基于YOLOv5的无人车自主目标识别优化算法

doi:10.12382/bgxb.2022.1161

摘要/Abstract

摘要：

复杂地面战场环境下的自主目标识别是未来无人车实现智能化作战的关键技术。如何将基于深度学习的自主目标识别算法进行资源受限情况下的嵌入式计算平台部署应用,已成为目前无人车智能化赋能亟待解决的核心问题。以用于地面自主目标识别的YOLOv5深度神经网络结构为基础,提出一种基于改进型注意力模块和BatchNorm层的多正则项自适应网络裁剪算法,在裁剪与训练的协同过程中实现网络结构的最优化裁剪;设计一种对权重实施不饱和映射以及对激活值实施饱和映射的组合式训练后INT8量化算法。将压缩优化后的YOLOv5x网络在基于Zynq UltraScale+MPSoC架构的嵌入式计算平台上进行应用部署及验证。验证结果表明:YOLOv5x网络在通道裁剪40%和INT8量化时,红外目标识别精度仅减少0.374%,可见光目标识别精度提升0.065%,目标识别帧频均可达到25帧/s,能够满足无人车复杂战场环境下自主目标识别的准确度及实时性要求;自主目标识别神经网络压缩优化算法可扩展应用于无人机、精确制导武器等其他作战平台。

关键词: 无人车, 目标识别, YOLOv5x, 正则化裁剪, INT8量化算法, 嵌入式平台

Abstract:

Autonomous target recognition is a key technology for enabling intelligent operation of unmanned vehicles in complex ground battlefield environments. However, deploying and applyingdeep learning-based autonomous target recognition algorithmson resource-constrained embedded computing platformsfor realizing intelligent unmanned vehicles remains a challenging task. Based on the YOLOv5 deep neural network structure used for autonomous recognition of ground targets, this paper proposes a multi regularization adaptive network pruning algorithm based on animproved attention module and BatchNorm layer. The proposed algorithm achieves optimal pruning of the network structure during the collaborative process of pruning and training. A combined posttraining INT8 quantization algorithm is designed, employing unsaturated mapping for weights and saturated mapping for activation values. The compressed and optimized YOLOv5x network is then deployed and verified on the embedded computing platform based on the Zynq UltraScale+ MPSoC architecture.The verification results show that when YOLOv5x network prunes 40% of the channel and quantizes with INT8 strategy, the recognition accuracy for infrared dataset is only reduced by 0.374%. The recognition accuracy for visible light dataset is increased by 0.065%, and the target recognition frame rate can reach 25 frames per second. The optimized network can meet the accuracy and real-time requirements of autonomous target recognition in the complex battlefield environment of unmanned vehicles. The proposed network optimization algorithm can be extended to other combat platforms, such as unmanned aerial vehiclesand precision-guided weapons.

Key words: unmanned vehicles, target recognition, YOLOv5x, regularized pruning, INT8 quantization algorithm, embedded platform

中图分类号:

TP391.4

赵晓冬, 张洵颖. 基于YOLOv5的无人车自主目标识别优化算法[J]. 兵工学报, 2023, 44(9): 2732-2744.

ZHAO Xiaodong, ZHANG Xunying. Optimization Algorithm of Autonomous Target Recognition for Unmanned Vehicles Based on YOLOv5[J]. Acta Armamentarii, 2023, 44(9): 2732-2744.

