| Abstrakt: |
With traditional survey methods such as ground-based counting, camera trapping, and aerial surveys, monitoring wild deer in Nepal's Chitwan National Park is challenging due to the dense tall vegetation that often conceals them. However, the thermal signatures of wild deer contrast sharply against the cooler background, facilitating detection via thermal imaging. This study explores the use of Unmanned Aerial Vehicles (UAVs) equipped with thermal cameras to monitor wild deer. A large volume of images can be captured, where wild animals appear as small objects. Reviewing these images manually is labor-intensive and time-consuming. To address this, we developed an object detection model using modified Faster R-CNN that automatically identifies small deer objects in the thermal images. Instead of VGG 16, the Feature Pyramid Network and Residual Neural Network (ResNet152) were employed to enhance feature extraction from these images, constructing multi-scale feature maps that enrich the feature information for small object detection. Customized anchor boxes were also designed to handle the wide variation in object scale and aspect ratios. To improve species identification accuracy for small Regions of Interest, a multi-scale aggregation method was proposed, which fuses features from multiple feature maps via Multi-scale RoIAlign pooling. The model proposed in this paper was evaluated by the COCO metrics. The experimental results obtained for the detection of deer and other animals in UAV thermal images with the resolution of 640 × 512 , showing mean Average Precision of 92.3% for all objects, 78.9% for small objects, 94.6% for medium objects, and 95.8% for large objects. This research provides a valuable means for detecting small objects in thermal images and contributes significantly to the field of wildlife monitoring. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink