Counting cattle Dataset: Supporting livestock Multi Objetct Tracking

Abstract Livestock farming has made technological advancements in recent times; however, there is still much to improve, as AI and computer vision have not yet fully penetrated the area. Livestock farming involves many routine tasks that require automation for improved control over production processes. Although the scientific community has made significant efforts to support precision livestock farming in recent years, artificial intelligence-based solutions for livestock counting and management remain limited (Myat Noe et al., 2023). Myat Noe et al. (2023) proposed a method for counting black cattle counting using the YOLO (You Only Look Once) object detection framework, in combination with the Detectron2 segmentation model and the DeepSORT (Wojke, Bewley and Paulus, 2017) and StrongSORT (Du et al., 2023) Multi Object Tracking (MOT) solutions. The study highlights the challenges faced in supporting cattle tracking as MOT solutions generally target objects with distinctly visual features. In the context of cattle detection and tracking, the requirements are more demanding due to the high visual similarity in color and shape among individuals. As a result, additional effort is still required to support precision agriculture through cattle identification and monitoring. This includes the creation of custom datasets for livestock MOT since existing public repositories do not meet the specific needs of this application. Moreover, retraining the models and applying additional techniques in combination are still necessary to achieve more satisfactory results (Myat Noe et al., 2023). In this context, the present dataset provides a series of videos and ground truth information for tracking Brangus cattle in a ranch environment in Uruguay. The dataset maintains consistent bounding boxes and identifiers for individual animals, being specifically useful for evaluating the effectiveness of tracking algorithms in support of cattle monitoring through MOT metrics. DataRecords The dataset contains 10 preselected videos, totaling a recording time of 3.17 minutes and a total of 93 cattle. It is organized into folders, one containing the videos used for tracking and another folder containing the ground truth, which consists of a CSV file with the processed information and the video on which the analysis was performed. The detection values were stored in the mentioned CSV file, so they can be processed later. Each line represents a detection, with the following information per line: -The frame number corresponding to the detection -The identifier number corresponding to the detected object -The x-coordinate value of the top-left corner of the detected object -The y-coordinate value of the top-left corner of the detected object -The width of the bounding box -The height of the bounding box -The confidence of the detection Further information on dataset construction and organization available on the dataset description attached file.摘要 近年来，畜牧养殖业已取得技术进步，但由于人工智能（AI）与计算机视觉尚未完全渗透该领域，仍有诸多环节有待完善。畜牧养殖包含诸多常规作业，需通过自动化手段以强化对生产流程的管控。尽管近年来科学界为支撑精准畜牧养殖已付出诸多努力，但基于人工智能的牲畜计数与管理解决方案仍较为有限（Myat Noe等人，2023）。 Myat Noe等人（2023）提出了一种基于YOLO（You Only Look Once，目标检测框架）的黑牛计数方法，结合了Detectron2分割模型以及DeepSORT（Wojke、Bewley与Paulus，2017）与StrongSORT（Du等人，2023）多目标跟踪（Multi Object Tracking, MOT）方案。该研究指出了肉牛跟踪任务所面临的挑战：多目标跟踪方案通常针对具备显著视觉特征的目标对象。在肉牛检测与跟踪场景中，由于个体间色彩与外形高度相似，对算法的要求更为严苛。因此，若要通过牲畜识别与监测支撑精准农业，仍需付出额外的研发努力。这包括为畜牧养殖多目标跟踪任务构建定制化数据集，因为现有的公开数据集仓库无法满足该应用场景的特定需求。此外，为获得更理想的效果，仍需对模型进行重新训练并结合多种额外技术手段（Myat Noe等人，2023）。 基于此背景，本数据集提供了一系列在乌拉圭牧场环境中跟踪婆罗门牛（Brangus cattle）的视频数据与真值标注信息。该数据集为每头牲畜提供了一致的边界框与唯一标识符，尤其适用于通过多目标跟踪（MOT）评估指标来验证跟踪算法在肉牛监测场景中的有效性。 数据记录 本数据集包含10条预筛选视频，总录制时长为3.17分钟，共涵盖93头肉牛。数据集以文件夹结构组织：一个文件夹存放用于跟踪任务的原始视频，另一个文件夹存放真值标注数据，其中包含一份记录了处理后信息的CSV文件以及本次分析所对应的视频文件。 检测相关的数值已存储于上述CSV文件中，可供后续处理使用。文件中每一行对应一次检测结果，每行包含以下信息： - 对应检测的帧编号 - 被检测对象的唯一标识符编号 - 被检测对象边界框左上角的X轴坐标值 - 被检测对象边界框左上角的Y轴坐标值 - 边界框的宽度 - 边界框的高度 - 本次检测的置信度得分 关于数据集构建与组织方式的更多细节，请参阅附件中的数据集说明文档。