like problems

下载: 全尺寸图片幻灯片

图 5 保留框和高效框示例

Figure 5. Examples of resbox and high-efficiency box

下载: 全尺寸图片幻灯片

图 6 α_i随|θ_i-θ_m|变化的曲线

Figure 6. Curve of α_i as a function of |θ_i-θ_m|

下载: 全尺寸图片幻灯片

图 7 MineR数据集的类别示例

Figure 7. Category example of the MineR dataset

下载: 全尺寸图片幻灯片

图 8 各类别样本数量统计

Figure 8. Statistics on the number of samples of each category

下载: 全尺寸图片幻灯片

图 9 MineR数据集的可视化效果对比

Figure 9. Comparison of visualizations of the MineR dataset

下载: 全尺寸图片幻灯片

图 10 DOTAv1.0数据集可视化效果对比

Figure 10. Comparison of visualizations of the DOTAv1.0 dataset

下载: 全尺寸图片幻灯片

表 1 RTMDet-R各尺寸模型对比

Table 1. Comparison of RTMDet-R models of each size

算法模型	平均精度/%	参数量/10⁶
RTMDet-R-tiny	75.36	4.88
RTMDet-R-s	76.93	8.86
RTMDet-R-m	78.24	24.67
RTMDet-R-l	78.85	52.27

下载: 导出CSV

表 2 图 2中检测框的空间回归代价值

Table 2. Spatial regression value of the detection frame in Fig. 2

检测框	C_reg	C_ADPreg（β=0.15）
红色框	1.23	1.50
蓝色框	1.28	1.29

下载: 导出CSV

表 3 MineR数据集上的消融试验结果

Table 3. Ablation study results on the MineR dataset

算法模型	FTNMS	ADPreg	AP₅₀/%	AP₇₅/%	参数量/10⁷
RTMDet-R-m	×	×	75.45	42.83	2.467
	√	×	75.52	43.30	2.467
	×	√	76.20	43.90	2.467
AFTDet	√	√	76.25	44.38	2.467
注：AP₅₀和AP₇₅分别为在IoU阈值为0.50和0.75时计算的平均精度，下同。

下载: 导出CSV

表 4 算法对检测速度的影响

Table 4. Influence of the algorithm on the detection speed

算法模型	延迟/ms
RTMDet-R-m	19.2
AFTDet	19.8

下载: 导出CSV

表 5 FTNMS和NMS在不同IoU阈值下的效果对比

Table 5. Comparison of the effects of FTNMS and NMS at different IoU thresholds

算法模型	后处理算法	平均计算精度/%
算法模型	后处理算法	AP₅₀	AP₆₅	AP₇₅	AP₈₅
RTMDet-R-m	NMS	75.45	64.20	42.83	14.16
RTMDet-R-m	FTNMS	75.52	64.54	43.30	14.71

下载: 导出CSV

表 6 ADPreg对各类别召回率的影响

Table 6. Effect of ADPreg on the recall rate of each category

算法模型	ADPreg	召回率/%
算法模型	ADPreg	乘用车辆	小型挖掘机	装载机	自卸卡车
RTMDet-R-m	×	85.4	91.0	85.3	92.1
RTMDet-R-m	√	87.1	92.8	90.7	91.7

下载: 导出CSV

表 7 先进算法在DOTAv1.0数据集上的对比试验结果

Table 7. Comparative experiment results of advanced algorithms on DOTAv1.0 dataset

算法模型	AP₅₀/%	帧率/（帧·s^-1）	参数量/10⁷
KLD^[38]	77.36	25.8	4.190
Oriented-RepPoint^[39]	77.63	23.5	3.661
CFA^[40]	76.67	24.1	3.661
CSL^[41]	76.21	24.5	3.735
Gliding-Vertex^[42]	75.02	17.4	6.013
PSC^[29]	71.83	26.1	3.193
S2ANet^[43]	76.11	17.4	3.624
RTMDet-R-m^[16]	78.24	49.5	2.467
RTMDet-R-l^[16]	78.85	29.1	5.227
AFTDet	78.61	45.6	2.467

