基于改进YOLOX的无人机航拍图像密集小目标车辆检测

doi:10.13229/j.cnki.jdxbgxb.20230779

吉林大学学报(工学版) ›› 2025, Vol. 55 ›› Issue (4): 1307-1318.doi: 10.13229/j.cnki.jdxbgxb.20230779

• 交通运输工程·土木工程 • 上一篇下一篇

基于改进YOLOX的无人机航拍图像密集小目标车辆检测

张河山^1,²(),范梦伟¹,谭鑫¹,郑展骥^1,²,寇立明³,徐进^1,²()

^1.重庆交通大学交通运输学院，重庆 400074
^2.重庆交通大学山区复杂道路环境“人-车-路”协同与安全重庆市重点实验室，重庆 400074
^3.重庆市交通规划研究院，重庆 400074

收稿日期:2023-07-25 出版日期:2025-04-01 发布日期:2025-06-19
通讯作者: 徐进 E-mail:hszhang@cqjtu.edu.cn;yhnl_996699@163.com
作者简介:张河山（1988-），男，讲师，博士.研究方向：视频图像处理，道路交通安全.E-mail： hszhang@cqjtu.edu.cn
基金资助:
教育部人文社会科学研究青年基金项目(24YJCZH412);国家自然科学基金项目(52172340);重庆市教育委员会青年项目(KJQN202200710);重庆市博士后科学基金项目(CSTB2022NSCQ-BHX0731);重庆交通大学研究生科研创新项目(CYS23498)

Dense small object vehicle detection in UAV aerial images using improved YOLOX

He-shan ZHANG^1,²(),Meng-wei FAN¹,Xin TAN¹,Zhan-ji ZHENG^1,²,Li-ming KOU³,Jin XU^1,²()

^1.School of Traffic & Transportation，Chongqing Jiaotong University，Chongqing 400074，China
^2.Chongqing Key Laboratory of "Human-Vehicle-Road" Cooperation and Safety for Mountain Complex Environment，Chongqing Jiaotong University，Chongqing 400074，China
^3.Chongqing Transportation Planning and Research Institute，Chongqing 400074，China

Received:2023-07-25 Online:2025-04-01 Published:2025-06-19
Contact: Jin XU E-mail:hszhang@cqjtu.edu.cn;yhnl_996699@163.com

摘要/Abstract

摘要：

针对无人机航拍视角下对小目标的检测仍存在漏检现象严重、检测精度低等问题，提出一种改进的YOLOX网络，用于无人机航拍图像的检测。为了增强网络的特征学习能力，在特征融合部分引入自适应空间特征融合（ASFF）模块，并在网络的颈部（Neck）嵌入坐标注意力机制（CA）。为了加强网络对正样本的学习，将二元交叉熵损失函数替换为变焦距损失函数。实验结果表明：改进后的YOLOX网络具有更好的检测效能，其mAP_@50和mAP_{@50_95}分别达到了91.50%和79.65%。在多种交通场景下的可视化结果表明：相较于其他算法，优化后的网络具有更低的漏检率以及更高的检测精度，能够胜任小目标车辆的检测任务，可为高空视角下的车辆多目标跟踪应用提供参考。

关键词: 交通运输系统工程, 小目标车辆检测, 损失函数, 坐标注意力机制, 自适应空间特征融合, YOLOX

Abstract:

Aiming at the issues of severe missed detections and low detection accuracy for small targets in the perspective of drone aerial photography， an improved YOLOX network is proposed for the detection of drone aerial images. To enhance the feature learning ability of the network， the ASFF module is introduced in the feature fusion part， and the CA mechanism is embedded in the neck of the network. To enhance the network's learning of positive samples， the binary cross-entropy loss function is replaced with the varifocal loss function. Experimental results show that the improved YOLOX network has better detection efficiency， and its mAP_@50 reaches 91.50% and mAP_{@50_95} reached 79.65%. The visualization results in various traffic scenarios show that compared with other algorithms， the optimized network has a lower missed detection rate and higher detection accuracy， which can be competent for the detection task of small target vehicles， and can provide a reference for vehicle multi-target tracking applications from a high-altitude perspective.

