深度学习在畜禽典型行为识别中的研究进展

doi:10.13304/j.nykjdb.2023.0168

摘要/Abstract

摘要：

采食、饮水、站立和打斗等典型行为与畜禽生产能力、健康状况和福利密切相关，影响畜禽产量与经济效益。当前，畜禽养殖规模化趋势加速，传统人工观察畜禽行为不仅费时费力，而且主观性较强。随着机器学习的快速发展，神经网络、算法和算力不断优化，计算机视觉、语音识别、生物识别、自然语言处理等技术能准确高效地监测畜禽信息，分析畜禽生理和健康状况，在畜禽领域展现出广阔的应用前景。介绍了深度学习技术的发展历程，阐述了深度学习技术在常见畜禽种类（牛、猪、羊、鸡）行为识别方面的研究进展，为未来研究和实际应用提供了技术参考；总结了深度学习技术在畜禽行为识别中关于模型通用性、数据集多样性和数字化行为结果全面性等方面存在的问题并提出改进策略，旨在推动深度学习在畜禽典型行为中的进一步应用。

关键词: 畜禽, 深度学习, 行为识别

Abstract:

The typical behaviors， such as feeding， drinking， standing and fighting， are closely related to the production capacity， health status as well as welfare of livestock and poultry， which affecting the production and economic benefits of livestock and poultry in farms. In fact， the traditional manual observation of livestock and poultry is not only time-consuming and laborious， but also highly subjective. So currently， the trend of large-scale livestock and poultry farming is accelerating. With the rapid development of machine learning as well as continuous optimization of neural networks， algorithms and computility， technologies such as computer vision， speech recognition， biometric recognition and natural language processing can accurately and efficiently monitor the information of livestock and poultry as well as analyze the physiological and health status of livestock and poultry， showing broad application prospects in the field of livestock and poultry. This article introduced the development history of deep learning technology， and then expounded the research progress of deep learning technology in behavior recognition of common livestock and poultry species such as cattle， pig， sheep and chicken， providing technical reference for future researches and practical applications. Meanwhile， this article summarized the problems and improvement strategies of deep learning technology in behavior recognition of common livestock and poultry from aspects of model versatility， data set diversity as well as the comprehensiveness of digital behavior results， aiming to provide theoretical reference for technicians to promote the further development of deep learning in the application of typical behavior of livestock and poultry.

Key words: livestock and poultry, deep learning, behavior recognition

中图分类号:

S126

朱芷芫, 王海峰, 李斌, 赵文文, 朱君, 贾楠, 赵宇亮. 深度学习在畜禽典型行为识别中的研究进展[J]. 中国农业科技导报, 2024, 26(10): 110-124.

Zhiyuan ZHU, Haifeng WANG, Bin LI, Wenwen ZHAO, Jun ZHU, Nan JIA, Yuliang ZHAO. Research Progress of Deep Learning in Typical Behavior Recognition of Livestock and Poultry[J]. Journal of Agricultural Science and Technology, 2024, 26(10): 110-124.

