Deer survey from drone thermal imagery using enhanced faster R-CNN based on ResNets and FPN

Deer surveys play an important role in the estimation of local ecological balance. In the Chitwan National Park of Nepal, the dense tree canopies and tall vegetation often obscure the presence of wild deer, which has a negative effect on the accurate population surveys of wild deer. DJI drones equipped with infrared sensors have been widely used to regularly monitor wild deer conservation areas and capture a lot of thermal images. How to automatically review and count the number of deer objects from thermal images is becoming more and more important. Due to the difference between thermal images and RGB images, as well as the distinct variations in object sizes in these two types of imagery, current ready-to-use object detection models, trained on true-color imagery, are ill-suited for the task of detecting small deer objects within thermal imagery. In this paper, an enhanced Faster R-CNN was constructed to detect small deer objects from thermal images, in which a Feature Pyramid Network (FPN) based on a residual network is used to improve the feature information of small deer objects and construct multi-scale feature maps for the localization of deer objects, bounding box regression, and regions of interest (RoIs) classification. In addition, small-scaled anchor boxes and a multi-scale feature map selection criterion are designed to improve the detection ability of small objects. Finally, based on Faster R-CNN, FPN, and different residual networks including ResNet18, ResNet34, ResNet50, ResNet101, and ResNet152, we constructed five object detection models, and respectively evaluated their detection performance by using COCO evaluation matrix. Under the condition of IoU≥0.5, the integration of Faster R-CNN, FPN, and ResNet18 is proved to be better than others. Specifically, The COCO evaluation results revealed an Average Precision (AP) score of 91.6% for all deer objects. Small deer objects (area ≤ 200 pixels) achieved an AP score of 73.6%, medium deer objects (200 < area ≤ 400) demonstrated an AP score of 93.4%, and large deer objects (area > 400) achieved the highest AP score of 94.3%. Our research is helpful for effective wild deer monitoring and conservation and can be a valuable reference for the exploration of small object detection from low-resolution thermal images.