Key words: engineering of communications and transportation system, small target vehicle detection, loss function, coordinate attention mechanism, adaptive spatial feature fusion, YOLOX

中图分类号:

U491

张河山,范梦伟,谭鑫,郑展骥,寇立明,徐进. 基于改进YOLOX的无人机航拍图像密集小目标车辆检测[J]. 吉林大学学报(工学版), 2025, 55(4): 1307-1318.

He-shan ZHANG,Meng-wei FAN,Xin TAN,Zhan-ji ZHENG,Li-ming KOU,Jin XU. Dense small object vehicle detection in UAV aerial images using improved YOLOX[J]. Journal of Jilin University(Engineering and Technology Edition), 2025, 55(4): 1307-1318.

图/表 16

图1

图2

图3

图4

表1

图5

图6

表2

网络检测性能的评价指标"

指标	公式	意义
Precision	$P = T P T P + F P$	在所有检测到的目标中检测到正确的概率
Recall	$R = T P T P + F N$	在所有阳性样本中正确识别的概率
F₁_score	$F 1 = 2 × R × P R + P$	准确率和召回率的调和均值
AP	$A P = ∫ P R d R$	同一召回率下精度的平均值
mAP	$m A P = 1 C ∑ K C A P K$	数据集中所有平均精度之和的平均值