图/表 9

参考文献 57

1	国家统计局.中华人民共和国2022年国民经济和社会发展统计公报[N].人民日报,2023-02-28(009).
2	滕光辉,冀横溢,庄晏榕,等.深度学习在猪只饲养过程的应用研究进展[J].农业工程学报, 2022, 38(14):235-249.
	TENG G H, JI H Y, ZHUANG R Y, et al.. Research progress of deep learning in the process of pig feeding [J]. Trans. Chin. Soc. Agric. Eng., 2022, 38(14): 235-249.
3	Chen C, Zhu W X, Norton T. Behaviour recognition of pigs and cattle: Journey from computer vision to deep learning [J/OL]. Comput. Electron. Agric., 2021,187(1-3):106255 [2024-01-13]. .
4	代东亮,刘志红,赵存,等.人工智能技术在畜牧业中的应用进展[J].畜牧与饲料科学,2021, 42(5):112-119.
	DAI D L, LIU Z H, ZHAO C, et al.. Research advances on application of artificial intelligence technology in animal husbandry [J]. Anim. Husbandry Feed Sci., 2021,42(5):112-119.
5	字吉湖,谭利辉,赵永聚,等.机器视觉在猪行为识别中的应用进展[J].现代畜牧科技, 2022(11):26-28.
	ZI J H, TAN L H, ZHAO Y J, et al.. Application progress of machine vision in pig behavior recognition [J]. Modern Anim. Husbandry Sci. Technol., 2022(11):26-28.
6	薛芳芳,王月明,李琦.基于计算机视觉技术的牲畜行为识别研究进展[J].黑龙江畜牧兽医, 2021(11):33-38.
	XUE F F, WANG Y M, LI Q. Research progress of animal behavior recognition based on computer vision technology [J]. Heilongjiang Anim. Sci. Veterinary, 2021(11):33-38.
7	姚州,谭焓,田芳,等.计算机视觉技术在智慧羊场中的研究进展[J].中国饲料, 2021(7):7-12.
	YAO Z, TAN H, TIAN F, et al.. Research progress of computer vision technology in wisdom sheep farm [J].China Feed, 2021(7):7-12.
8	刘峰,吴文杰,刘小磊,等.计算机视觉与深度学习在猪只识别中的研究进展[J].华中农业大学报,2023,42(3):47-56.
	LIU F, WU W J, LIU X L, et al.. Progress of computer vision and deep learning methods for pig's identity and behavior recognition [J]. J. Huazhong Agric. Univ., 2023,42(3):47-56.
9	张军阳,王慧丽,郭阳, 等.深度学习相关研究综述[J].计算机应用研究,2018,35(7): 1921-1928, 1936.
	ZHANG J Y, WANG H L, GUO Y, et al.. Review of deep learning [J]. Appl. Res. Comput., 2018,35(7):1921-1928, 1936.
10	GOODFELLOW I, BENGIO Y, COURVILLE A. Deep Learning [M]. Cambridge: MIT Press, 2016:1-800.
11	LECUN Y, BENGIO Y, HINTON G. Deep learning [J]. Nature, 2015, 521(7553): 436-444.
12	QIAO Y L, KONG H, CLARK C, et al.. Intelligent perception for cattle monitoring: A review for cattle identification, body condition score evaluation, and weight estimation [J/OL]. Comput. Electron. Agric., 2021, 185: 106143 [2024-01-13]. .
13	WU Y, SCHUSTER M, CHEN Z, et al.. Google’s neural machine translation system: Bridging the gap between human and machine translation [EB/OL]. (2016-09-26) [2024-01-13]. .
14	KRIZHEVSKY A, SUTSKEVER I, HINTON E G. ImageNet classification with deep convolutional neural networks [M]// Advances in Neural Information Processing Systems, 2012:1097-1105.
15	SZEGEDY C,0015 L W,JIA Y,et al.. Going deeper with convolutions [C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015: 1-9.
16	SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition [J/OL]. Comput. Electron. Agric.,2014(9): 1556 [2024-01-13]. .
17	ZIN T T, PHYO C N, TIN P, et al.. Image technology based cow identification system using deep learning [C]// Proceedings of International Multi Conference of Engineers and Computer Scientists. Piscataway, 2018:14-16.
18	周飞燕,金林鹏,董军.卷积神经网络研究综述[J].计算机学报,2017,40(6):1229-1251.
	ZHOU F Y, JIN L P, DONG J. Review of convolutional neural network [J]. Chin. J. Comput., 2017, 40(6): 1229-1251.
19	GOODFELLOW J I, POUGET-ABADIE J, MIRZA M, et al.. Generative adversarial networks [J]. Adv. Neural Inform. Proc. Syst.,2014,3:2672-2680.
20	李丹,张凯锋,李行健,等.基于 Mask R-CNN 的猪只爬跨行为识别[J].农业机械学报, 2019, 50(S1): 261-266, 275
	LI D, ZHANG K F, LI X J, et al. Mounting behavior recognition for pigs based on mask R-CNN [J]. Trans. Chin. Soc. Agric. Mach., 2019, 50(S1): 261-266, 275.
