How to achieve accurate wildlife detection by using vehicle‐mounted mobile monitoring images and deep learning?

Abstract

With the advancement of artificial intelligence (AI) technologies, vehicle‐mounted mobile monitoring systems have become increasingly integrated into wildlife monitoring practices. However, images captured through these systems often present challenges such as low resolution, small target sizes, and partial occlusions. Consequently, detecting animal targets using conventional deep‐learning networks is challenging. To address these challenges, this paper presents an enhanced YOLOv7 model, referred to as YOLOv7(sr‐sm), which incorporates a super‐resolution (SR) reconstruction module and a small object optimization module. The YOLOv7(sr‐sm) model introduces a super‐resolution reconstruction module that leverages generative adversarial networks (GANs) to reconstruct high‐resolution details from blurry animal images. Additionally, an attention mechanism is integrated into the Neck and Head of YOLOv7 to form a small object optimization module, which enhances the model's ability to detect and locate densely packed small targets. Using a vehicle‐mounted mobile monitoring system, images of four wildlife taxa—sheep, birds, deer, and antelope —were captured on the Tibetan Plateau. These images were combined with publicly available high‐resolution wildlife photographs to create a wildlife test dataset. Experiments were conducted on this dataset, comparing the YOLOv7(sr‐sm) model with eight popular object detection models. The results demonstrate significant improvements in precision, recall, and mean Average Precision (mAP), with YOLOv7(sr‐sm) achieving 93.9%, 92.1%, and 92.3%, respectively. Furthermore, compared to the newly released YOLOv8l model, YOLOv7(sr‐sm) outperforms it by 9.3%, 2.1%, and 4.5% in these three metrics while also exhibiting superior parameter efficiency and higher inference speeds. The YOLOv7(sr‐sm) model architecture can accurately locate and identify blurry animal targets in vehicle‐mounted monitoring images, serving as a reliable tool for animal identification and counting in mobile monitoring systems. These findings provide significant technological support for the application of intelligent monitoring techniques in biodiversity conservation efforts.