An Embedded AI System for Pet Activity Recognition and Nutrition Management

This paper introduces an embedded, low-power artificial intelligence (AI) system designed for real-time pet activity recognition and personalized nutritional guidance. Utilizing a 3-axis accelerometer-based inertial measurement unit (IMU), the system applies frequency-domain features extracted via a 512-point FFT combined with an enhanced multilayer perceptron (MLP) deployed on a Seeed XIAO nRF52840 microcontroller. The inference results are categorized into predefined activity levels and communicated in real time via Bluetooth Low Energy (BLE) to a mobile application. Local data backup is provided on a microSD card. GPS-derived speed data provide reliable ground-truth labeling for objective and scalable model training. Following American Animal Hospital Association (AAHA) guidelines, maintenance energy requirements (MER) are estimated using activity classification outcomes and individual pet profiles. The proposed architecture achieves high classification accuracy (94.3%), low inference latency (16 ms), and efficient resource utilization, making it particularly suited for practical integration into home-based care, veterinary clinical assessments, and pet insurance applications.