表2

表3

图7

图8

表4

图9

图10

表5

图11

参考文献 29

1	Liu C, Yang D G, Tang L, et al. A lightweight object detector based on spatial-coordinate self-attention for UAV aerial images[J]. Remote Sensing, 2023, 15(1): No.83.
2	Liu W, Wang M, Zhang S, et al. Research on vehicle target detection technology based on UAV aerial images[C]∥IEEE International Conference on Mechatronics and Automation (ICMA), Guilin,China, 2022: 412-416.
3	李旭, 宋世奇, 殷晓晴. 基于目标空间分布特征的无人机航拍车辆实时检测技术研究[J]. 中国公路学报, 2022, 35(12): 193-204.
	Li Xu, Song Shi-qi, Yin Xiao-qing. Research on real-time detection technology of UAV aerial photography vehicles based on target spatial distribution characteristics[J]. China Journal of Highway and Transport, 2022, 35(12): 193-204.
4	Girshic R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]∥IEEE Conference on Computer Vision and Pattern Recognition,Columbus, USA, 2014: 580-587.
5	Girshick R. Fast R-CNN[C]∥IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 2015: 1440-1448.
6	Ren S, He K, 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.
7	Ge Z, Liu S, Wang F, et al. YOLOX: exceeding YOLO series in 2021[DB/OL]. [2023-06-27].
8	Wang Y, Li J, Chen Z, et al. Ships' small target detection based on the CBAM-YOLOX algorithm[J]. Journal of Marine Science and Engineering, 2022, 10(12): No. 2013.
9	Shen C, Ma C, Gao W. Multiple attention mechanism enhanced YOLOX for remote sensing object detection[J]. Sensors, 2023, 23(3): No. 1261.
10	赵振兵, 马迪雅, 石颖, 等. 基于改进YOLOX的变电站仪表外观缺陷检测算法[J]. 图学学报, 2023(5):937-946.
	Zhao Zhen-bing, Ma Di-ya, Shi Ying, et al. Appearance defect detection algorithm for substation instrument based on improved YOLOX[J]. Journal of Graphics, 2023(5): 937-946.
11	Yi K, Luo K, Chen T, et al. An improved YOLOX model and domain transfer strategy for nighttime pedestrian and vehicle detection[J]. Applied Sciences,2022, 12(23): No.12476.
12	Xiong C, Yu A, Yuan S, et al. Vehicle detection algorithm based on lightweight YOLOX[J]. Signal, Image and Video Processing, 2023, 17(5): 1793-1800.
13	Luo Q, Wang J, Gao M, et al. G-YOLOX: a lightweight network for detecting vehicle types[J]. Journal of Sensors, 2022, 2022: 1-10.
14	黄键, 徐伟峰, 苏攀, 等. 基于YOLOX-S的车窗状态识别算法[J]. 吉林大学学报: 理学版, 2023, 61(4):875-882.
	Huang Jian, Xu Wei-feng, Su Pan, et al. Window state recognition algorithm based on YOLOX-S[J].Journal of Jilin University (Science Edition), 2023, 61(4):875-882.
15	Han S, Yoo J, Kwon S. Real-time vehicle-detection method in bird-view unmanned-aerial-vehicle imagery[J]. Sensors, 2019, 19(18): No.3958.
16	刘军黎, 刘晓锋, 邱洁, 等. YOLOX-IM: 一种无人机航拍视频的轻量化交通参数提取模型[J]. 国外电子测量技术, 2023, 42(1): 159-169.
	Liu Jun-li, Liu Xiao-feng, Qiu Jie,et al. YOLOX-IM:a lightweight traffic parameter extraction model for UAV aerial video[J].Foreign Electronic Measurement Technology, 2023, 42(1): 159-169.
17	张稀柳, 张晓玲, 何敏军. 基于改进YOLOX-s的车辆检测方法研究[J]. 系统仿真学报, 2024(2): 487-496.
	Zhang Xi-liu, Zhang Xiao-ling, He Min-jun. Research on vehicle detection method based on improved YOLOX-s[J]. Journal of System Simulation, 2024(2):487-496.
18	Peng H, Tan X. Improved YOLOX's anchor-free SAR image ship target detection[J].IEEE Access, 2022, 10: 70001-70015.
19	Luo X, Wu Y, Zhao L. YOLOD: a target detection method for UAV aerial imagery[J]. Remote Sensing, 2022, 14(14): No.3240.
20	Hu J, Shen L, Albanie S, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011-2023.
21	Woo S, Park J, Lee J Y, et al. CBAM: convolutional block attention module[C]∥European Conference on Computer Vision, Cham,Germany, 2018: 3-19.
22	Hou Q, Zhou D, Feng J. Coordinate attention for efficient mobile network design[C]∥IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Nashville, USA,2021: 13708-13717.
23	Zhang F, Cao W, Wang S, et al. Improved YOLOv4 recognition algorithm for pitaya based on coordinate attention and combinational convolution[J]. Frontiers in Plant Science, 2022, 13:No.1030021.
24	Xuan W, Gao J Z, Hou B J, et al.A lightweight modified YOLOX network using coordinate attention mechanism for PCB surface defect detection[J]. IEEE Sensors Journal, 2022, 22(21): 20910-20920.
25	Song Z, Huang X, Ji C, et al. Deformable YOLOX: detection and rust warning method of transmission line connection fittings based on image processing technology[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-21.
26	Song C Y, Zhang F, Li J S, et al. Detection of maize tassels for UAV remote sensing image with an improved YOLOX model[J]. Journal of Integrative Agriculture, 2023, 22 (6): 1671-1683.
27	Qu Z, Shang X, Xia S F, et al. A method of single-shot target detection with multi-scale feature fusion and feature enhancement[J]. IET Image Processing, 2022, 16(6): 1752-1763.
28	Wang G, Liu Z, Sun H, et al. Yolox-BTFPN:an anchor-free conveyor belt damage detector with a biased feature extraction network[J]. Measurement, 2022, 200: No.111675.
29	Kumar S, Jain A, Rani S, et al. Deep neuralnetwork based vehicle detection and classification of aerial images[J]. Intelligent Automation & Soft Computing, 2022, 34(1): 119-131.

Metrics

Viewed

Full text

Abstract

Cited

Shared

Discussed

实验平台	型号参数
CPU	12 vCPU Intel（R） Xeon（R） Platinum 8255C CPU @2.50 GHz
GPU	GeForce RTX 3080（10 GB）
操作系统	Windows 10
框架	Pytorch
编程环境	Python

参数	值
训练轮次	200
类别	3
批次	8
初始学习率	0
动量	0.9
权重衰减	5×10^-4
检测阈值	0.35
解冻训练	30
非极大值抑制	0.65

网络	mAP_@50/%	mAP_{@50_95}/%	Parameters/M
Original YOLOX	89.3	78.44	8.94
ASFF	89.7	78.83	14.38
ASFF+CA	89.9	79.35	14.40
Improved YOLOX	91.5	79.65	14.40