21	CHENG M, YUAN H, WANG Q, et al.. Application of deep learning in sheep behaviors recognition and influence analysis of training data characteristics on the recognition effect [J/OL]. Comput. Electron. Agric.,2022,198: 107010 [2024-01-13]. .
22	SONG S, LIU T, WANG H, et al.. Using pruning-based YOLOv3 deep learning algorithm for accurate detection of sheep face [J/OL]. Animals, 2022,12(11):1465 [2024-01-13]. .
23	DENG Z W, VAHDAT A, HU H X, et al.. Structure inference machines: Recurrent neural networks for analyzing relations in group activity recognition [C]// Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Piscataway, NJ, USA: IEEE, 2016: 4772-4781.
24	KARIM F, MAJUMDAR S, DARABI H, et al. Multivari‐ate LSTM-FCNs for time series classification [J]. Neural Networks, 2019, 116: 237-245.
25	LI W J, QI F, TANG M, et al.. Bidirectional LSTM with self-attention mechanism and multi-channel features for sentiment classification [J]. Neurocomputing, 2020, 387:63-77.
26	ACHOUR B, BELKADI M, FILALI I, et al.. Image analysis for individual identification and feeding behaviour monitoring of dairy cows based on Convolutional Neural Networks (CNN) [J]. Biosyst. Eng., 2020,198:31-49.
27	BEZEN R, EDAN Y, HALACHMI I. Computer vision system for measuring individual cow feed intake using RGB-D camera and deep learning algorithms [J/OL]. Comput. Electron. Agric., 2020,172I:105345 [2024-01-13]. .
28	LI Z, SONG L, DUAN Y, et al.. Basic motion behaviour recognition of dairy cows based on skeleton and hybrid convolution algorithms [J/OL]. Comput. Electron. Agric.,2022,196:106889 [2024-01-13]. .
29	王少华,何东健,刘冬.基于机器视觉的奶牛发情行为自动识别方法[J].农业机械学报,2020, 51(4):241-249.
	WANG S H, HE J D, LIU D. Automatic eecognition method of dairy cow estrus behavior based on machine vision [J].Trans. Chin. Soc. Agric. Mach., 2020, 51(4):241-249.
30	王少华,何东健.基于改进YOLOv3模型的奶牛发情行为识别研究[J].农业机械学报,2021,52(7):141-150.
	WANG S H, HE J D. Estrus behavior recognition of dairy cows based on improved YOLO v3 model [J].Trans. Chin. Soc. Agric. Mach.,2021,52(7):141-150.
31	WANG R, BAI Q, GAO R, et al.. Oestrus detection in dairy cows by using atrous spatial pyramid and attention mechanism [J]. Biosyst. Eng., 2022, 223:259-276.
32	王政,许兴时,华志新,等.融合YOLO v5n与通道剪枝算法的轻量化奶牛发情行为识别方法[J].农业工程学报,2022,38(23) :134-144.
	WANG Z, XU X S, HUA Z X, et al.. Lightweight recognition for the oestrus behavior of dairy cows combining YOLO v5n and channel pruning [J]. Trans. Chin. Soc. Agric. Eng.,2022,38(23):130-140.
33	FUENTES A, YOON S, PARK J. Deep learning-based hierarchical cattle behavior recognition with spatiotemporal information [J/OL]. Comput. Electron. Agric.,2020,177(1):105627 [2024-01-13]. .
34	YIN X, WU D, SHANG Y, et al.. Using an EfficientNet-LSTM for the recognition of single Cow’s motion behaviours in a complicated environment [J/OL]. Comput. Electron. Agric., 2020,177:105707 [2024-01-13]. .
35	YANG Q, XIAO D, LIN S. Feeding behavior recognition for group-housed pigs with the Faster R-CNN [J]. Computers Electron. in Agric.,2018,155:453-460.
36	CHEN C, ZHU W, STEIBEL J, et al.. Recognition of feeding behaviour of pigs and determination of feeding time of each pig by a video- based deep learning method [J/OL]. Comput. Electron. Agric.,2020,176:105642 [2024-01-13]. .
37	CHEN C, ZHU W, STEIBEL J, et al.. Classification of drinking and drinker-playing in pigs by a video-based deep learning method [J]. Biosys. Eng., 2020,196 :1-14.
38	杨秋妹,肖德琴,张根兴.猪只饮水行为机器视觉自动识别[J].农业机械学报,2018,49(6):232-238.
	YANG Q, XIAO D, ZHANG G. Automatic pig drinking behavior recognition with machine vision [J]. Trans. Chin. Soc. Agric. Mach., 2018,49(6):232-238.
39	ZHENG C, ZHU X, YANG X, et al.. Automatic recognition of lactating sow postures from depth images by deep learning detector [J]. Comput. Electron. Agric., 2018,147:51-63.
