Journal of Experimental Agriculture International

Animal intrusion in agricultural fields can contribute to crop damage and financial loss, while many conventional deterrent methods are labour-intensive, inconsistent or unsuitable for non-lethal field protection. This study evaluated a low-cost image-detection system based on ESP32-CAM and a lightweight neural-network model for detecting selected animals under field conditions. Images of target animals were collected from different angles and views, annotated using bounding boxes and used to train a FOMO (Faster Objects, More Objects) MobileNetV2 0.35 model. The trained model was exported to the ESP32-CAM platform and tested for real-time detection of cows, elephants and deer. During field deployment, the system captured images, processed them on the embedded platform and generated detection outputs with confidence values and bounding-box coordinates. The detector identified cows in all ten trials, with confidence values ranging from 57.3% to 95.3%. For elephants and deer, performance was more variable, with several zero-detection cases, indicating sensitivity to lighting, camera angle, object distance, background complexity and partial occlusion. The system also triggered an LED alert following detection, demonstrating its practical potential for automated field monitoring. Overall, the results indicate that ESP32-CAM integrated with a lightweight edge-AI model can support affordable real-time animal-intrusion detection, although further improvement in dataset diversity, environmental robustness and model optimisation is required before wider deployment.