A Deep Learning Approach to Detect Wounded Animals

The escalating incidents of wildlife and stray animals being fatally injured have caused determination for an automatic and machine-driven approach to pan out poorly wounded animals to take on even more urgency. Traditional monitoring methods involving direct observation or manually analyzing camera trap images are very labor-intensive, timeconsuming, and prone to human error. This paper proposes an integrated deep learning-based solution for real-time detection of wounded animals using the YOLOv8 algorithm. The model was fine-tuned on a specifically composed dataset of images of healthy and injured animals from different species and environments. Data annotation was accomplished using Roboflow tools, whereby certain regions of interest, visible injuries in particular, were marked for improved detection accuracy. A strong training pipeline was put in place for all scenarios of varying illumination, angles, and occlusions for high precision, recall, and mean Average Precision (mAP) scores. The implementation of the trained model is through a user-friendly web application-based model created in Flask, which allows a user to upload images for analysis or for a live video. The system detects injuries automatically, marks the affected areas, and raises alerts for timely rescue and medical assistance, therefore helping ameliorate wildlife conservation.