40	薛月菊,朱勋沐,郑婵,等. 基于改进Faster R-CNN识别深度视频图像哺乳母猪姿态[J]. 农业工程学报,2018,34(9) :189-196.
	XUE Y J, ZHU X M, ZHENG C, et al.. Lactating sow postures recognition from depth image of videos based on improved faster R-CNN [J]. Trans. Chin. Soc. Agric. Eng.,2018, 34(9):189-196.
41	NASIRAHMADI A, STURM B, EDWARDS S,et al.. Deep learning and machine vision approaches for posture detection of individual pigs [J/OL]. Sensors, 2019, 19(17):3738 [2024-01-13].
42	ZHU X, CHEN C, ZHENG B, et al.. Automatic recognition of lactating sow postures by refined two-stream RGB-D Faster R-CNN [J]. Biosyst. Eng., 2020,189:116-132.
43	RIEKERT M, KLEIN A, ADRION F, et al.. Automatically detecting pig position and posture by 2D camera imaging and deep learning [J/OL]. Comput. Electron. Agric., 2020,174:105391 [2024-01-13]. .
44	CHEN C, ZHU W, STEIBEL J, et al.. Recognition of aggressive episodes of pigs based on convolutional neural network and long short-term memory [J]. Comput. Electron. Agric., 2020,169:105166 [2024-01-13]. .
45	高云,陈斌,廖慧敏,等. 群养猪侵略性行为的深度学习识别方法[J]. 农业工程学报, 2019, 35(23)192-200.
	GAO Y, CHEN B, LIAO H, et al.. Recognition method for aggressive behavior of group pigs based on deep learning [J]. Trans. Chin. Soc. Agric. Eng., 2019,35(23):192-200.
46	ZHANG Y, CAI J, XIAO D, et al.. Real-time sow behavior detection based on deep learning [J/OL]. Comput. Electron. Agric., 2019,163:104884 [2024-01-13]. .
47	YANG A, HUANG H, ZHENG B, et al.. An automatic recognition framework for sow daily behaviours based on motion and image analyses[J]. Biosystems Engineering,2020,192:56-71.
48	LI D, ZHANG K, LI Z, et al.. A spatiotemporal convolutional network for multi-behavior recognition of pigs [J/OL]. Sensors, 2020,20(8):2381 [2024-01-13]. .
49	YANG A, HUANG H, YANG X, et al.. Automated video analysis of sow nursing behavior based on fully convolutional network and oriented optical flow [J/OL]. Comput. Electron. Agric., 2019,167:105048 [2024-01-13]. .
50	MALLY C I, WURTZ K E, STEIBEL J P, et al.. Relationships among aggressiveness, fearfulness and response to humans in finisher pigs [J]. App. Anim. Behaviour Sci., 2018,205:194-201.
51	WANG K, WU P, CUI H, et al.. Identification and classification for sheep feeding behavior based on acoustic signal and deep learning [J/OL]. Comput. Electron. Agric.,2021,187(195):106275 [2024-01-13]. .
52	陆明洲,梁钊董, NORTON T .等.基于EfficientDet网络的湖羊短时咀嚼行为识别方法[J]. 农业机械学报,2021,52(8):248-254, 426.
	LU M Z, LIANG Z D, NORTON T, et al.. Automatic identification method of short-term chewing behaviour for sheep based on EfficientDet Network [J]. Trans. Chin. Soc. Agric. Mach., 2021, 52(8):248-254, 426.
53	李小迪,王天一.基于改进卷积神经网络的羊行为识别[J].智能计算机与应用,2022,12(12):226-230.
	LI X D, WANG T Y. Sheep behavior recognition based on improved convolutional neural network [J]. Intelligent Comput. Appl., 2022,12(12): 226-230.
54	JIANG M, RAO Y, ZHANG J, et al.. Automatic behavior recognition of group-housed goats using deep learning [J/OL]. Comput. Electron. Agric.,2020,177(1):105706 [2024-01-13]. .
55	CHENG M, YUAN H, WANG Q, et al.. Application of deep learning in sheep behaviors recognition and influence analysis of training data characteristics on the recognition effect [J]. Comput. Electron. Agric., 2022, 198 :107010 [2024-01-13]. .
56	CHENG F, ZHANG T, ZHENG H, et al.. Pose estimation and behavior classification of broiler chickens based on deep neural networks [J/OL]. Computers Electron. Agric., 2021,180:105863 [2024-01-13]. .
57	李娜,任昊宇,任振辉.基于深度学习的群养鸡只行为监测方法研究[J]. 河北农业大学学报,2021,44(2):117-121.
	LI N, REN H, REN Z. Research of behavior monitoring method of flock hens based on deep learning [J]. J. Hebei Agric. Univ., 2021,44(2):117-121.

