Dog behaviour classification with movement sensors placed on the harness and the collar

Unsupervised Domain Adaptation for Mitigating Sensor Variability and Interspecies Heterogeneity in Animal Activity Recognition

Animal activity recognition (AAR) using wearable sensor data has gained significant attention due to its applications in monitoring and understanding animal behavior. However, two major challenges hinder the development of robust AAR models: domain variability and the difficulty of obtaining labeled datasets. To address this issue, this study intensively investigates the impact of unsupervised domain adaptation (UDA) for AAR. We compared three distinct types of UDA techniques: minimizing divergence-based, adversarial-based, and reconstruction-based approaches. By leveraging UDA, AAR classifiers enable the model to learn domain-invariant features, allowing classifiers trained on the source domain to perform well on the target domain without labels. We evaluated the effectiveness of UDA techniques using dog movement sensor data and additional data from horses. The application of UDA across sensor positions (neck and back), sizes (middle-sized and large-sized), and gender (female and male) within the dog data, as well as across species (dog and horses), exhibits significant improvements in the classification performance and reduced the domain discrepancy. The results highlight the potential of UDA to mitigate the domain shift and enhance AAR in various settings and for different animal species, providing valuable insights for practical applications in real-world scenarios where labeled data is scarce.

The Use of Triaxial Accelerometers and Machine Learning Algorithms for Behavioural Identification in Domestic Cats (Felis catus): A Validation Study

Animal behaviour can be an indicator of health and welfare. Monitoring behaviour through visual observation is labour-intensive and there is a risk of missing infrequent behaviours. Twelve healthy domestic shorthair cats were fitted with triaxial accelerometers mounted on a collar and harness. Over seven days, accelerometer and video footage were collected simultaneously. Identifier variables (n = 32) were calculated from the accelerometer data and summarized into 1 s epochs. Twenty-four behaviours were annotated from the video recordings and aligned with the summarised accelerometer data. Models were created using random forest (RF) and supervised self-organizing map (SOM) machine learning techniques for each mounting location. Multiple modelling rounds were run to select and merge behaviours based on performance values. All models were then tested on a validation accelerometer dataset from the same twelve cats to identify behaviours. The frequency of behaviours was calculated and compared using Dirichlet regression. Despite the SOM models having higher Kappa (>95%) and overall accuracy (>95%) compared with the RF models (64-76% and 70-86%, respectively), the RF models predicted behaviours more consistently between mounting locations. These results indicate that triaxial accelerometers can identify cat specific behaviours.

Machine learning based canine posture estimation using inertial data

The aim of this study was to design a new canine posture estimation system specifically for working dogs. The system was composed of Inertial Measurement Units (IMUs) that are commercially available, and a supervised learning algorithm which was developed for different behaviours. Three IMUs, each containing a 3-axis accelerometer, gyroscope, and magnetometer, were attached to the dogs' chest, back, and neck. To build and test the model, data were collected during a video-recorded behaviour test where the trainee assistance dogs performed static postures (standing, sitting, lying down) and dynamic activities (walking, body shake). Advanced feature extraction techniques were employed for the first time in this field, including statistical, temporal, and spectral methods. The most important features for posture prediction were chosen using Select K Best with ANOVA F-value. The individual contributions of each IMU, sensor, and feature type were analysed using Select K Best scores and Random Forest feature importance. Results showed that the back and chest IMUs were more important than the neck IMU, and the accelerometers were more important than the gyroscopes. The addition of IMUs to the chest and back of dog harnesses is recommended to improve performance. Additionally, statistical and temporal feature domains were more important than spectral feature domains. Three novel cascade arrangements of Random Forest and Isolation Forest were fitted to the dataset. The best classifier achieved an f1-macro of 0.83 and an f1-weighted of 0.90 for the prediction of the five postures, demonstrating a better performance than previous studies. These results were attributed to the data collection methodology (number of subjects and observations, multiple IMUs, use of common working dog breeds) and novel machine learning techniques (advanced feature extraction, feature selection and modelling arrangements) employed. The dataset and code used are publicly available on Mendeley Data and GitHub, respectively.

Margined Horn-Shaped Air Chamber for Body-Conduction Microphone

The sound amplification ratios of sealed air chambers with different shapes were quantitatively compared to design a body-conduction microphone to measure animal scratching sounds. Recently, quantitative monitoring of scratching intensity in dogs has been required. We have already developed a collar with a body-conduction microphone to measure body-conducted scratching sounds. However, the air chamber, one of the components of the body-conduction microphone, has not been appropriately designed. This study compared the amplification ratios of air chambers with different shapes through numerical analysis and experiments. According to the results, the horn-shaped air chamber achieved the highest amplification performance, at least for sound frequencies below 3 kHz. The simulated amplification ratio of the horn-shaped air chamber with a 1 mm height and a 15 mm diameter was 52.5 dB. The deformation of the bottom of the air chamber affected the amplification ratio. Adjusting the margin of the margined horn shape could maintain its amplification ratio at any pressing force. The simulated and experimental amplification ratios of the margined horn-shaped air chamber were 53.4 dB and 19.4 dB, respectively.

