A machine learning-based positioning method for poultry in cage environments

Highlights

• A signal strength-based model works for problems with low accuracy requirements.

• The proposed method quantifies the regional activity of an individual caged chicken.

• The results from the proposed method are highly correlated with the video statistics.

Abstract

Individuals or groups of animals exhibit different activities that characterize domain behavior. Rapid and accurate localization of poultry in small and complex cage environments helps analyze the poultry domain behavior. This study proposes a machine-learning-based method for locating poultry in small and complex cage environments. Here, the characteristics of ultra-high frequency-radio frequency identification devices were determined, received signal strength indicator values were collected, and the tag-coordinate regression problem was converted into a multi-area classification problem. Different models were used to predict the target position. The results revealed that the neural network model yielded the best prediction, locating the target within a 40 cm × 40 cm area with 88.74% accuracy or within a 30 cm × 30 cm area with 76.81% accuracy, with average errors of 7.61 cm and 7.97 cm, respectively. Finally, experiments with live chickens were performed, and the results were verified using synchronized video, obtaining a Pearson correlation coefficient exceeding 0.909. This study presents a feasible method for target localization in small and complex cage environments, providing valuable modal information for multimodal learning.

Introduction

Individuals or groups of animals engage in a variety of activities. As their ranges often overlap, they generally tend to mutual avoidance rather than repelling each other. Numerous studies have demonstrated a connection between domain behavior and animal social hierarchies (Milewski et al., 2022). A dominance hierarchy develops among congregated roosters, with the highest-ranking rooster being dominant and having territorial priority (Collias et al., 1966). Dominant roosters maintain a fixed territory, whereas lower-ranking roosters retain only a small area adjacent to themselves (McBride et al., 1969). Higher-rank roosters occupy significantly more cage space and dominant areas than lower-rank roosters (Odén et al., 2004). Individuals in social species frequently form dominance relationships, in which dominant individuals have greater access to resources than subordinate individuals (Favati et al., 2014). Male domestic fowl explore new areas more quickly and maintain long-term vigilance than female domestic fowl. In contrast, abnormal walking patterns are a strong indicator of decreased welfare in broiler chickens. It is critical to understand how captive birds utilize their space to promote welfare by optimizing space quality and meeting their biological needs (Aydin, 2016).

Various technologies have been developed for indoor positionings, such as satellite-based systems (Obeidat et al., 2021), inertial navigation systems (INS), magnet-based systems, and radio frequency-based systems (Denis et al., 2019). However, many of these methods are ineffective in locating small animals. For instance, because of building exterior walls, the global positioning system (GPS), one of the most widely used satellite-based navigation systems, cannot locate indoor objects (Nirjon et al., 2014). INS can be error-prone, necessitating sophisticated filtering techniques such as the Kalman filter (Hu et al., 2020). Magnetic technology is precise but susceptible to conductive and ferromagnetic materials at low frequencies (Diaz et al., 2019). In addition, magnetic-based navigation systems typically rely on disturbances in the Earth’s magnetic field within enclosed environments, resulting from the ferromagnetic nature of metal structures within buildings (Shu et al., 2015). As the Internet of Things (IoT) technology has advanced in recent decades, radio frequency-based systems have been rapidly developed, becoming more suitable for indoor positionings, such as frequency modulation technology (Popleteev, 2017), Wi-Fi (Bagosi and Baruch, 2011), ZigBee (Bianchi et al., 2018), Bluetooth (Wang et al., 2015), radio frequency identification (RFID) (Tesoriero et al., 2010), and LoRa (Islam et al., 2019).

Among these technologies, ultra-high frequency (UHF)-RFID has been applied to agricultural positioning research because of its long identification distance, high accuracy, rapid response time, and strong anti-interference capability. UHF-RFID-based positioning methods are classified as range-based (distance measurement and angle measurement methods) and range-free (fingerprinting and non-fingerprinting methods) (Li et al., 2019). Based on the electromagnetic propagation regulation, conventional range-based methods can convert the received signal strength indicator (RSSI), time-of-flight, and phase information to the distance between the reader antenna and the target tag. The geometric characteristics infer the position of the target tag. Ma et al. (2016) increased the accuracy to 0.769 m using the phase of arrival (POA) and time of arrival (TOA) modeling. Povalac and Sebesta (2011) achieved an average absolute error of 0.14 m by improving the phase-based ranging technology.

Subsequently, methods for localization based on reference tags and fingerprinting, which are more suitable for complex environments, have been extensively investigated. Ni et al. (2003) proposed a LANDMARC system based on reference tags, which involved deploying numerous fixed-position reference tags, reading RSSI values relative to target tags, and calculating target tag positions using k-nearest neighbor (KNN). Xu et al. (2017) improved the LANDMARC system (Ni et al., 2003), achieving an average positioning error of less than 15 cm. Zhao et al. (2017) and Mo and Li (2019) proposed improved methods based on virtual reference tags and phase variations that improved localization accuracy to within 10 cm.

However, the environment in massive farms is highly complex, with issues such as temperature and humidity fluctuations, signal masking, and target tag rotation. Most methods previously discussed are based on sophisticated computational models and require stringent experimental conditions. Moreover, they lack real-time performance and robustness. Therefore, this study proposes a machine learning-based localization method to achieve accurate localization in a complex cage environment using a relatively simple experimental design. With the aid of UHF-RFID devices and our method, it is possible to precisely locate multiple chickens in metal cages. To the best of our knowledge, no other study has used the same device to achieve localization accuracy comparable to that of our method in a similar environment.