AUTOMATED DETECTION OF LARGE ANIMALS IN ROAD SCENE ENVIRONMENTS USING DEEP LEARNING

Automated detection of large animals in road scenes plays a crucial role in enhancing the safety of autonomous vehicles, particularly in regions where wildlife-related accidents are common. This paper introduces a deep learning-based explanation for detecting and classifying ten large animal classes within road scene environments, such as dogs, horses, cows, and bears. A specialized dataset was fetched using selected classes from the COCO and Open Images V5 datasets, annotated in the COCO format. Four advanced object detection models were trained and evaluated with the EfficientDet-D1, RetinaNet R-50-FPN, Faster R-CNN R-50-FPN, and Cascade R-CNN R-50-FPN. Results show that RetinaNet R-50-FPN achieved the highest mean Average Precision (mAP) of 0.83 for one joint class and 0.69 for ten classes while also delivering the fastest inference speed at 50.6 FPS for one-class detection and 45.2 FPS for multi-class detection. EfficientDet-D1 achieved a mAP of 0.89 for one joint class and 0.77 for ten classes,offering competitive performance but with slightly slower inference speeds. The findings highlight RetinaNet as the most effective and efficient model for real-time large animal detection in road scenes, offering significant potential for integration into modern autonomous driving systems.