Integration of Computer Vision AI (Camera + Image Processing) into an IoT Calving Detection Device

Abstract

Calving is one of the most critical events in dairy cattle farming. Complications during delivery, if undetected in time, lead to significant calf and cow mortality, substantial economic losses, and increased veterinary costs. Traditional monitoring methods, based on manual farm inspections, are labor-intensive and unreliable, particularly during nighttime hours when the majority of calvings occur. This memoir presents CalvingSafe, a low-cost, IoT-based intelligent system designed to automatically detect pre-calving behavioral patterns and deliver real-time alerts to farmers and veterinarians. The system integrates three MPU6050 sixaxis inertial measurement units mounted along the cow’s tail, connected through a TCA9548A I2C multiplexer to an ESP8266 NodeMCUWi-Fi microcontroller. Sensor data is transmitted via UDP over a localWi-Fi network to a Python backend, where it is processed through a sliding window of 20 consecutive readings and fed into a hybrid artificial intelligence engine. The AI engine combines a feedforward Artificial Neural Network (ANN), implemented as a two-hidden-layer MLPClassifier (128 × 64 neurons, tanh activation, threshold = 0.2, trained with SMOTE balancing on a Kaggle-published dataset), with a seven-input Mamdani Fuzzy Logic System (FLS) operating on eight handcrafted inference rules. A two-model voting mechanism with a 10-consecutive-frame temporal filter produces three confidence levels: HIGH, MEDIUM, and LOW, each mapped to a specific alert behavior. Detected events are pushed to a Firebase Realtime Database and reflected in real time on the CalvingSafewebdashboard, deployed on Vercel, which provides role-differentiated views for farmers and veterinarians. A three-channel alert system combining in-dashboard banners, Firebase event logging, and Blynk mobile push notifications ensures that no calving event goes unnoticed. System validation was conducted through a CSV simulation mode exercising the complete inference and cloud pipeline on a labeled IMU dataset. Results confirmed correct end-to-end behavior across all subsystems, with an estimated total pipeline latency below one second. The system demonstrates the feasibility of deploying an intelligent, low-cost calving detection solution adapted to Algerian farming conditions.