下载: 导出CSV

表 8 算法在DOTAv1.0数据集上的消融试验结果

Table 8. Ablation study results of the algorithm on the DOTAv1.0 dataset

算法模型	FTNMS	ADPreg	AP₅₀/%	AP₇₅/%
RTMDet-R-m	×	×	78.24	54.47
	√	×	78.29	54.86
	×	√	78.60	55.13
AFTDet	√	√	78.61	55.45

下载: 导出CSV

参考文献(43)

[1]	李宏刚, 王云鹏, 廖亚萍, 等. 无人驾驶矿用运输车辆感知及控制方法[J]. 北京航空航天大学学报, 2019, 45(11): 2335-2344. LI Hong-gang, WANG Yun-peng, LIAO Ya-ping, et al. Perception and control method of driverless mining vehicle [J]. Journal of Beijing University of Aeronautics and Astronautics, 2019, 45(11): 2335-2344.
[2]	REN S Q, HE K M, GIRSHICK R, 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. doi: 10.1109/TPAMI.2016.2577031
[3]	DAI J F, LI Y, HE K M, et al. R-FCN: Object detection via region-based fully convolutional networks[C]//LEE D D, SUGIYAMA M, LUXBURG U V, et al. Advances in Neural Information Processing Systems 29. Red Hook: Curran Associates, 2016: 379-387.
[4]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: Single shot MultiBox detector[C]// LEIBE B, MATAS J, SEBE N, et al. Computer Vision – ECCV 2016. Cham: Springer, 2016: 21-37.
[5]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection [C]// IEEE. 2017 IEEE International Conference on Computer Vision (ICCV). New York: IEEE, 2017: 2999-3007.
[6]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection [C]// IEEE. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2016: 779-788.
[7]	HE Y, JIN L S, GUO B C, et al. Density-based road segmentation algorithm for point cloud collected by roadside LiDAR [J]. Automotive Innovation, 2023, 6(1): 116-130.
[8]	张涛, 路向阳, 李雷, 等. 露天矿山运输无人驾驶关键技术与标准[J]. 控制与信息技术, 2019(2): 13-19. ZHANG Tao, LU Xiang-yang, LI Lei, et al. Key technologies and standards of autonomous driving system applied in surface mines [J]. Control and Information Technology, 2019(2): 13-19.
[9]	FU Y, GAO M, XIE G T, et al. Density-aware U-Net for unstructured environment dust segmentation [J]. IEEE Sensors Journal, 2024, 24(6): 8210-8226. doi: 10.1109/JSEN.2024.3355388
[10]	彭仲仁, 刘晓锋, 张立业, 等. 无人飞机在交通信息采集中的研究进展和展望[J]. 交通运输工程学报, 2012, 12(6): 119-126. doi: 10.19818/j.cnki.1671-1637.2012.06.018 PENG Zhong-ren, LIU Xiao-feng, ZHANG Li-ye, et al. Research progress and prospect of UAV applications in transportation information collection [J]. Journal of Traffic and Transportation Engineering, 2012, 12(6): 119-126. doi: 10.19818/j.cnki.1671-1637.2012.06.018
[11]	ZHOU Y, YANG X, ZHANG G F, et al. MMRotate: A rotated object detection benchmark using PyTorch [C]// ACM. Proceedings of the 30th ACM International Conference on Multimedia. New York: ACM, 2022: 7331-7334.
[12]	XIE X X, CHENG G, RAO C F, et al. Oriented object detection via contextual dependence mining and penalty-incentive allocation [J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5618010.
[13]	ZHAO J Q, DING Z Y, ZHOU Y, et al. OrientedFormer: An end-to-end transformer-based oriented object detector in remote sensing images [J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5640816.
[14]	ZHOU M H, LI T Y, QIAO C F, et al. DMM: Disparity-guided multispectral mamba for oriented object detection in remote sensing [J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 5404913.
