Real-time black ice detection using YOLOX on drone

Detailed Info: https://github.com/hsh060824/blackice-drone-dataset Dataset Type: Object Detection Dataset (with bounding boxes) Overview Road safety during winter months remains a critical concern due to the elusive nature of black ice, a thin layer of ice that forms on road surfaces, making it challenging for drivers to identify and navigate safely. In an effort to address this issue, our research team at Cheongshim International Academy (CSIA) has conducted extensive studies on real-time black ice detection utilizing YOLOX, a state-of-the-art object detection algorithm, deployed on drones. As a significant contribution to the research community, we are pleased to share our meticulously curated image dataset, which encapsulates diverse scenarios and conditions representative of real-world black ice occurrences. Background Black ice poses a significant threat to road safety, especially during winter, as it is often challenging for drivers to detect, leading to increased risks of accidents and hazardous road conditions. Our dataset aims to fill the gap in existing resources by providing a comprehensive collection of images showcasing various instances of black ice under different environmental conditions. The dataset covers diverse scenarios, including different lighting conditions, road surfaces, and black ice formations, making it a valuable resource for developing and testing robust black ice detection models. Significances of the Dataset The significance of this dataset lies in its potential to advance the development of effective black ice detection algorithms. By sharing our dataset with the research community, we aim to facilitate the creation of more accurate and reliable models for real-time detection of black ice using drone technology. The dataset includes annotations in COCO format, providing detailed information about the location and characteristics of black ice instances in each image. Categorization In our pursuit of advancing the field of computer vision and contributing to ongoing research endeavors, we proudly introduce three distinct image datasets meticulously curated by our research team. These datasets, categorized as "White," "Black," and "Outdoors (OD)," cater to unique scenarios and are designed to fuel the development of specialized models addressing specific challenges in visual recognition. White Dataset: Composition: This dataset comprises 413 images, each meticulously annotated with an average of 1.1 annotations per image, depicting the unique optical characteristics of black ice. Properties: The average proportion of instance pixel area is 3.16%, emphasizing the subtlety of the black ice formations. The average image brightness is measured at 149.358. Capture Environment: The images were taken in controlled indoor laboratory conditions, ensuring consistency and repeatability. Creation Method: The dataset was generated by cooling asphalt samples in a freezer to temperatures ranging from -4°C to -20°C. Subsequently, 4°C water was sprayed onto the sample surfaces, creating black ice. The dataset captures the optical properties of black ice, showcasing its interaction with light. Significance: Valuable for highlighting the optical characteristics of black ice, enhancing model accuracy in well-lit scenarios. Black Dataset: Composition: This dataset comprises 814 images, with a detailed annotation structure averaging 3.5 annotations per image, showcasing the challenges of recognition in low-light conditions. Properties: The average proportion of instance pixel area is notably higher at 12.37%, reflecting the complex and varied formations of black ice. The average image brightness is measured at 123.028. Capture Environment: Similar to the White Dataset, images were captured in a controlled indoor laboratory environment. Asphalt pelt was placed under the black iced asphalt pieces to replicate realistic scenarios. Creation Method: The dataset creation involved the same process of cooling asphalt samples, followed by spraying water to create black ice. To simulate real-world conditions, asphalt pelt was used as a background, and various shapes of black ice were randomly placed in each image. Significance: Realistic emulation of black ice using backgrounds made up of asphalt pelts, providing essential drark images for robust model training. Outdoor (OD) Dataset Composition: This dataset is the most extensive, consisting of 1624 images, with an average of 1.5 annotations per image, capturing the challenges of recognizing black ice in outdoor winter conditions. Properties: The average proportion of instance pixel area is 12.34%, mirroring the complexity of real-world outdoor scenarios. The average image brightness is significantly lower at 56.575. Capture Environment: Unlike the indoor datasets, the OD dataset was captured outdoors in winter conditions where black ice naturally forms. Creation Method: Black ice was created on the asphalt road of Cheongsim International High School by spraying +4°C water onto the surface. DJI Tello's built-in camera was used for capturing images from various angles, simulating drone-like perspectives. This dataset is designed to closely replicate real-world scenarios, providing a valuable resource for training models for outdoor applications. Significance: Represents real-world outdoor scenarios, offering a unique perspective for developing models capable of handling diverse and challenging conditions. Cameras: iPhone SE2 (Apple, California) iPhone SE3 (Apple, California) iPhone 12 (Apple, California) iPhone 14 Pro (Apple, California) Q9 (LG Electronics, Seoul, Korea) V30 (LG Electronics, Seoul, Korea) Tello (DJI, ShenZhen, China)