网络	mAP/%	Parameters/M	Precision/%	Recall/%	F₁_score	Gflops/G
Original YOLOX	89.3	8.94	95.89	93.69	93.97	26.8
Faster R-CNN	62.8	137.00	46.20	72.47	56.33	185.1
SSD	88.6	26.29	78.38	93.75	85.33	140.9
YOLOv5	89.3	7.10	84.91	91.33	81.00	16.5
Improved YOLOX	91.5	14.40	95.33	94.94	94.13	35.2

[1]	潘义勇,徐家聪,尤逸文,全勇俊. 网约车出行需求影响因素多尺度空间异质性分析[J]. 吉林大学学报(工学版), 2025, 55(5): 1567-1575.
[2]	卢凯明,陈艳艳,仝瑶,张健,李永行,罗莹. 数据驱动的信号交叉口排队尾车驶离状态预测[J]. 吉林大学学报(工学版), 2025, 55(4): 1275-1286.
[3]	周成栋,宋菲,赵小梅,姚俊杰. 基于多模式双动态演化的拥堵收费模型[J]. 吉林大学学报(工学版), 2025, 55(4): 1319-1327.
[4]	秦严严,肖腾飞,罗钦中,王宝杰. 雾天高速公路车辆跟驰安全分析与控制策略[J]. 吉林大学学报(工学版), 2025, 55(4): 1241-1249.
[5]	宋现敏,湛天舒,李海涛,刘博,张云翔. 考虑用户成本和泊位利用率的停车预约分配模型[J]. 吉林大学学报(工学版), 2025, 55(4): 1287-1297.
[6]	郭祎,魏书威,姜涛. 基于区位势能和多源数据的城市客运交通规划算法[J]. 吉林大学学报(工学版), 2025, 55(4): 1328-1335.
[7]	潘义勇,徐翔宇. 数据不平衡的MobileViT网络交通事故严重程度预测模型[J]. 吉林大学学报(工学版), 2025, 55(3): 947-953.
[8]	陈永恒,杨家伟,孙经宇. 借道左转交叉口的网联左转车辆最佳轨迹控制[J]. 吉林大学学报(工学版), 2025, 55(2): 614-622.
[9]	陈发城,鲁光泉,林庆峰,张浩东,马社强,刘德志,宋会军. 有条件自动驾驶下驾驶人接管行为综述[J]. 吉林大学学报(工学版), 2025, 55(2): 419-433.
[10]	何永明,冯佳,魏堃,万亚楠. 超高速公路曲线路段车辆制动侧滑影响因素分析[J]. 吉林大学学报(工学版), 2025, 55(2): 591-602.
[11]	吴娇蓉,刘旭东. 不同住房类型空间单元的建成环境对通勤方式选择的影响分析[J]. 吉林大学学报(工学版), 2025, 55(2): 554-565.
[12]	王长帅,徐铖铖,任卫林,彭畅,佟昊. 自动驾驶接管过程中驾驶能力恢复状态对交通流振荡特性的影响[J]. 吉林大学学报(工学版), 2025, 55(1): 150-161.
[13]	张曦,库少平. 基于生成对抗网络的人脸超分辨率重建方法[J]. 吉林大学学报(工学版), 2025, 55(1): 333-338.
[14]	张娜,陈峰,王剑坡,朱亚迪. 基于时空序列相似性的城轨乘客出行模式识别[J]. 吉林大学学报(工学版), 2024, 54(9): 2588-2599.
[15]	周锡浈,宫贺,李敦敦,季彦婕,严杰. 建成环境对路内停车泊位使用率的非线性影响模型[J]. 吉林大学学报(工学版), 2024, 54(9): 2520-2530.

基于改进YOLOX的无人机航拍图像密集小目标车辆检测

Dense small object vehicle detection in UAV aerial images using improved YOLOX

RICH HTML

PDF (PC)

摘要/Abstract

引用本文

使用本文

图/表 16

参考文献 29

相关文章 15

Metrics

本文评价

推荐阅读 10