Enhanced Yolov8 network with Extended Kalman Filter for wildlife detection and tracking in complex environments

Abstract

Amid a growing global focus on ecological conservation and biodiversity monitoring, the efficient identification and tracking of wildlife are essential for environmental research, wildlife protection, and habitat management. Nevertheless, intricate landscapes, varied animal sizes, and obstructions obstruct wildlife detection and tracking. This study introduces the wilDT-YOLOv8n model, specifically engineered for the effective identification and tracking of animals. Initially, the Stable Diffusion model augments the dataset, establishing a basis for training data. Subsequently, enhancements to the Yolov8n model are implemented through the incorporation of the deformable convolutional network DCNv3 and the utilization of the C2f_DCNV3 layer to augment feature extraction efficacy, while addressing detection challenges associated with small targets and intricate backgrounds by integrating the EMGA attention mechanism and the ASPFC feature fusion module. Enhancing the Extended Kalman Filter algorithm guarantees reliable and precise tracking. The research findings reveal that the wilDT-YOLOv8n model attained an average detection accuracy (mAP50) of 88.54 % on the custom dataset, reflecting a 4.57 % enhancement over the original YOLOv8n model; the refined Extended Kalman Filter realizes a Multi-Object Tracking Accuracy (MOTA) of 40.35 %, representing a 3.923 % advancement over the original Kalman Filter. The results indicate the feasibility of accurately detecting and monitoring wildlife in intricate environments, offering significant insights for ecological research and biodiversity conservation, and aiding in the protection of endangered species.