图/表 22

参考文献 26

[1]	李良福, 陈卫东, 高强, 等. 基于深度学习的光电系统智能目标识别[J]. 兵工学报, 2022, 43(增刊1) :162-168.
	LI L F, CHEN W D, GAO Q, et al. Deep learning-based intelligent target recognition technology for electro-optical system[J]. Acta Armamentarii, 2022, 43(S1):162-168. (in Chinese) doi: 10.12382/bgxb.2022.A004
[2]	李博, 王博, 韩京冶, 等. 基于车载计算机的红外图像移动目标检测[J]. 兵工学报, 2022, 43(增刊1): 66-73.
	LI B, WANG B, HAN J Y, et al. Infrareds image moving object detection technology based on onboard computer[J]. Acta Armamentarii, 2022, 43(S1): 66-73. (in Chinese)
[3]	于博文, 吕明. 改进的YOLOv3 算法及其在军事目标检测中的应用[J]. 兵工学报, 2022, 43(2): 345-354. doi: 10.3969/j.issn.1000-1093.2022.02.012
	YU B W, LÜ M. Improved YOLOv3 algorithm and its application in military target detection[J]. Acta Armamentarii, 2022, 43(2): 345-354. (in Chinese) doi: 10.3969/j.issn.1000-1093.2022.02.012
[4]	杨传栋, 钱立志, 薛松, 等. 图像自寻的弹药目标检测方法综述[J]. 兵工学报, 2022, 43(10): 2687-2704.
	YANG C D, QIAN L Z, XUE S, et al. Review on target detection of image homing ammunition[J]. Acta Armamentarii, 2022, 43(10): 2687-2704. (in Chinese) doi: 10.12382/bgxb.2021.0610
[5]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-Time object detection[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Las Vegas, NV, US:IEEE, 2016: 779-788.
[6]	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Honolulu, HI,US:IEEE, 2017: 6517-6525.
[7]	REDMON J, FARHADI A.YOLOv3: an incremental improvements:arXiv: 1804. 02767[R]. Ithaca,NY, US: Cornell University, 2018:1804.02767.
[8]	BOCHKOVSKIY A, WANGC Y, LIAO H M.Yolov4: optimal speed and accuracy of object detection:arXiv: 2004. 10934[R]. Ithaca,NY, US: Cornell University, 2020:2004.10934.
[9]	GE Z, LIU S T, WANG F, et al.YOLOX: exceeding YOLO series in 2021:arXiv: 2107. 08430[R]. Ithaca,NY, US: Cornell University, 2021:2107-08430.
[10]	WANG C Y, BOCHKOVSKIY A, LIAO H M.YOLOv7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors:arXiv: 2207. 02696[R]. Ithaca,NY, US: Cornell University, 2022:2207.02696.
[11]	褚文杰. 基于YOLOv5的坦克装甲车辆目标检测关键技术的研究[D]. 北京: 北京交通大学, 2021.
	ZHU W J. Research on the key technology of tank and armored vehicle target detection based on YOLOv5[D]. Beijing: Beijing Jiaotong University, 2021. (in Chinese)
[12]	林建宇. 基于神经网络的特定目标车辆检测与匹配算法研究[D]. 长春: 吉林大学, 2022.
	LIN J Y. Research on specific target vehicle detection and matching algorithm based on neural network[D]. Changchun: Jilin University, 2022. (in Chinese)
[13]	龙赛. 基于YOLOv5s的航拍图像车辆检测算法研究[D]. 武汉: 华中师范大学, 2022.
	LONG S. Research on vehicle detection algorithm in aerial image based on YOLOv5s[D]. Wuhan: Central China Normal University, 2022. (in Chinese)
[14]	李垠汛. 基于无人机视角的车牌识别技术研究与实现[D]. 银川: 宁夏大学, 2022.
	LI Y X. Research and implementation of license plate recognition technology based on UAV perspective[D]. Yinchuan: Ningxia University, 2022. (in Chinese)
[15]	董光辉, 陈星宇. YOLOv5定位多特征融合的车标识别[J]. 计算机工程与应用, 2023, 59(5):176-193. doi: 10.3778/j.issn.1002-8331.2207-0389
	DONG G H, CHEN X Y. Vehicle logo recognition with YOLOv5 location and multi-feature fusion[J]. Computer Engineering and Applications, 2023, 59(5):176-193. (in Chinese) doi: 10.3778/j.issn.1002-8331.2207-0389
[16]	邓天民, 刘旭慧, 王丽, 等. 结合级联注意力机制的车辆检测算法[J]. 计算机工程与应用.
	DENG T M, LIU X H, WANG L, et al. Vehicle detection algorithm combined with cascading attention mechanism[J]. Computer Engineering and Applications. in Chinese)
[17]	WEN W J, YANG F, SU Y F, et al. Learning sparse patterns in deep neural networks[C]//Proceedings of the IEEE 13rd International Conference on ASIC.Chongqing, China:IEEE, 2019:1-4.
[18]	YU R, LI A, CHEN C F, et al. NISP: pruning networks using neuron importance score propagation[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, US:IEEE, 2018: 9194-9203.
[19]	LIU Z, LI J G, SHEN Z Q, et al. Learning efficient convolutional networks through network slimming[C]//Proceedings of IEEE International Conference on Computer Vision. Venice, Italy:IEEE, 2017: 2755-2763.
[20]	ZHOU A, YAO A, GUO Y W, et al. Incremental network quantization: Towards lossless cnns with low-precision weights:arXiv:1702.03044[R]. Ithaca,NY, US: Cornell University, 2017:1702.03044.
[21]	JACOB B, KLIGYS S, CHEN B, et al. Quantization and training of neural networks for efficient integer-arithmetic-only inference[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, US:IEEE, 2018: 2704-2713.
[22]	汪枭杰. 卷积神经网络压缩技术的研究与实现[D]. 北京: 北京邮电大学, 2019.
	WANG X J. Research and implementation of convolutional neural network compression technique[D]. Beijing: Beijing University of Posts and Telecommunications, 2019. (in Chinese)
[23]	WANG C Y, LIAO M H, WU Y H, et al. CSPNet: a new backbone that can enhance learning capability of CNN[C]//Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Seattle, WA, US:IEEE, 2020: 1571-1580.
[24]	LIN T Y, DOLLAR P, GIRSHICKR. Feature pyramid networks for object detection[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition.Honolulu, HI,US:IEEE, 2017: 936-944.
[25]	LIU S, QI L, QIN H, et al. Path aggregation network for instance segmentation[C]//Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT, US:IEEE, 2018: 8759-8768.
[26]	MERCIONI M A, HOLBAN S. Soft clipping mish-a novel activation function for deep learning[C]//Proceeding of the 4th International Conference on Information and Computer Technologies. HI, US:ICICT, 2021: 13-17.

