Hjort-YOLO: a lightweight and explainable computer vision framework for wildlife detection
Abstract
Introduction Automated animal detection is an important component of ecological monitoring and may in the future be used also for wildlife vehicle collision mitigation by constructing efficient alert systems. However, the often complex environmental background in habitats along roads and different sizes due to the distance from the camera lens of animals in the images may reduce the detection accuracy, whereas high-precision detection requires computationally expensive object detection methods. This paper presents Hjort-YOLO, an efficient computer vision framework for detecting wildlife from camera-trap images acquired under challenging background conditions. Methodology The proposed framework was developed by enhancing the lightweight YOLOv11n architecture through targeted modifications to the backbone and detection head by improving the cross-stage partial convolutional block (C2f-W), an area attention block (A2C2f-W), and reducing a detection head compared to the YOLOv11n algorithm, enabling an optimal balance between detection accuracy and computational efficiency. Hjort-YOLO is evaluated on the custom Svanøy deer dataset and the African Wildlife dataset, which include complex forest scenes, occlusions, background clutter, and motion blur. Results Experimental findings indicate that the Hjort-YOLO achieves detection accuracy comparable to state-of-the-art object detection models while significantly improving runtime performance. Hjort-YOLO provides 0.2 percent, 13 percent, and 17 percent better mean average precision, frames per second, and inference time, respectively, while detecting wildlife compared to YOLOv11n. The computation complexity, such as Giga Floating Operations per Second, was reduced by 19 percent compared to the YOLOv11n algorithm. Discussion These performance gains show the suitability of the proposed framework for detecting wildlife from ground-based sensors like camera traps, providing real-time performance and high detection accuracy in resource-constrained platforms.