Video Validation of Tri-Axial Accelerometer for Monitoring Zoo-Housed Tamandua tetradactyla Activity Patterns in Response to Changes in Husbandry Conditions

Accelerometers are a technology that is increasingly used in the evaluation of animal behaviour. A tri-axial accelerometer attached to a vest was used on Tamandua tetradactyla individuals (n = 10) at Biodiversity Park. First, the influence of using a vest on the animals’ behaviour was evaluated (ABA-type: A1 and A2, without a vest; B, with a vest; each stage lasted 24 h), and no changes were detected. Second, their behaviour was monitored using videos and the accelerometer simultaneously (experimental room, 20 min). The observed behaviours were correlated with the accelerometer data, and summary measures (X, Y and Z axes) were obtained. Additionally, the overall dynamic body acceleration was calculated, determining a threshold to discriminate activity/inactivity events (variance = 0.0055). Then, based on a 24 h complementary test (video sampling every 5 min), the sensitivity (85.91%) and precision (100%) of the accelerometer were assessed. Animals were exposed to an ABA-type experimental design: A1 and A2: complex enclosure; B: decreased complexity (each stage lasted 24 h). An increase in total activity (%) was revealed using the accelerometer (26.15 ± 1.50, 29.29 ± 2.25, and 35.36 ± 3.15, respectively). Similar activity levels were detected using video analysis. The results demonstrate that the use of the accelerometer is reliable to determine the activity. Considering that the zoo-housed lesser anteaters exhibit a cathemeral activity pattern, this study contributes to easily monitoring their activities and responses to different management procedures supporting welfare programs, as well as ex situ conservation.

Dog–Owner Relationship, Owner Interpretations and Dog Personality Are Connected with the Emotional Reactivity of Dogs

Simple Summary

The relationship between owner and the dog affects the dog’s attachment behaviors and stress coping. In turn, the quality of the relationship may affect owner’s interpretations about their dog’s behavior. Here, we assessed dogs’ emotional responses from heart rate variability and behavioral changes during five different situations. Dog owners evaluated the emotion (valence and arousal) of their dog after each situation. We found that both negative and positive incidents provoked signs of emotional arousal in dogs. Owners detected the dog’s arousal especially during fear- and stress-evoking situations. The dog–owner relationship did not affect owners’ interpretation of dogs’ emotion. However, the dog–owner relationship was reflected in the dog’s emotional reactions. Close emotional bond with the owner appeared to decrease the arousal of the dogs. Dog owners’ frequent caregiving of their dog was associated with increased attachment behaviors and heightened arousal of dogs. Owners rated the disadvantages of the dog relationship higher for the dogs that were less owner-oriented and less arousable. Dog’s arousal may provoke dog’s need to seek human attention, which in turn may promote the development of emotional bond.

Abstract

We evaluated the effect of the dog–owner relationship on dogs’ emotional reactivity, quantified with heart rate variability (HRV), behavioral changes, physical activity and dog owner interpretations. Twenty nine adult dogs encountered five different emotional situations (i.e., stroking, a feeding toy, separation from the owner, reunion with the owner, a sudden appearance of a novel object). The results showed that both negative and positive situations provoked signs of heightened arousal in dogs. During negative situations, owners’ ratings about the heightened emotional arousal correlated with lower HRV, higher physical activity and more behaviors that typically index arousal and fear. The three factors of The Monash Dog–Owner Relationship Scale (MDORS) were reflected in the dogs’ heart rate variability and behaviors: the Emotional Closeness factor was related to increased HRV (p = 0.009), suggesting this aspect is associated with the secure base effect, and the Shared Activities factor showed a trend toward lower HRV (p = 0.067) along with more owner-directed behaviors reflecting attachment related arousal. In contrast, the Perceived Costs factor was related to higher HRV (p = 0.009) along with less fear and less owner-directed behaviors, which may reflect the dog’s more independent personality. In conclusion, dogs’ emotional reactivity and the dog–owner relationship modulate each other, depending on the aspect of the relationship and dogs’ individual responsivity.

