Peer-reviewed scientific articles describing the use of sensors for detecting lying, rest and sleep behavior in healthy calves, defined as bovines with less than 6 months of age, were eligible for inclusion. Literature search was conducted through Scopus and Web of Science on 13 November 2023, to include articles written in English, with full text available and based upon original data.

Since this critical evaluation focused on the use of sensors to detect lying, rest or sleep behavior in calves, all studies evaluating other animal species or adult bovines, as also the alterations of these behaviors depending on management practices or the presence of disease in calves were excluded.

The search was conducted applying the following search terms: (calf OR calves OR dairy cal* OR veal cal* OR young cattle) and (sensor OR accelerometer OR activity sensor OR artificial intelligence OR computer vision OR non-invasive technology OR logger OR machine learning) and (rest* OR sleep* OR lying). The selection of these search terms was based on initial screening of relevant articles to gain general background information and expert opinion. The terms were searched within article title, abstract and keywords.

All papers obtained from the database searches were exported into a Microsoft Excel file, and a first cleaning of the dataset was conducted to remove papers that did not report the authors’ names or abstract and to maintain only peer-reviewed papers. Before screening, all duplicates were removed. The screening process was carried out by the first author (G.P.R.). Firstly, any record with a title that clearly did not fit the eligibility criteria was excluded. The remaining papers were screened based on the abstract, and subsequently on the full text. The information regarding animal age and weight, type of sensor and site of application (when wearable), type of behavior recorded, sampling interval and other measures taken for the study, were tracked and tabulated.

Among the 16 papers included, 10 articles dealt with the use of tri-axial accelerometers. This type of sensor measures the acceleration forces towards the three axes (x, y, z) which occurs during animal movement at determined sampling intervals, and then statistical analysis or algorithm were applied to translate the measurements into posture and behavior output files.

Seven papers (20, 21, 31–35) quantified the time spent lying (both in percentage of the day and as the duration in minutes), referring to the sole posture of the animal, and quantifying the so-called “bout,” defined as the transition between lying and standing, also estimated in terms of mean duration and frequency in the considered timespan. We can notice that this aspect defines only the posture of animal and is generally linked to a state of inactivity or rest. It is therefore indirectly related to resting behavior and sleep. For this kind of evaluation, accelerometers were placed on the right hind leg as a pedometer (20, 21, 31, 33, 35), or on the ear as an ear tag (32, 34).

Temporal distribution of resting activity was deduced by the measurement of total locomotor behavior recorded by activity data logger in the work of Giannetto et al. (36).

On the other hand, two papers focused on the specific evaluation of sleep (37, 38) and its stages (37). As previously reported by Hänninen et al. (28), in fact, sleeping behavior and its stages can be defined as follows: NREM sleep when the calf is resting head up, being still, and REM sleep when the calf is resting neck relaxed, with the head against the floor or flank. Considering the implication of head and neck in sleep evaluation, accelerometers were placed, respectively, on the middle occipital using a halter and on a collar in the works of Hokkanen et al. (37) and Fukasawa (38).

Within the reviewed studies, sensors could be differentiated based on their applications: specifically, sensors can be categorized into those with research features (R), such as customizable sampling frequency and the ability to download and analyze data, and those designed for on-farm use with real-time software applications, referred to as ‘commercially available sensors’ (C).

In particular, considering the different commercial accelerometers found in literature search, it must be noted that none of them was purposely designed for calves: it was indeed used a sensor which had been previously validated on cows. This implies that the features of the sensor are relevant for their practical use on calves: when considering pedometer and ear tag, in fact, sensors must be light and little enough, not to cause disturb to animals, both for their welfare, but also for the quality of data acquired from technology.

The distinction based on the application of accelerometers also relates to real-time availability of data and the necessity of subsequent data elaboration (R) or automatic algorithm application (C).