算法	像素	参数量/10⁶	FLOPs/10⁹
YOLOv5s	416×416	7.2	6.9
YOLOv5m	416×416	21.2	20.7
YOLOv5l	416×416	46.5	46.1
YOLOv5x	416×416	86.7	86.8

算法	像素	参数量/10⁶	FLOPs/10⁹
YOLOv5s	416×416	7.2	6.9
YOLOv5m	416×416	21.2	20.7
YOLOv5l	416×416	46.5	46.1
YOLOv5x	416×416	86.7	86.8

CSP编号	网络类型
CSP编号	YOLOv5s	YOLOv5m	YOLOv5l	YOLOv5x
CSP1_1	1	2	3	4
CSP1_2	3	6	9	12
CSP1_3	3	6	9	12
CSP2_1	1	2	3	4
CSP2_2	1	2	3	4
CSP2_3	1	2	3	4
CSP2_4	1	2	3	4
CSP2_5	1	2	3	4

CSP编号	网络类型
CSP编号	YOLOv5s	YOLOv5m	YOLOv5l	YOLOv5x
CSP1_1	1	2	3	4
CSP1_2	3	6	9	12
CSP1_3	3	6	9	12
CSP2_1	1	2	3	4
CSP2_2	1	2	3	4
CSP2_3	1	2	3	4
CSP2_4	1	2	3	4
CSP2_5	1	2	3	4

数据集	裁剪比率/%	精度mAP(变化幅度/%)
		红外未裁剪0.9882		可见光未裁剪0.9215
		文献[19]算法	ReLU+Sigmoid		Mish+Sigmoid
红外数据集	20	0.9897 (+0.152)	0.9903 (+0.213)		0.9915 (+0.334)
	30	0.9848 (-0.344)	0.9855 (-0.273)		0.9861 (-0.213)
	40	0.9842 (-0.405)	0.9851 (-0.314)		0.9858 (-0.243)
	可见光数据集	20	0.9244 (+0.315)	0.9251 (+0.39)		0.9258 (+0.467)
30		0.9238 (+0.25)	0.9243 (+0.304)		0.9253 (+0.412)
40		0.9218 (+0.033)	0.9225 (+0.109)		0.9229 (+0.152)