详细信息：https://github.com/hsh060824/blackice-drone-dataset
数据集类型：目标检测数据集（含边界框标注）

概述：冬季道路安全始终是关键民生议题，因黑冰（black ice）——一种形成于路面的薄冰层——极具隐蔽性，驾驶员难以识别并安全通行。为解决这一难题，清心国际学院（Cheongshim International Academy, CSIA）的研究团队开展了大量研究，基于前沿目标检测算法YOLOX，在无人机平台上实现黑冰的实时检测。作为对学界的重要贡献，我们荣幸地分享精心遴选的图像数据集，该数据集涵盖了真实世界中黑冰出现的多样场景与环境条件。

背景：黑冰对道路安全构成重大威胁，尤其在冬季，因其难以被驾驶员察觉，极易引发交通事故与危险路况。现有相关研究资源存在缺口，本数据集旨在填补这一空白，提供全面的图像集合，展示不同环境条件下的各类黑冰实例。数据集覆盖多样化场景，包括不同光照条件、路面类型与黑冰形态，是开发与测试鲁棒性黑冰检测模型的宝贵资源。

数据集的意义：本数据集的核心价值在于有望推动高效黑冰检测算法的发展。通过向学界共享本数据集，我们期望助力基于无人机技术的黑冰实时检测模型实现更高的精度与可靠性。数据集采用COCO标注格式进行标注，提供每张图像中黑冰实例的位置与特征细节信息。

数据集分类：为推进计算机视觉领域发展并助力相关研究，我们的研究团队精心打造了三个独立的图像数据集，分别归类为“White”、“Black”与“Outdoors (OD)”，针对特定场景设计，旨在推动针对视觉识别特定挑战的专用模型开发。

### White数据集
组成：该数据集包含413张图像，每张图像平均标注1.1个边界框，用于呈现黑冰独特的光学特性。
属性：单实例像素面积占比均值为3.16%，凸显了黑冰形态的隐蔽性；图像平均亮度为149.358。
拍摄环境：所有图像均在可控的室内实验室环境中拍摄，确保实验的一致性与可重复性。
制作方法：通过将沥青样本置于冰柜中冷却至-4℃至-20℃的温度范围，随后在样本表面喷洒4℃的水以形成黑冰。该数据集记录了黑冰与光线交互的光学特性。
意义：有助于凸显黑冰的光学特性，提升模型在光照良好场景下的检测精度。

### Black数据集
组成：该数据集包含814张图像，每张图像平均标注3.5个边界框，用于展示低光照条件下的识别挑战。
属性：单实例像素面积占比均值显著更高，达12.37%，反映了黑冰形态的复杂性与多样性；图像平均亮度为123.028。
拍摄环境：与White数据集一致，所有图像均在可控的室内实验室环境中拍摄，以沥青块作为背景以还原真实场景。
制作方法：采用与White数据集相同的流程制备黑冰，通过在沥青背景上随机放置不同形态的黑冰，模拟真实世界的场景。
意义：通过沥青背景还原真实黑冰场景，为模型训练提供关键的低光照图像数据。

### 户外（Outdoors, OD）数据集
组成：该数据集规模最大，包含1624张图像，每张图像平均标注1.5个边界框，用于捕捉户外冬季环境中黑冰识别的挑战。
属性：单实例像素面积占比均值为12.34%，贴合真实户外场景的复杂性；图像平均亮度显著更低，为56.575。
拍摄环境：与前两个室内数据集不同，OD数据集拍摄于冬季户外环境，黑冰可自然形成于路面。
制作方法：在清心国际高中的沥青路面上喷洒4℃的水以形成黑冰，使用DJI Tello内置相机从不同角度拍摄，模拟无人机拍摄视角。该数据集旨在高度还原真实户外场景，为户外应用的模型训练提供宝贵资源。
意义：该数据集代表真实的户外场景，为开发能够应对多样复杂环境的检测模型提供了独特的研究视角。

拍摄设备：
- iPhone SE2（苹果公司，加利福尼亚州，美国）
- iPhone SE3（苹果公司，加利福尼亚州，美国）
- iPhone 12（苹果公司，加利福尼亚州，美国）
- iPhone 14 Pro（苹果公司，加利福尼亚州，美国）
- Q9（LG电子，首尔，韩国）
- V30（LG电子，首尔，韩国）
- Tello（大疆创新，深圳，中国）