UAD-YOLO: An Efficient Context-Aware and Shape-Aware Framework Based on YOLOv11 for Urban Anomaly Detection in Complex Scenes

Cities are living systems, constantly shifting and changing, yet they rely on a steady rhythm of maintenance to stay safe. When a road cracks, a tree falls, or a sign is damaged, the consequences can range from minor inconvenience to serious danger. For decades, keeping an eye on these urban hazards has been a job for human inspectors, who walk or drive through neighborhoods looking for trouble. This method is slow, expensive, and often misses small or hidden problems. In recent years, scientists have turned to computers to help, teaching them to "see" these issues through cameras. This field, known as object detection, allows machines to scan images and identify specific items, from cars to potholes. However, teaching a computer to spot a jagged crack in a busy street is far harder than finding a car in an empty parking lot. The background is cluttered, the lighting changes, and the objects themselves are often irregular, broken, or partially hidden.

A new study from researchers at Sichuan Vocational and Technical University of Communications addresses these difficulties by creating a smarter way for computers to watch our cities. The team developed a system called UAD-YOLO, designed specifically to find a wide variety of urban problems in real time. Instead of just looking for shapes that look like standard boxes, this system was built to understand the messy, complex reality of a city street. It can spot seven different types of anomalies, including road cracks, potholes, damaged signs, fallen trees, garbage, graffiti, and broken utility poles. The researchers did not just rely on existing data; they built a new, large collection of images containing thousands of examples of these specific problems to teach the computer what to look for. By combining several advanced techniques, they created a tool that is both highly accurate and fast enough to run on standard equipment, making it a practical solution for continuous city monitoring.

The core challenge the researchers faced was that urban anomalies do not look like neat, perfect objects. A road crack might stretch for meters in a thin, winding line, while a pile of garbage might be an amorphous blob. Traditional computer vision tools often struggle with these irregular shapes, especially when they are surrounded by other distractions like shadows, other vehicles, or complex textures. To solve this, the team modified a powerful existing detection framework, known as YOLOv11, which is famous for its speed. They added three specific improvements to help the computer "think" more like a human observer. First, they gave the system a better way to look at the big picture. By using a technique that allows the computer to see long-range connections in an image, it can understand how a small crack relates to the larger road surface, rather than just seeing a random line. This helps the system ignore background noise and focus on what actually matters.

Second, the researchers improved how the computer looks at fine details. They introduced a method that allows the system to learn complex textures during its training phase but then simplifies its structure when it is actually working. This is similar to how a student might study with many notes and diagrams but then take a test using only a streamlined mental map. This approach ensures the computer can spot small details, like a tiny pothole or a faint scratch, without slowing down the process. Third, they changed the way the computer draws the box around a detected object. Standard methods often force these boxes into rigid shapes, which can be inaccurate for irregular items. The new system uses a more flexible approach that respects the actual shape of the object, whether it is long and thin or short and wide, ensuring the location is marked precisely.

To test if these changes worked, the researchers trained their system on a new dataset they created, which included over 17,000 images of urban scenes. They tested the system against other popular detection methods and found that their new approach was significantly better at finding the right objects while missing fewer of them. In their tests, the system correctly identified urban anomalies in 83.1% of the cases where it was supposed to find them, a notable improvement over the standard models. It also managed to find 77.1% of all the actual problems present in the images, a key metric for safety applications where missing a hazard is dangerous. Perhaps most importantly for real-world use, the system remained incredibly fast. It could process an image and give an answer in just 2.31 milliseconds, meaning it could theoretically analyze hundreds of images every second. This speed suggests the technology could be installed on moving vehicles or drones to monitor city streets continuously without lag.

The study also explored how different settings affected the system's performance, confirming that the specific combination of techniques they chose was the most effective. They found that using a moderate size for the "long-range view" was crucial; looking too narrowly missed context, while looking too broadly introduced too much confusion. Similarly, they discovered that the most flexible shape-detection settings worked best for the irregular nature of city damage. While the system is not perfect and can still be confused by extreme lighting or heavy obstruction, the results show a clear path forward. The researchers acknowledge that more testing in different weather conditions and locations is needed, but the current findings demonstrate that a computer can now be trained to see the subtle, broken, and messy details of urban life with a level of accuracy and speed that was previously out of reach. This work offers a promising tool for cities to move from reactive repairs to proactive safety, ensuring that the infrastructure we rely on is watched over with a constant, unblinking eye.