Robust Deep Learning Models For Wildlife Detection: Adapting To Varying Environmental Conditions

Abstract-Wildlife conservation and the effective mitigation of human-wildlife conflicts heavily rely on the dependable detection and continuous monitoring of wildlife across a wide spectrum of ever-changing environments. In this study, we investigate the versatility of two deep learning models, RetinaNet-152 and DenseNet-201, in tackling the complexities presented by fluctuating environmental conditions in wildlife detection. RetinaNet-152, renowned for its depth and skip connections, is adept at capturing intricate features in images, making it an excellent candidate for scenarios with diverse species and habitats. On the other hand, DenseNet-201’s densely connected architecture enhances its adaptability to varying image quality, making it a robust choice for challenging field conditions. Our empirical experiments unveil the remarkable performance of both models under ideal circumstances. However, in demanding situations such as low-light conditions and adverse weather, DenseNet-201 consistently exhibits superior performance compared to RetinaNet-152.