Hobo Pendant G logger (R) was the most reported sensor in literature (4 papers out of 10): it was positioned as pedometer and validated for recording lying behavior of unweaned calves considering different anatomical site (i.e., forelimbs and hind limbs (20), but also considering a different environmental condition [i.e., validation in tropical weather (35)]. It was also positioned middle occipital on a halter and validated for measuring sleep-like postures as indicator of sleep behaviors (38). Moreover, it was also considered in the work of Swartz et al. (21), its data for measuring lying behavior were also used to compare and validate the sensor object of the paper (i.e., Afitag II). In all the aforementioned cases sampling frequency was decided by the authors and after the experimental period, data were downloaded to a computer using HOBOware Pro Software and subsequently elaborated with statistical programs (i.e., SPSS, R, SAS, and Excel).

On the contrary, data from C sensors found through literature search (namely Afitag II, IceQube, Icetag, Actiwatch Mini, SmartBow and SensOor) were automatically elaborated through private algorithm (unpublished), resulting in a potentially prompt instrument for on-farm application, while fewer information are available about how behavioral quantification was carried out. However, all the above-mentioned sensors were validated with visual observation or video recordings. Considering that raw data are not available and summarized at intervals of 15 min (as for Afitag II and IceQube), validation was made based on the sum of data, aligned with the timing of video recordings or visual observations.

Literature search also resulted in 2 papers (39, 40) focused on the approach and the algorithm used to classify and quantify multiple behaviors using accelerometer data, rather than focusing on the sole sensor and on one type of behavior. This implies a rising interest in the integration of different types of behaviors, using algorithms with the attempt of describing the complexity of animal behavior (also considering the implication of its alteration). Lying behavior was considered and most represented, and one of the two papers (40) also evaluated aspects of active and non-active lying with high levels of accuracy (90.38% both) although active lying had the worst performance in terms of sensitivity (64%) and precision (69%), probably because it was confused with similar behaviors as non-active lying or ruminating. Providing more detailed information about the level of activity performed while lying and what the calf is actually doing, is a largely unexplored aspect considering the results of the present review, and a reason might be linked to the necessity of finding sufficiently sensitive sensors.

Sturm et al. (39) applied a chaos theoretic approach to animal activity recognition: animal behavior was considered as a nonlinear dynamical system and time series derived from ear-tags with tri-axial accelerometers [Smartbow, previously validated by the same authors in the work of Roland et al. (34)] were used to associate two main mutually exclusive behaviors, namely, lying and standing/walking with six possible activities: feeding, drinking water, drinking milk, playing, rumination, and neutral behavior.

Carslake et al. (40) developed an algorithm to identify postures (namely lying and standing) and several behaviors (i.e., active lying, non-active lying, locomotor play, self-grooming, ruminating, non-nutritive and nutritive suckling) in 13 pre-weaned dairy calves with collar-mounted accelerometers (Sparkfun 9, research-based). In particular, a quantification algorithm for predicting behavior distribution was developed and validated. This approach, referred by authors as mostly ignored in precision livestock, has the advantage of showing high accuracy with relatively low overestimation in unseen real-world data despite very low behavior prevalence. The identification of rare behaviors might be of strong interest, considering its use for disease and welfare prediction and quantification.

Table 2 presents detailed information about the accelerometers used in tri-axial accelerometer studies and machine learning studies involving accelerometer data. It includes specifics on the sensor’s placement, its application, the type of behavior or posture recorded and the related behavioral measures, as well as data sampling frequency and elaboration.

Two papers (41, 42) evaluated the use of computer vision technology: through this method images or video information can be analyzed to recognize and classify specific animal behaviors or postures. It must be noted that both studies evaluated the use of this technology in individually housed calves, so new perspectives in this field might be the evaluation of group-housed calves’ behaviors, in particular considering the resting pattern and social behavior of these animals.

Guo et al. (41) evaluated a new method (i.e., Integrated Background Model), built by combining background-subtraction and inter-frame difference methods to monitor the behaviors of the dairy calf. By using the new model and motion characteristics of the calf in different areas of the enclosure, the authors successfully identified the behaviors of entering the resting area, leaving the resting area, remaining stationary, turning around, feeding and drinking.

