无人机人物热成像标注数据

本算法用于无人机热成像监测系统中的人车非（人、汽车、自行车/非机动车）目标检测，数据可用于无人机热成像监测、夜间搜救、城市安全管理等相关的目标检测模型训练。通过一定数量的专利进行验证，可以评估专利的可行性，也可以评估本数据集的多样性和实用性数据采集自我司结合若干专利进行技术中介、专利运营时的测试数据脱敏后的目标检测数据。利用无人机热成像摄像头在多样化场景中收集图像，预处理，保留原始热成像分辨率和热成像的特征信息。考虑到小目标检测的特殊性，采用多尺度训练策略。随机裁剪-随机旋转-随机翻转-热噪声模拟对图像进行增强。热噪声模拟通过添加高斯噪声来模拟热成像设备的噪声特性。使用LabelImg工具，用包围框(x,y,width,height)定位目标。用COCO格式进行多类目标检测标注，"Person"中，只要能辨认出人形即可标注，即使部分遮挡也标，标注时只包括露出部分；"Car"包括各种机动车，从轿车到卡车都算入，只要能看到车身一部分即可标注露出部分；"Bike"涵盖自行车和电动自行车，整车出现两个轮子才标注。
将COCO格式转换为CSV格式：
x1=x,y1=y,x2=x+width,y2=y+height,cls∈{"Person","Car","Bike"}分别对应行人、机动车辆、自行车及电动自行车等非机动车辆。cls为类别，x1,y1,x2,y2是坐标。

This algorithm is designed for pedestrian, motor vehicle and non-motor vehicle (person, car, bicycle/non-motor vehicle) object detection in unmanned aerial vehicle (UAV) thermal imaging monitoring systems. The dataset can be used for training object detection models related to UAV thermal imaging monitoring, night search and rescue, urban safety management and other scenarios.

The dataset can be validated using a certain number of patents, which can not only evaluate the feasibility of the patents, but also assess the diversity and practicality of this dataset. The data is de-identified object detection data collected from test data generated during our company's technical intermediary and patent operation services combined with several patents.

Images were collected using UAV thermal imaging cameras in diverse scenarios, then preprocessed while retaining the original thermal imaging resolution and characteristic thermal information. Considering the particularity of small object detection, a multi-scale training strategy is adopted. Image augmentation is implemented via random cropping, random rotation, random flipping and thermal noise simulation. The thermal noise simulation mimics the noise characteristics of thermal imaging equipment by adding Gaussian noise.

Object localization was conducted using the LabelImg tool with bounding boxes formatted as (x, y, width, height). Multi-class object detection annotations were produced in COCO format: For the "Person" category, annotation is allowed as long as the human silhouette can be recognized, even if partially occluded, and only the exposed part shall be included in the annotation; For the "Car" category, all types of motor vehicles from sedans to trucks are covered, and annotation shall be made for the exposed part as long as a portion of the vehicle body is visible; For the "Bike" category, it includes bicycles and electric bicycles, and annotation shall be performed only when the entire vehicle with two wheels is present.

Convert the COCO format annotations to CSV format: x1 = x, y1 = y, x2 = x + width, y2 = y + height, where cls ∈ {"Person", "Car", "Bike"} correspond to pedestrians, motor vehicles, bicycles and electric bicycles and other non-motor vehicles respectively. Here, cls represents the object category, and x1, y1, x2, y2 are the bounding box coordinate values.

搜集汇总

数据集介绍

特点

无人机人物热成像标注数据包含2883条数据，用于无人机热成像监测系统中的人车非目标检测，适用于夜间搜救和城市安全管理等场景。数据采用多尺度训练策略和多种图像增强技术，标注格式为COCO转换为CSV格式，包含行人、机动车辆和非机动车辆等类别。