Adapting YOLOv5 for Detection of Central European Wildlife

Wildlife detection is crucial for enhancing vehicle safety and accurate detection systems can help prevent wildlife-vehicle collisions and improve safety. Additionally, these systems provide valuable data for tracking wildlife movement patterns or projections that could significantly improve the effectiveness of wildlife detection system. The model was trained on a diverse dataset compiled from multiple sources, including COCO 2017, FLIR, and wildlife-specific datasets such as LILA BC, NTLNP, and LSOTB-TIR. A major challenge was the lack of annotated data for wildlife classes. To overcome this, the Grounding DINO framework was used for automatic annotation, with clustering techniques applied to remove incorrectly annotated images. Various training strategies were tested, including training without augmentation, using default parameters, applying weighted loss functions with and without high-resolution training, tuning hyperparameters, and implementing image resampling and downsampling techniques. The YOLOv5l model trained with a weighted loss function at high resolution showed the best performance, achieving the highest mean Average Precision (mAP), precision, and recall scores. While the model successfully detected some wildlife classes, it struggled with others. The reasons for these limitations were analyzed. Additionally, the performance of YOLOv5 was compared to the newer YOLOv8 model the results were discussed. This research aims to contribute to the development of automated wildlife detection systems, which have the potential to improve road safety. However, further improvements are needed to address the model’s limitations in detecting certain wildlife classes.