Implementation of Convolutional Neural Network to Minimize Wildlife-Vehicle Collisions

Wildlife-vehicle collisions (WVCs) have serious consequences for both humans and especially the wildlife that surrounds us. It is therefore imperative to address this issue in order to preserve the ecosystem. We propose a real-time animal detection model utilizing You Only Look Once (YOLO) v8, which is based on Convolutional Neural Network. The dataset for this research is obtained from the Roboflow and consists of 1,239 images of various wild animals captured by trail cameras. To facilitate the training and evaluation process, the dataset is divided into 866 images allocated for training purposes, 247 images for validation, and 126 images for testing. Post-training evaluation of the model demonstrates excellent performance in detecting animals in various scenarios, achieving mAP of 0.829 (at IoU of 0.5) and an impressive speed of approximately 133.99 frames per second. Although the results are suitable for real-time animal detection, further research may be required to ensure the successful execution of the model by training it over a more diverse dataset and hardware specifications.