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)