行为Behaviour		深度学习模型 Deep learning model	结果 Result	优缺点 Advantage and disadvantage	参考文献 Reference
单行为 Single behavior	采食 Feeding	CNN	准确率Accuracy：92.00%	无法统计具体采食量 Specific feed intake is not available	［26］
	采食 Feeding	Fast R-CNN	准确率Accuracy：93.65%	识别准确率高，但依赖位置信息，当采食槽为空时依然会识别为采食 The model has high recognition accuracy， but depending on location information， it can recognize the feeding situation even when the feeding tank is empty	［27］
	运动 Locomotion	HRNET	准确率Accuracy：91.80%	识别准确率待提高，使用时需结合目标检测算法 The accuracy of model recognition should be improved， and the target detection algorithm should be combined	［28］
	交配 Mating	GMM+YOLOv3	准确率Accuracy： 100.00%	识别准确率高，适用于复杂环境 The model recognition accuracy is high， and it is suitable for complex environment	［29］
		DenseBlock+ YOLOv3	准确率Accuracy： 99.15%	识别精度较高，模型泛化能力强 The model recognition accuracy is high， and the model generalization ability is strong	［30］
		C3GC3+YOLOv5	平均精度均值mAP： 94.30%	识别精度高，但模型大，检测速度慢 The model recognition accuracy is high， but the scale is large and the speed is slow	［31］
		YOLOv5n+通道剪枝 YOLOv5n and channel pruning	平均精度均值mAP： 97.9%	识别精度较高，检测速度高 The model recognition accuracy is relatively high， and the detection speed is high	［32］
多行为 Multi-behavior	攻击Aggressive	YOLOv3	平均精度均值mAP：76.00%	对于多目标牛只行为检测具有一定局限性 The model has some limitations for multi-target cattle behavior detection	［33］
	交配Mating		平均精度均值mAP：79.10%
	站立Standing		平均精度均值mAP：87.10%
	采食Feeding		平均精度均值mAP：81.10%
	行走Walking		平均精度均值mAP：83.10%
	躺卧Lying		平均精度均值mAP：82.30%
	睡觉Sleeping		平均精度均值mAP：74.30%
	甩尾Tail flicking		平均精度均值mAP：64.20%
	摇头Head shaking		平均精度均值mAP：75.30%
	舔舐Licking		平均精度均值mAP：85.10%
	自我梳理Self-grooming		平均精度均值mAP：80.40%
	采食Feeding	EfficientNet-LSTM	精度Precision：99.00%	识别准确率高，但只识别单只牛行为 The model has high recognition accuracy， but only recognizes the behavior of a single cow	［34］
	饮水Drinking		精度Precision：95.84%
	行走Walking		精度Precision：98.85%
	站立Standing		精度Precision：96.90%
	躺卧Lying		精度Precision：98.60%

