Monitoring horses’ behaviors through sensors can yield important information about their health and welfare. Sampling frequency majorly affects the classification accuracy in activity recognition and energy needs for the sensor. The aim of this study was to evaluate the effect of sampling rate reduction of a tri-axial accelerometer on the recognition accuracy by resampling a 50 Hz experimental dataset to four lower sampling rates (5 Hz, 10 Hz, 12.5 Hz and 25 Hz). Also, in this work we investigate the ‘reality gap’ that incorporates changes in the data that are primarily characterized as sensor rotations or measurement noise through various data augmentation techniques such as rotation and jittering. Finally, another factor influencing activity recognition are the subjects themselves and therefore the model is evaluated on different horse types. A deep learning-based approach for activity detection of equines is proposed to automatically classify 2238 manually annotated 2 s samples tri-axial accelerometer leg data data of seven different activities performed by six different subjects. The raw data are preprocessed and fed into a convolutional neural network (CNN) from which features are extracted automatically by using strong computing capabilities. Furthermore, the neural network was intentionally designed to minimize running time, enabling us to imagine the future use of the built model in embedded constrained devices. The complexity of these automatic learning techniques can be decreased while achieving high accuracies using ten-fold-cross validation using a computationally less intensive received signal length data (99.32% at 5 Hz vs 99.74% at 25 Hz). This indicates that sampling at 5 Hz with a 2 s window will offer advantages for activity surveillance thanks to decreased energy requirements, since validation time decreases 16-fold (784 microseconds at 50 Hz to 48 microseconds at 5 Hz). Moreover, in this work we show that rotating the training or validation signal with 10 degrees over the X, Y and Z-axis increases the generalization capabilities of our model (99.61 % vs 99.93%) while adding small amounts of noise (smaller than 0.3 standard deviation (STD)) does not decrease the classification accuracy under 99%. Finally, the performance and ability of the model to generalize is validated on data from unseen horses at the cost of only 4.1% and 2.45% reduction in accuracy when validated on a pony and a lame horse, respectively.

Equine training activity detection will help to track and enhance the performance and fitness level of riders and their horses. Currently, the equestrian world is eager for a simple solution that goes beyond detecting basic gaits, yet current technologies fall short on the level of user friendliness and detection of main horse training activities. To this end, we collected leg accelerometer data of 14 well-trained horses during jumping and dressage trainings. For the first time, 6 jumping training and 25 advanced horse dressage activities are classified using specifically developed models based on a neural network. A jumping training could be classified with a high accuracy of 100 %, while a dressage training could be classified with an accuracy of 96.29%. Assigning the dressage movements to 11, 6 or 4 superclasses results in higher accuracies of 98.87%, 99.10% and 100%, respectively. Furthermore, during dressage training, the side of movement could be identified with an accuracy of 97.08%. In addition, a velocity estimation model was developed based on the measured velocities of seven horses performing the collected, working, and extended gaits during a dressage training. For the walk, trot, and canter paces, the velocities could be estimated accurately with a low root mean square error of 0.07 m/s, 0.14 m/s, and 0.42 m/s, respectively.

To detect behavioral anomalies (disease/injuries), 24 h monitoring of horses each day is increasingly important. To this end, recent advances in machine learning have used accelerometer data to improve the efficiency of practice sessions and for early detection of health problems. However, current devices are limited in operational lifetime due to the need to manually replace batteries. To remedy this, we investigated the possibilities to power the wireless radio with a vibrational piezoelectric energy harvester at the leg (or in the hoof) of the horse, allowing perpetual monitoring devices. This paper reports the average power that can be delivered to the node by energy harvesting for four different natural gaits of the horse: stand, walking, trot and canter, based on an existing model for a velocity-damped resonant generator (VDRG). To this end, 33 accelerometer datasets were collected over 4.5 h from six horses during different activities. Based on these measurements, a vibrational energy harvester model was calculated that can provide up to $64.04$ μW during the energetic canter gait, taking an energy conversion rate of 60% into account. Most energy is provided during canter in the forward direction of the horse. The downwards direction is less suitable for power harvesting. Additionally, different wireless technologies are considered to realize perpetual wireless data sensing. During horse training sessions, BLE allows continues data transmissions (one packet every 0.04 s during canter), whereas IEEE 802.15.4 and UWB technologies are better suited for continuous horse monitoring during less energetic states due to their lower sleep current.

Automatic horse gait classification offers insights into training intensity, but direct sensor attachment to horses raises concerns about discomfort, behavioral disruption, and entanglement risks. To address this, our study leverages rider-centric accelerometers for movement classification. The position of a sensor, sampling frequency, and window size of segmented signal data have a major impact on classification accuracy in activity recognition. Yet, there are no studies that have evaluated the effect of all these factors simultaneously using accelerometer data from four distinct rider locations (the knee, backbone, chest, and arm) across five riders and seven horses performing three gaits. A total of eight models were compared, and an LSTM-convolutional network (ConvLSTM2D) achieved the highest accuracy, with an average accuracy of 89.72% considering four movements (halt, walk, trot, and canter). The model performed best with an interval width of four seconds and a sampling frequency of 25 Hz. Additionally, an F1-score of 86.18% was achieved and validated using LOSOCV (Leave One Subject Out Cross-Validation).