Pose estimation and behavior classification of broiler chickens based on deep neural networks
Introduction
In many studies on animal health and welfare, quantitative behavioral analysis plays a vital role (Koolhaas and Van Reenen, 2016, Rushen et al., 2011). Important information about animal behavior can make production managers evaluate animal welfare better (Janczak and Riber, 2015, Peden et al., 2018). Pose estimation is an integral part of behavioral analysis (Li et al., 2019). In some animal posture recognition studies, some additional markers are usually placed on the animal to enhance the recognition precision (Maghsoudi et al., 2017). Although markers can extract animal poses with high recognition precision, this type of system is expensive and distracts animals because they are invasive. Video cameras have the advantage of low invasiveness, which reduces the impact on the animal’s life. Among them, the skeleton fitting and motion contour model is an excellent alternative to physical marking. Although modern high-speed video cameras can accurately record animal behavior, manually analyzing these records requires large amounts of time and labor (Egnor and Kristin, 2016). Video recording and description can take a long time, and these observations are subjective.
With the continuous advancement of computer vision and deep learning technology, research on animals is being improved. Uhlmann et al. (2017) proposed an automatic tracking tool called FlyLimbTracker, which is used to track flies and analyze the behaviors of various flies. Zhuang and Zhang (2019) classified healthy and suspected sick chickens by using a target detection algorithm IFSSD. Fang et al. (2020) used the Tbroiler tracking algorithm to track a single chicken in a flock and analyzed the relate indexes, such as the overlap rate (OR) and pixel error (PE). Lin et al. (2018) used a deep convolutional neural network (CNN) to track broiler chickens by combining heat stress with the broiler chicken activity index as a new predictor, which was used to predict the heat stress response of broiler chickens.
Previously, the postures of animals such as bats and mice were estimated based on traditional image-processing algorithms (Mikhail et al., 2014, Salem et al., 2019, Shen et al., 2017). The DNN is commonly used in animal behavioral classification (Ravbar et al., 2019, Stern et al., 2015). Mathis et al., 2018, Pereira et al., 2019 successfully applied CNNs to animal pose estimation for the first time. Mathis and Pereira's work provided the basis for our research. In general, training a deep CNN requires a large number of marker images (Graving et al., 2019, Günel et al., 2019). Deeplabcut uses ResNet-50 as the initial network for pretraining and overcomes this problem using transfer learning (He et al., 2016). By marking a small number of images, we can accurately track the specific parts of the body of the target chicken, such as chicken legs, chicken neck, chicken head, and so on.
Different from previous broiler chicken detection methods, this study focuses on six common behaviors of broiler chickens, and it implements pose detection methods to detect broiler chickens’ behaviors. The main objective of this article is to combine broiler chickens’ pose estimation networks and classification networks to detect broiler chickens’ behaviors. By combining a neural network based on feature points and a classification network based on the naive bayesian model (NBM), we aimed to analyze six behaviors of broiler chickens.
This paper proposes a user-defined broiler chicken pose skeleton and behavior classification method by using Deeplabcut as the basis for the tracking framework. Through this model, we monitor and track specific parts of the chicken's body in a dynamic environment. Broiler chickens were tracked by pose estimation, and their behaviors were classified.
Section snippets
The pose skeleton model of a broiler chicken
First, we established a simple map of a broiler chicken pose skeleton. As seen from Fig. 1, the plan mainly includes ten custom feature points and broiler chicken pose skeletons. Point 1 is the center point of the chicken, which is the center of the largest inscribed circle in the contour boundary of the broiler chicken (Zhuang et al., 2018). The tail point is the point at which the longest line between the center point and the tail area of the broiler chicken intersects the outline of the
Broiler chicken pose estimation results
The premise of behavioral classification is to accurately estimate the posture of the chicken, and thus, we must analyze the results of the chicken tracking feature points. Fig. 7 shows the loss map for the deep neural network training.
Behavior classification results
Each of the six behaviors after the data expansion includes 100 pose skeletons. The posture characteristics of the chicken skeleton points show different weight values in different behaviors. We obtained Fig. 8 by counting the different feature points of the
Conclusions
This study shows that we can further analyze the behavior of chickens by using deep learning to track specific points in the body of the chicken. Through the preliminary exploration and analysis of the six behavioral postures of chickens, we filled in the blanks of using skeleton feature points to calculate the behaviors of chickens. In addition, the pose estimation and behavior classification methods of chickens based on deep neural networks proposed in this research have certain reference
CRediT authorship contribution statement
Cheng Fang: Conceptualization, Methodology, Software, Writing - original draft. Tiemin Zhang: Supervision, Conceptualization. Haikun Zheng: Visualization, Software. Junduan Huang: Validation, Writing - review & editing. Kaixuan Cuan: Methodology, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
Funding: This work was supported by the National Key Research and Development Plan of China [grant No. 2018YFD0500705] and the Guangdong Province Special Fund for Modern Agricultural Industry Common Key Technology R&D Innovation Team of China [grant No. 2019KJ129].
References (31)
- et al.
Classifier evaluation under limited resources
Pattern Recogn. Lett.
(2006) - et al.
Application of a fully automatic analysis tool to assess the activity of broiler chickens with different gait scores
Comput. Electron. Agric.
(2010) - et al.
Comparative study on poultry target tracking algorithms based on a deep regression network
Biosyst. Eng.
(2020) - et al.
Method for detecting avian influenza disease of chickens based on sound analysis
Biosyst. Eng.
(2019) - et al.
Review of rearing-related factors affecting the welfare of laying hens
Poult. Sci.
(2015) - et al.
The translation of animal welfare research into practice: The case of mixing aggression between pigs
Appl. Animal Behav. Sci.
(2018) - et al.
A systematic analysis of performance measures for classification tasks
Inf. Process. Manage.
(2009) - et al.
Relationships between beak condition, preening behavior and ectoparasite infestation levels in laying hens
Poult. Sci.
(2015) - et al.
Feature selection for multi-label naive Bayes classification
Inf. Sci.
(2009) - et al.
Development of an early warning algorithm to detect sick broilers
Comput. Electron. Agric.
(2018)