An Efficient Algorithm for Small Livestock Object Detection in Unmanned Aerial Vehicle Imagery

TL;DR

This paper introduces a new algorithm for detecting small livestock in drone images, improving accuracy and efficiency for population surveys.

Contribution

The novel LSNET algorithm enhances small-object detection in UAV imagery with a lightweight model and improved accuracy.

Findings

The proposed LSNET algorithm increases mean Average Precision (mAP) by 1.47% compared to YOLOv7.

The method effectively detects small livestock in dense and challenging UAV imagery conditions.

A new dataset of grazing livestock was developed for deep learning using UAV images from Inner Mongolia.

Abstract

Precise livestock detection is crucial for livestock population surveys. However, livestock in unmanned aerial vehicle imagery are often small and densely distributed, leading to suboptimal detection performance. To address this issue, we propose a novel small-object livestock detection method. Experimental results demonstrate that proposed method significantly enhances the accuracy of livestock detection while achieving a lightweight model, providing an effective technical solution for livestock population surveys.

Livestock population surveys are crucial for grassland management tasks such as health and epidemic prevention, grazing prohibition, rest grazing, and forage–livestock balance assessment. These tasks are integral to the modernization and upgrading of the livestock industry and the sustainable development of grasslands. Unmanned aerial vehicles (UAVs) provide significant advantages in flexibility and maneuverability, making them ideal for livestock population surveys. However, grazing livestock in UAV images often appear small and densely packed, leading to identification errors. To address this challenge, we propose an efficient Livestock Network (LSNET) algorithm, a novel YOLOv7-based network. Our approach incorporates a low-level prediction head (P2) to detect small objects from shallow feature maps, while removing a deep-level prediction head (P5) to mitigate the effects of excessive down-sampling. To capture high-level semantic features, we introduce the Large Kernel Attentions Spatial Pyramid Pooling (LKASPP) module. In addition, we replaced the original CIoU with the WIoU v3 loss function. Furthermore, we developed a dataset of grazing livestock for deep learning using UAV images from the Prairie Chenbarhu Banner in Hulunbuir, Inner Mongolia. Our results demonstrate that the proposed module significantly improves the detection accuracy for small livestock objects, with the mean Average Precision (mAP) increasing by 1.47% compared to YOLOv7. Thus, this work offers a novel and practical solution for livestock detection in expansive farms. It overcomes the limitations of existing methods and contributes to more effective livestock management and advancements in agricultural technology.

Figures