Dog Behavior Recognition Based on Multimodal Data from a Camera and Wearable Device

Although various studies on monitoring dog behavior have been conducted, methods that can minimize or compensate data noise are required. This paper proposes multimodal data-based dog behavior recognition that fuses video and sensor data using a camera and a wearable device. The video data represent the moving area of dogs to detect the dogs. The sensor data represent the movement of the dogs and extract features that affect dog behavior recognition. Seven types of behavior recognition were conducted, and the results of the two data types were used to recognize the dog’s behavior through a fusion model based on deep learning. Experimentation determined that, among FasterRCNN, YOLOv3, and YOLOv4, the object detection rate and behavior recognition accuracy were the highest when YOLOv4 was used. In addition, the sensor data showed the best performance when all statistical features were selected. Finally, it was confirmed that the performance of multimodal data-based fusion models was improved over that of single data-based models and that the CNN-LSTM-based model had the best performance. The method presented in this study can be applied for dog treatment or health monitoring, and it is expected to provide a simple way to estimate the amount of activity.

Description of movement sensor dataset for dog behavior classification

Movement sensor data from seven static and dynamic dog behaviors (sitting, standing, lying down, trotting, walking, playing, and (treat) searching i.e. sniffing) was collected from 45 middle to large sized dogs with six degree-of-freedom movement sensors attached to the collar and the harness. With 17 dogs the collection procedure was repeated. The duration of each of the seven behaviors was approximately three minutes. The order of the tasks was varied between the dogs and the two repetitions (for the 17 dogs). The behaviors were annotated post-hoc based on the video recordings made with two camcorders during the tests with one second resolution. The annotations were accurately synchronized with the raw movement sensors data. The annotated data was originally used for training behavior classification machine learning algorithms for classifying the seven behaviors. The developed signal processing and classification algorithms are provided together with the raw measurement data and reference annotations. The description and results of the original investigation that the dataset relates to are found in: P. Kumpulainen, A. Valldeoriola Cardó, S. Somppi, H. Törnqvist, H. Väätäjä, P. Majaranta, Y. Gizatdinova, C. Hoog Antink, V. Surakka, M. V. Kujala, O. Vainio, A. Vehkaoja, Dog behavior classification with movement sensors placed on the harness and the collar, Applied Animal behavior Science, 241 (2021), 105,393.

Deep Learning Empowered Wearable-Based Behavior Recognition for Search and Rescue Dogs

Search and Rescue (SaR) dogs are important assets in the hands of first responders, as they have the ability to locate the victim even in cases where the vision and or the sound is limited, due to their inherent talents in olfactory and auditory senses. In this work, we propose a deep-learning-assisted implementation incorporating a wearable device, a base station, a mobile application, and a cloud-based infrastructure that can first monitor in real-time the activity, the audio signals, and the location of a SaR dog, and second, recognize and alert the rescuing team whenever the SaR dog spots a victim. For this purpose, we employed deep Convolutional Neural Networks (CNN) both for the activity recognition and the sound classification, which are trained using data from inertial sensors, such as 3-axial accelerometer and gyroscope and from the wearable’s microphone, respectively. The developed deep learning models were deployed on the wearable device, while the overall proposed implementation was validated in two discrete search and rescue scenarios, managing to successfully spot the victim (i.e., obtained F1-score more than 99%) and inform the rescue team in real-time for both scenarios.

Cross-Modality Interaction Network for Equine Activity Recognition Using Imbalanced Multi-Modal Data

With the recent advances in deep learning, wearable sensors have increasingly been used in automated animal activity recognition. However, there are two major challenges in improving recognition performance—multi-modal feature fusion and imbalanced data modeling. In this study, to improve classification performance for equine activities while tackling these two challenges, we developed a cross-modality interaction network (CMI-Net) involving a dual convolution neural network architecture and a cross-modality interaction module (CMIM). The CMIM adaptively recalibrated the temporal- and axis-wise features in each modality by leveraging multi-modal information to achieve deep intermodality interaction. A class-balanced (CB) focal loss was adopted to supervise the training of CMI-Net to alleviate the class imbalance problem. Motion data was acquired from six neck-attached inertial measurement units from six horses. The CMI-Net was trained and verified with leave-one-out cross-validation. The results demonstrated that our CMI-Net outperformed the existing algorithms with high precision (79.74%), recall (79.57%), F1-score (79.02%), and accuracy (93.37%). The adoption of CB focal loss improved the performance of CMI-Net, with increases of 2.76%, 4.16%, and 3.92% in precision, recall, and F1-score, respectively. In conclusion, CMI-Net and CB focal loss effectively enhanced the equine activity classification performance using imbalanced multi-modal sensor data.