无人机车辆重识别数据集

UAV-VeID是基于无人机视频构建的车辆重识别数据集。其包含41,917张有标注的图像和4,601辆各型车辆。数据拍摄场景包括复杂的背景，以及不同的视角、光照和遮挡。数据集还添加了300,000张干扰图像作为干扰数据集，以增加数据的多样性。在视频采集过程中，无人机视频采集的地点、背景和光照条件不同，以获得车辆更多的观角、尺度和背景的变化，并尽可能的模拟真实场景。无人机的飞行高度在15至60米之间，无人机摄像头的垂直角控制在40至80度之间。视频以30帧每秒（fps）的速度和2704×1520、4096×2160两种分辨率录制。项目组在原始视频中挑选了80个视频序列用以构建UAV-VeID数据集。项目组从收集的无人机视频中对车辆ID进行了人工标注。首先，对视频进行每秒一帧采样作为关键帧。在关键帧中进一步标注车辆检测框和ID信息。为完成车辆标注工作，项目组邀请了6位数据标注人员。整个数据标注过程耗时大约1000个工时，最终得到了包含41917个车辆边界框和4601个车辆ID的图片数据集。每个车辆ID至少有两个边界框标注，大多数车辆ID都有3个以上不同视角下的图像。UAV-VeID随机分为三个部分，用以分别进行训练、验证和测试。其中训练集包含1797个车辆ID的18709个图像，验证集包含596个车辆ID的4150个图像，测试集包含2208个车辆ID的19058个图像。对于UAV-VeID中的验证集和测试集，项目组随机选择每个ID的一张图像放入待查询图像库中，其他图像作为查询图像。

UAV-VeID is a vehicle re-identification dataset constructed based on unmanned aerial vehicle (UAV) videos. It contains 41,917 annotated images and 4,601 vehicles of various types. The data capture scenarios include complex backgrounds, as well as diverse viewpoints, lighting conditions and occlusions. Additionally, 300,000 distractor images are added to the dataset to increase data diversity.

During video collection, the UAV video acquisition locations, backgrounds and lighting conditions were varied to obtain more variations in vehicle viewpoints, scales and backgrounds, and to simulate real-world scenarios as much as possible. The flight altitude of the UAV ranges from 15 to 60 meters, and the vertical angle of the UAV camera is controlled between 40 and 80 degrees. The videos are recorded at 30 frames per second (fps) with two resolutions: 2704×1520 and 4096×2160.

The project team selected 80 video sequences from the original videos to construct the UAV-VeID dataset. Vehicle IDs were manually annotated from the collected UAV videos. First, one frame per second was sampled from the videos as key frames. Then, vehicle bounding boxes and ID information were annotated in the key frames. To complete the vehicle annotation work, 6 data annotators were invited. The entire data annotation process took approximately 1,000 man-hours, and finally an image dataset containing 41,917 vehicle bounding boxes and 4,601 vehicle IDs was obtained.

Each vehicle ID has at least two bounding box annotations, and most vehicle IDs have images captured from more than three different viewpoints.

UAV-VeID is randomly divided into three subsets for training, validation and testing respectively. Specifically, the training set contains 18,709 images of 1,797 vehicle IDs, the validation set contains 4,150 images of 596 vehicle IDs, and the test set contains 19,058 images of 2,208 vehicle IDs. For the validation and test subsets of UAV-VeID, the project team randomly selected one image per ID for the gallery set, while the remaining images were used as query images.

搜集汇总

背景与挑战

背景概述

无人机车辆重识别数据集（UAV-VeID）是一个基于无人机视频构建的大规模车辆重识别数据集，包含41,917张标注图像和4,601辆车辆，覆盖复杂背景、多视角、光照变化和遮挡条件，并添加了300,000张干扰图像以增强数据多样性。数据采集参数包括飞行高度15-60米、高分辨率视频（最高4096×2160），从80个视频序列中人工标注生成，耗时约1000工时，每个车辆ID至少有2个边界框标注，多数有3个以上不同视角图像。数据集已划分为训练、验证和测试集，适用于车辆重识别算法的开发和评估，模拟真实场景以提升模型鲁棒性。