行为Behaviour		深度学习模型 Deep learning model	结果 Result	优缺点 Advantage and disadvantage	参考文献 Reference
单行为 Single behavior	采食 Feeding	CNN	准确率Accuracy：92.00%	无法统计具体采食量 Specific feed intake is not available	［26］
	采食 Feeding	Fast R-CNN	准确率Accuracy：93.65%	识别准确率高，但依赖位置信息，当采食槽为空时依然会识别为采食 The model has high recognition accuracy， but depending on location information， it can recognize the feeding situation even when the feeding tank is empty	［27］
	运动 Locomotion	HRNET	准确率Accuracy：91.80%	识别准确率待提高，使用时需结合目标检测算法 The accuracy of model recognition should be improved， and the target detection algorithm should be combined	［28］
	交配 Mating	GMM+YOLOv3	准确率Accuracy： 100.00%	识别准确率高，适用于复杂环境 The model recognition accuracy is high， and it is suitable for complex environment	［29］
		DenseBlock+ YOLOv3	准确率Accuracy： 99.15%	识别精度较高，模型泛化能力强 The model recognition accuracy is high， and the model generalization ability is strong	［30］
		C3GC3+YOLOv5	平均精度均值mAP： 94.30%	识别精度高，但模型大，检测速度慢 The model recognition accuracy is high， but the scale is large and the speed is slow	［31］
		YOLOv5n+通道剪枝 YOLOv5n and channel pruning	平均精度均值mAP： 97.9%	识别精度较高，检测速度高 The model recognition accuracy is relatively high， and the detection speed is high	［32］
多行为 Multi-behavior	攻击Aggressive	YOLOv3	平均精度均值mAP：76.00%	对于多目标牛只行为检测具有一定局限性 The model has some limitations for multi-target cattle behavior detection	［33］
	交配Mating		平均精度均值mAP：79.10%
	站立Standing		平均精度均值mAP：87.10%
	采食Feeding		平均精度均值mAP：81.10%
	行走Walking		平均精度均值mAP：83.10%
	躺卧Lying		平均精度均值mAP：82.30%
	睡觉Sleeping		平均精度均值mAP：74.30%
	甩尾Tail flicking		平均精度均值mAP：64.20%
	摇头Head shaking		平均精度均值mAP：75.30%
	舔舐Licking		平均精度均值mAP：85.10%
	自我梳理Self-grooming		平均精度均值mAP：80.40%
	采食Feeding	EfficientNet-LSTM	精度Precision：99.00%	识别准确率高，但只识别单只牛行为 The model has high recognition accuracy， but only recognizes the behavior of a single cow	［34］
	饮水Drinking		精度Precision：95.84%
	行走Walking		精度Precision：98.85%
	站立Standing		精度Precision：96.90%
	躺卧Lying		精度Precision：98.60%

