Abstract
This research studies dog behavior classification using inertial motion sensors and a transformer-based neural architecture. The model uses self-attention to capture temporal dependencies across accelerometer and gyroscope signals collected from back-mounted sensors, focusing on seven behaviors such as walking, trotting, galloping, sniffing, lying, sitting, and standing. The proposed encoder-only transformer architecture adds a lightweight classifier head to produce behavior probabilities from 2-second sensor windows. On an open-access dataset of motion data from 45 dogs, the method reaches 98.5% accuracy, outperforming time-series DNN baselines while reducing model size, computational complexity, and latency for resource-constrained deployment.