[15]	NEUBECK A, VAN GOOL L. Efficient non-maximum suppression [C]// IEEE. 18th International Conference on Pattern Recognition (ICPR'06). New York: IEEE, 2006: 850-855.
[16]	LYU C Q, ZHANG W W, HUANG H A, et al. RTMDet: An empirical study of designing real-time object detectors [EB/OL]. (2022-10-16). https://arxiv.org/abs/2212.07784.
[17]	LI G F, CHI X Y, QU X D. Depth estimation based on monocular camera sensors in autonomous vehicles: A self-supervised learning approach [J]. Automotive Innovation, 2023, 6(2): 268-280. doi: 10.1007/s42154-023-00223-6
[18]	王仁炎, 陆占国, 胡振涛, 等. 基于改进YOLOv3的露天矿卡车目标检测方法[J]. 矿业研究与开发, 2024, 44(2): 164-169. WANG Ren-yan, LU Zhan-guo, HU Zhen-tao, et al. Target detection method of truck in open-pit mine based on improved YOLOv3 algorithm [J]. Mining Research and Development, 2024, 44(2): 164-169.
[19]	秦晓辉, 黄启东, 常灯祥, 等. 基于改进YOLOv5的露天矿山目标检测方法[J]. 湖南大学学报(自然科学版), 2023, 50(2): 23-30. QIN Xiao-hui, HUANG Qi-dong, CHANG Deng-xiang, et al. Object detection method in open-pit mine based on improved YOLOv5 [J]. Journal of Hunan University (Natural Sciences), 2023, 50(2): 23-30.
[20]	岳伟, 林军, 康高强, 等. 基于改进DeepSORT的路侧感知方法在露天矿山中的应用[J]. 控制与信息技术, 2023(3): 89-94. YUE Wei, LIN Jun, KANG Gao-qiang, et al. Application of roadside perception method based on improved DeepSORT in surface mine [J]. Control and Information Technology, 2023(3): 89-94.
[21]	WOJKE N, BEWLEY A, PAULUS D. Simple online and realtime tracking with a deep association metric [C]// IEEE. 2017 IEEE International Conference on Image Processing (ICIP). New York: IEEE, 2017: 3645-3649.
[22]	阮顺领, 张回国, 顾清华, 等. 基于双目视觉的露天矿无人车前障碍检测研究[J]. 煤炭学报, 2024, 49(增2): 1285-1294. RUAN Shun-ling, ZHANG Hui-guo, GU Qing-hua, et al. Research on the detection of obstacles in front of unmanned vehicles in opencast mines based on binocular vision [J]. Journal of China Coal Society, 2024, 49(S2): 1285-1294.
[23]	陈婷, 姚大春, 高涛, 等. 基于PReNet和YOLOv4融合的雨天交通目标检测网络[J]. 交通运输工程学报, 2022, 22(3): 225-237. doi: 10.19818/j.cnki.1671-1637.2022.03.018 CHEN Ting, YAO Da-chun, GAO Tao, et al. A fused network based on PReNet and YOLOv4 for traffic object detection in rainy environment [J]. Journal of Traffic and Transportation Engineering, 2022, 22(3): 225-237. doi: 10.19818/j.cnki.1671-1637.2022.03.018
[24]	陈龙, 司译文, 田滨, 等. 基于3D LiDAR的矿山无人驾驶车行驶边界检测[J]. 煤炭学报, 2020, 45(6): 2140-2146. CHEN Long, SI Yi-wen, TIAN Bin, et al. Boundary detection of mine drivable area based on 3D LiDAR [J]. Journal of China Coal Society, 2020, 45(6): 2140-2146.
[25]	孟德将, 田滨, 蔡峰, 等. 面向无人驾驶矿车的露天矿山道路坡度实时检测方法[J]. 测绘学报, 2021, 50(11): 1628-1638. MENG De-jiang, TIAN Bin, CAI Feng, et al. Road slope real-time detection for unmanned truck in surface mine [J]. Acta Geodaetica et Cartographica Sinica, 2021, 50(11): 1628-1638.
[26]	YANG X, YANG J R, YAN J C, et al. SCRDet: Towards more robust detection for small, cluttered and rotated objects [C]// IEEE. 2019 IEEE/CVF International Conference on Computer Vision (ICCV). New York: IEEE, 2019: 8231-8240.
[27]	YANG X, YAN J C. On the arbitrary-oriented object detection: Classification based approaches revisited [J]. International Journal of Computer Vision, 2022, 130(5): 1340-1365. doi: 10.1007/s11263-022-01593-w