行为 Behaviour			深度学习模型 Deep learning model	结果 Result	优缺点 Advantage and disadvantage	参考文献 Reference
采食Feeding			Faster R-CNN+ZfNet	准确度Accuracy：99.60%	识别准确率高，但依赖位置信息，当采食槽为空时依然会识别为采食 The model has high recognition accuracy， but depends on the location information， it can still recognize the feeding when the feeding tank is empty	［35］
采食Feeding			CNN-LSTM	准确度Accuracy： 98.40%	识别准确率高 The recognition accuracy of the model is high	［36］
饮水Drinking			GoogLeNet	准确度Accuracy： 92.11%	可识别饮水和玩耍行为，但猪只饮水器更换则需要重新训练 The model recognizes drinking and play behaviors， but needs to be retrained after the drinker is replaced	［37］
饮水Drinking			ResNet50-LSTM	准确度Accuracy： 92.50%	识别准确率较高，但依赖位置信息 Recognition accuracy is high， but it depends on location information	［38］
站立Standing			Faster R-CNN	精度Precision： 99.10%	识别准确率高，但模型体积较大 The recognition accuracy is high， but the model volume is large	［39］
			ZF+Faster R-CNN	平均精度AP： 96.73%	［40］
			Faster R-CNN/SSD/R-FCN	平均精度AP： 93.00%	［41］
			Faster R-CNN	平均精度AP： 99.74%	［42］
坐立Sitting			Faster R-CNN	精度Precision： 77.60%	基于深度图像识别猪坐立行为，识别精度高，但计算量大 The recognition of pig sitting and standing behavior based on depth image has high recognition accuracy but large computation	［39］
			ZF+Faster R-CNN	平均精度AP： 94.62%	［40］
			Faster R-CNN	平均精度AP： 99.74%	［42］
单行为Single-behavior		躺卧Lying	Faster R-CNN+NAS	平均精度AP： 80.20%	可长时间多视角识别，但准确率较低 The model can recognize multiple perspectives for a long time， but the accuracy is low	［43］
		俯卧Sternal recumbency	Faster R-CNN	精度Precision： 76.30%	检测速度快，但识别精度较低The model detection speed is fast，but the accuracy is low	［39］
		ZF+Faster R-CNN	平均精度AP： 86.28%	［40］
		Faster R-CNN	平均精度AP： 90.77%	［42］
		腹卧Ventral recumbency	Faster R-CNN	精度Precision： 97.20%	识别准确率高，但模型体积较大 The model recognition accuracy is high， but the model volume is large	［39］
		ZF+Faster R-CNN	平均精度AP： 89.57%	［40］
		Faster R-CNN/SSD/R-FCN	平均精度AP： 92.00%	［41］
		Faster R-CNN	平均精度AP： 90.91%	［42］
		侧卧Lateral recumbency	Faster R-CNN	精度Precision： 98.70%	可识别猪姿态变化情况，识别准确率高，但模型体积较大The model can recognize the change of pig posture with high recognition accuracy，but the model volume is large	［39］
		ZF+Faster R-CNN	平均精度AP： 99.04%	［40］
		Faster R-CNN/SSD/R-FCN	平均精度AP： 92.00%	［41］
		Faster R-CNN	平均精度AP： 99.45%	［42］
		攻击Aggressive	CNN-LSTM	准确度Accuracy： 97.20%	识别准确率高，但数据量少且单一，模型易过拟合 The model recognition accuracy is high， but the data amount is small and single， and the model is easy to overfit	［44］
		3DConvNet	准确度Accuracy： 95.70%	识别准确率高，但无法精确到个体Recognition accuracy is high， but not accurate to the individual	［45］
多行为Multi-behavior	饮水 Drinking		SSD+MobileNet	平均精度均值mAP： 96.50%	识别准确率高，但仅使用空间特征 Recognition accuracy is high， but only spatial features are used	［46］
	交配 Mating		平均精度均值mAP： 92.30%
	排泄Urination		平均精度均值mAP： 91.40%
	采食Feeding		FCN	准确度Accuracy： 95.36%	识别准确率较高，但需人工标记时空特征 The recognition accuracy is higher， but the temporal and spatial features are marked manually	［47］
	饮水Drinking		准确度Accuracy： 97.49%
	哺乳Suckling		准确度Accuracy： 97.60%
	采食Feeding		PMB-SCN	准确度Accuracy： 94.59%	模型泛化能力强 The model has strong generalization ability	［48］
	躺卧 Lying		平均精度均值mAP： 97.73%
	攻击 Aggressive		准确度Accuracy： 100.00%
	交配 Mating		准确度Accuracy： 95.65%

