The authors of the Cattle Detection and Counting in UAV Images dataset propose a system for cattle detection and counting to assist with grazing cattle management. Their approach relies on Convolutional Neural Networks (CNNs) and utilizes aerial images captured by an Unmanned Aerial Vehicle (UAV). To enhance the accuracy of cattle detection, they leverage the predictability of object size in UAV images, assuming a relatively constant UAV height from the ground. Their method involves resizing images to an optimal resolution based on object size and network down-sampling rates, both during training and testing.

Cattle counting is a crucial task in grazing cattle management, aiding farmers in efficiently organizing pasturing layouts and promptly identifying cattle-related incidents. However, manually counting cattle by walking across the pasture is both labor-intensive and time-consuming. Therefore, the authors advocate for computer vision-based cattle monitoring as a cost-effective and labor-saving alternative. They highlight the potential of using aerial images captured by UAVs for monitoring animals over extensive and hard-to-reach areas. UAVs offer advantages such as wide coverage and reduced object overlap from a bird’s-eye perspective, making them suitable for object detection and counting.

The authors describe two fundamental approaches for estimating object counts from images: count-by-regression and count-by-detection. Given the need to cover large grazing areas and ensure cattle visibility in each image, they adopt the count-by-detection approach. This involves first detecting cattle in individual images and then aggregating the results to estimate the total count.

In their study, the authors employ a DJI Phantom 4 UAV, weighing 1,380 g and capable of approximately 28 minutes of flight time. They use the UAV’s still photography mode to capture images with a maximum size of 4,000x3,000 pixels. The UAV supports live viewing during photography through the DJI GO 4 mobile application.

The dataset used in their research comprises two sets of pasture aerial images. Dataset1 consists of 656 images, captured on 18 May 2016, in Aso, Kumamoto, Japan, under sunny and clear weather conditions. The UAV was flown at an altitude of about 50 m during seven flights, labeled A to G. Each image covers an area of approximately 80 m x 60 m. Images were taken at 2-second intervals but sometimes had intervals of up to 12 seconds due to UAV turning. The images have an 80% overlap.

Dataset2 includes 14 images taken on 24 August 2016 during a single flight, also in sunny and clear conditions. This dataset covers a different area within the same pasture in Kumamoto, Japan. Images were captured every 6 seconds. The purpose of this dataset is to evaluate the robustness of the authors’ cattle detection and counting system.

Figure 1. Examples of flight route and images in our two datasets. (a) shows one example of flight route. (b) shows examples of dataset 1. © shows examples of dataset 2.

For annotation purposes, the authors manually labeled all cattle in the images, providing details such as the top-left position, width, and height of the bounding boxes, data quality labels, and cattle IDs. The data quality label include Normal, Truncated, Blurred, and Occluded categories and examples of each label are provided. Multiple images of the same target were labeled with the same ID, allowing for individual cattle counting in each flight area. Flights F and G in Dataset1, containing more images, were divided into flight sections in the experiment. The areas covered by each flight section range from 1.5 to 5 hectares.

Dataset1 comprises 1,886 annotations of 212 individual cattle targets, while Dataset2 contains 62 annotations of 6 individual targets. The median cattle size is 87x90 pixels in Dataset1 and 59x101 pixels in Dataset2. The histogram of the longer edge length of cattle bounding boxes with the data quality label Normal in Dataset1 is presented. Among the 218 individual targets, 210 appeared in multiple images. The authors also assess the moving distance of cattle during each flight section, using body length as a reference.