[28]	YU Y, DA F P. Phase-shifting coder: Predicting accurate orientation in oriented object detection [C]// IEEE. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2023: 13354-13363.
[29]	MING Q, ZHOU Z Q, MIAO L J, et al. Dynamic anchor learning for arbitrary-oriented object detection [J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35(3): 2355-2363. doi: 10.1609/aaai.v35i3.16336
[30]	ZHOU X Y, YAO C, WEN H, et al. EAST: An efficient and accurate scene text detector [C]// IEEE. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2017: 2642-2651.
[31]	ZHU Y X, DU J, WU X Q. Adaptive period embedding for representing oriented objects in aerial images [J]. IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(10): 7247-7257. doi: 10.1109/TGRS.2020.2981203
[32]	LI Z H, HOU B, WU Z T, et al. FCOSR: A simple anchor-free rotated detector for aerial object detection [J]. Remote Sensing, 2023, 15(23): 5499. doi: 10.3390/rs15235499
[33]	GE Z, LIU S T, WANG F, et al. YOLOX: Exceeding YOLO series in 2021 [EB/OL]. (2021-08-06). https://arxiv.org/abs/2107.08430.
[34]	DING X H, ZHANG X Y, HAN J G, et al. Scaling up your kernels to 31 × 31: Revisiting large kernel design in CNNs [C]// IEEE. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2022: 11953-11965.
[35]	XIA G S, BAI X, DING J, et al. DOTA: A large-scale dataset for object detection in aerial images [C]// IEEE. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New York: IEEE, 2018: 3974-3983.
[36]	BODLA N, SINGH B, CHELLAPPA R, et al. Soft-NMS: Improving object detection with one line of code [C]// IEEE. 2017 IEEE International Conference on Computer Vision (ICCV). New York: IEEE, 2017: 5562-5570.
[37]	NING C C, ZHOU H J, SONG Y, et al. Inception single shot MultiBox detector for object detection [C]// IEEE. 2017 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). New York: IEEE, 2017: 549-554.
[38]	YANG X, YANG X J, YANG J R, et al. Learning high-precision bounding box for rotated object detection via Kullback-Leibler divergence[C]// RANZATO M, BEYGEIR H, DAUPHIN Y, et al. Advances in Neural Information Processing Systems 34. Red Hook: Curran Associates, 2021: 18381-18394.
[39]	LI W T, CHEN Y J, HU K X, et al. Oriented RepPoints for aerial object detection [C]// IEEE. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2022: 1819-1828.
[40]	GUO Z H, LIU C, ZHANG X S, et al. Beyond bounding-box: Convex-hull feature adaptation for oriented and densely packed object detection [C]// IEEE. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New York: IEEE, 2021: 8788-8797.
[41]	YANG X, YAN J C. Arbitrary-oriented object detection with circular smooth label [C]// VEDALDI A, BISCHOF H, BROX T, et al. Computer Vision – ECCV 2020. Cham: Springer International Publishing, 2020: 677-694.
[42]	XU Y C, FU M T, WANG Q M, et al. Gliding vertex on the horizontal bounding box for multi-oriented object detection [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 43(4): 1452-1459. doi: 10.1109/TPAMI.2020.2974745
[43]	HAN J M, DING J, LI J, et al. Align deep features for oriented object detection [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5602511.