Earticle

Animal detection and classification are crucial for effective wildlife management (WM) and reducing risks associated with animals related road accidents and attacks. Previous attempts trained the models using imbalanced data with fewer representative features and baseline models without improvement. This paper presents a new dataset of five animal classes captured in various poses, lighting conditions, and intraclass variations. The standard coupled detection head of the YoloV4 algorithm faces limitations when performing simultaneous classification and localization due to shared parameters and inputs. To address this issue, we propose a decoupled detection head (DDH) that handles these tasks separately, improving performance. We conducted extensive experiments using the proposed dataset. We found that the optimal backbone features marginally improve the performance of the modified network compared to state-of-the-art (SOTA) works in the subject domain. Our work contributes by addressing the limitations of the standard YoloV4 algorithm and proposing a new dataset for researchers to use in future studies.