International Journal of Science, Engineering and TechnologyAuthors: Gabhale Ankita Dattatray, Lawande Sayali Prabhakar, Mahabare Pratham Ravindra, Prof. A. P. Banger, Prof. K. D. Dere
Abstract: Human–wildlife conflict in forest-border regions has emerged as a critical concern for both rural communities and conservation agencies, with increasing incidents of large carnivores such as tigers, leopards, and hyenas entering human settlements. Conventional surveillance methods—watchtowers, perimeter fencing, and passive CCTV monitoring—rely heavily on continuous human observation and are therefore prone to delayed responses, fatigue-induced errors, and high operational cost. This paper presents the design, implementation, and experimental evaluation of an AI powered Animal Intrusion Detection System (AniDet) that integrates Internet of Things (IoT) hardware, embedded processing, and a deep-learning vision pipeline. The prototype employs six RCWL-0516 Doppler radar modules to provide 360° motion coverage, an MLX90614 non-contact infrared (IR) thermometer to confirm the presence of a warm body, and an ESP32- CAM module that streams JPEG frames over Wi-Fi/GSM to a Flask backend running a YOLOv8m classifier trained on a custom three-class dataset (tiger, leopard, hyena). On a held-out test set of 200 frames, the system achieved 92.0 % overall classification accuracy, mAP@0.5 of 0.93, and a mean detection confidence of 0.81 across the three target classes. End-to-end latency from motion to alert was measured at ≈ 1.68 s, while the multi-stage sensor fusion (radar + thermal + vision) reduced false alarms from 38 % (radar-only) to 6 %. The complete system operates on a 6 W solar–TP4056–Li-ion subsystem, demonstrating its suitability for off-grid deployment in forest-fringe regions.
DOI: http://doi.org/10.5281/zenodo.20176466