行为 Behaviour			深度学习模型 Deep learning model	结果 Result	优缺点 Advantage and disadvantage	参考文献 Reference
采食Feeding			Faster R-CNN+ZfNet	准确度Accuracy：99.60%	识别准确率高，但依赖位置信息，当采食槽为空时依然会识别为采食 The model has high recognition accuracy， but depends on the location information， it can still recognize the feeding when the feeding tank is empty	［35］
采食Feeding			CNN-LSTM	准确度Accuracy： 98.40%	识别准确率高 The recognition accuracy of the model is high	［36］
饮水Drinking			GoogLeNet	准确度Accuracy： 92.11%	可识别饮水和玩耍行为，但猪只饮水器更换则需要重新训练 The model recognizes drinking and play behaviors， but needs to be retrained after the drinker is replaced	［37］
饮水Drinking			ResNet50-LSTM	准确度Accuracy： 92.50%	识别准确率较高，但依赖位置信息 Recognition accuracy is high， but it depends on location information	［38］
站立Standing			Faster R-CNN	精度Precision： 99.10%	识别准确率高，但模型体积较大 The recognition accuracy is high， but the model volume is large	［39］
			ZF+Faster R-CNN	平均精度AP： 96.73%	［40］
			Faster R-CNN/SSD/R-FCN	平均精度AP： 93.00%	［41］
			Faster R-CNN	平均精度AP： 99.74%	［42］
坐立Sitting			Faster R-CNN	精度Precision： 77.60%	基于深度图像识别猪坐立行为，识别精度高，但计算量大 The recognition of pig sitting and standing behavior based on depth image has high recognition accuracy but large computation	［39］
			ZF+Faster R-CNN	平均精度AP： 94.62%	［40］
			Faster R-CNN	平均精度AP： 99.74%	［42］
单行为Single-behavior		躺卧Lying	Faster R-CNN+NAS	平均精度AP： 80.20%	可长时间多视角识别，但准确率较低 The model can recognize multiple perspectives for a long time， but the accuracy is low	［43］
		俯卧Sternal recumbency	Faster R-CNN	精度Precision： 76.30%	检测速度快，但识别精度较低The model detection speed is fast，but the accuracy is low	［39］
		ZF+Faster R-CNN	平均精度AP： 86.28%	［40］
		Faster R-CNN	平均精度AP： 90.77%	［42］
		腹卧Ventral recumbency	Faster R-CNN	精度Precision： 97.20%	识别准确率高，但模型体积较大 The model recognition accuracy is high， but the model volume is large	［39］
		ZF+Faster R-CNN	平均精度AP： 89.57%	［40］
		Faster R-CNN/SSD/R-FCN	平均精度AP： 92.00%	［41］
		Faster R-CNN	平均精度AP： 90.91%	［42］
		侧卧Lateral recumbency	Faster R-CNN	精度Precision： 98.70%	可识别猪姿态变化情况，识别准确率高，但模型体积较大The model can recognize the change of pig posture with high recognition accuracy，but the model volume is large	［39］
		ZF+Faster R-CNN	平均精度AP： 99.04%	［40］
		Faster R-CNN/SSD/R-FCN	平均精度AP： 92.00%	［41］
		Faster R-CNN	平均精度AP： 99.45%	［42］
		攻击Aggressive	CNN-LSTM	准确度Accuracy： 97.20%	识别准确率高，但数据量少且单一，模型易过拟合 The model recognition accuracy is high， but the data amount is small and single， and the model is easy to overfit	［44］
		3DConvNet	准确度Accuracy： 95.70%	识别准确率高，但无法精确到个体Recognition accuracy is high， but not accurate to the individual	［45］
多行为Multi-behavior	饮水 Drinking		SSD+MobileNet	平均精度均值mAP： 96.50%	识别准确率高，但仅使用空间特征 Recognition accuracy is high， but only spatial features are used	［46］
	交配 Mating		平均精度均值mAP： 92.30%
	排泄Urination		平均精度均值mAP： 91.40%
	采食Feeding		FCN	准确度Accuracy： 95.36%	识别准确率较高，但需人工标记时空特征 The recognition accuracy is higher， but the temporal and spatial features are marked manually	［47］
	饮水Drinking		准确度Accuracy： 97.49%
	哺乳Suckling		准确度Accuracy： 97.60%
	采食Feeding		PMB-SCN	准确度Accuracy： 94.59%	模型泛化能力强 The model has strong generalization ability	［48］
	躺卧 Lying		平均精度均值mAP： 97.73%
	攻击 Aggressive		准确度Accuracy： 100.00%
	交配 Mating		准确度Accuracy： 95.65%

行为 Behaviour		深度学习模型 Deep learning model	结果 Result	优缺点 Advantage and disadvantage	参考文献 Reference
多行为 Multi-behavior	采食Feeding	FCN+YOLOv4	精度Precision：93.61%	基于骨架识别，易混淆识别行走和跑步行为The research is based on skeleton， easily confused to recognize walking and running behavior	［56］
	站立Standing		精度Precision：75.11%
	跑步Running		精度Precision：60.27%
	行走Walking		精度Precision：51.35%
	休息Resting		精度Precision：96.23%
	啄羽Feather pecking		精度Precision：92.58%
	采食Feeding	NBSFCN+YOLOv4	平均精度AP：96.67%	基于2D图像，部分行为识别平均精度较低，图像采集角度单一 The research is based on 2D images， the average accuracy of some behavior recognition is low， and the image acquisition angle is single	［57］
	站立Standing		平均精度AP：90.34%
	梳羽Preening		平均精度AP：82.01%
	啄羽Feather pecking		平均精度AP：63.38%
	攻击Aggressive		平均精度AP：67.14%