Tung et al. (42) developed a deep learning algorithm for calf posture recognition, in order to classify whether the calf is standing or lying based on images collected with cameras in two different positions, above the calf.

One study (43) was found through literature search for behavior identification with the use of wearable cameras, a method where the camera is attached to the animals and moves with them, circumventing identification problems and allowing a closer look for targeted behavior. This study aimed to verify if the images obtained from wearable cameras can accurately record the behavior of calves, in order to use the videos for automatic analyses using AI in future. The wearable camera was placed in a protective case and fixed to the calf’s right cheek with a commercially available calf halter. Postures such as standing and lying and behaviors such as feeding and rumination could be observed as accurately as through direct observations.

Ueda et al. (44) assessed the usefulness of a commercially available indoor positioning system for monitoring the resting time and moving distance in group-housed dairy calves. The method predicted lying time using the recorded displacement by IPS, a commercially available real-time locating system, including a tag (transmitter), locator (receiver), and data acquisition and processing software.

The application of accelerometers encompasses several critical considerations. Firstly, memory limitations are a significant factor, particularly when contrasting data loggers with wireless data acquisition systems. Data loggers possess restricted memory capacity, whereas wireless systems can transmit data in real-time and store a substantially larger volume of information.

Another relevant aspect is the impact of sampling frequencies on the accuracy of actigraphy measures. Higher sampling frequencies can provide more precise insights into specific behaviors, such as sleep, compared to more general behaviors, such as lying down. It is also essential for researchers to filter the data collected by accelerometers. This process of filtering facilitates the removal of potentially erroneous readings, thereby enhancing the overall accuracy of the measurements.

The tolerance of wearable devices is another crucial consideration. These devices must be accepted by animals without causing discomfort. For instance, devices like pedometers should not cause signs of distress, whereas ear tags were described as subject to this occurrence, requiring corrective actions.

Finally, features derived from both accelerometers and gyroscopes are indispensable for achieving high levels of accuracy in the classification of behavioral data. The integration of these features markedly improves the precision of measurements, as highlighted by Carslake et al. (40).

Providing more detailed information about the level of activity performed while lying down and understanding the specific actions of the calf through data integration remains a largely unexplored area. This gap may be attributable to the challenge of identifying sufficiently sensitive sensors.

Counting behaviors that do not frequently occur based on a prediction classifier can result in overestimation. This issue is highlighted by the Classify and Count Method [as noted by Carslake et al. (40)], which fails to account for the fact that the positive predictive value decreases with prevalence. Additionally, there may be discrepancies in behavior prevalence between the training/test datasets and a new unlabeled dataset, which further complicates the accuracy of predictions.

Moreover, there is a risk of overfitting, where a model performs well on training data but fails to generalize to new data, potentially compromising the validity of the machine learning model.

Issues encountered in recognizing calves’ behaviors were attributed to their darker images and the calf’s black and white coat. Detection challenges arose, particularly when the calf’s image overlapped with the area defined as resting zone. In their study, Guo et al. (41) utilized the average of 10 consecutive frames to derive characteristic values for behavior recognition. Consequently, instances of static behavior were occasionally misclassified as entering or leaving the resting area when the calf transitioned from a stationary state.

Moreover, the analysis of models trained on images from different cameras underscores the critical role of image quality. The findings of Tung et al. (42) emphasize that high-quality images are indispensable for enabling deep learning models to learn and accurately predict the distinctive features of calf postures.

Wearable cameras study had some critical outcomes related to the risk of entrapment (43), as the halter expands around the cameras. Therefore, considering their welfare, calves need to be regularly welfare-checked by an individual.

Ueda et al. (44) underline that while Integrated Positioning Systems (IPS) demonstrated efficacy in predicting resting time and movement distance, there is a noted need for improved accuracy in the prediction of lying time. Secondly, the effectiveness of IPS has yet to be validated in large-scale dairy farming operations with group-housed calves, necessitating further research in such settings.