Abstract

Wild Cervidae(deer and their relatives) play a crucial role in maintaining ecological balance and are integral components of ecosystems. However, factors such as environmental changes and poaching behaviors have resulted in habitat degradation for Cervidae. The protection of wild Cervidae has become urgent, and Cervidae monitoring is one of the key means to ensure the effectiveness of wild Cervidae protection. Object detection algorithms based on deep learning offer promising potential for automatically detecting and identifying animals. However, when those algorithms are used for inference in unseen background environments, there will be a significant decrease in accuracy, especially in the situation that a certain type of Cervidae images are collected from single scene for algorithm training. In this paper, a two-stage localization and classification pipeline for Cervidae monitoring is proposed. The pipeline effectively reduces background interference in Cervidae monitoring and enhances monitoring accuracy. In the first stage, the YOLOv7 network is designed to automatically locate Cervidae in UAV infrared images, while implementing improved bounding box regression through the α-IoU loss function enables the network to locate Cervidae more accurately. Then, Cevdidae objects are extracted to eliminate the background information. In the second stage, a classification network named CA-Hybrid, based on Convolutional Neural Networks(CNN) and Vision Transformer(ViT), as well as Channel Attention Mechanism(CAM) enhances the expression of key features, is constructed to accurately identify Cervidae categories. Experimental results indicate that this method achieves an Average Precision (AP) of 95.9% for Cervidae location and a top-1 accuracy of 77.73% for Cervidae identification. This research contributes to a more comprehensive and accurate monitoring of wild Cervidae, and provides valuable references for subsequent UAV-based wildlife monitoring.