Software Prototype for Smart Traffic Monitoring Using YOLOv11: Baseline Evaluation and Limitations

In the bustling arteries of modern cities, the flow of vehicles is a constant, chaotic rhythm that often grinds to a halt. For decades, managing this flow has relied on rigid systems: traffic lights that change on fixed timers regardless of how many cars are waiting, and human officers watching from towers, unable to see every intersection at once. As cities grow denser, these old methods struggle to keep pace, leading to longer commutes, wasted fuel, and increased danger on the roads. To solve this, scientists are turning to a branch of artificial intelligence known as computer vision. This technology allows machines to "see" the world through cameras, identifying objects like cars, buses, and motorcycles in real-time. By combining this visual intelligence with the internet of things—where devices talk to each other wirelessly—researchers hope to create traffic systems that react instantly to changing conditions, much like a nervous system responding to a sudden touch.

A researcher at Sanjivani University in India recently built a software prototype to test how well this vision works for traffic management. Their goal was to create a system that could watch a video feed of a street, count the vehicles, estimate how crowded the road is, and display that information on a simple web dashboard. They used a powerful, pre-trained artificial intelligence model called YOLOv11, which is designed to spot objects in images, and connected it to a web interface built with standard programming tools. The design also included a plan for a small, low-cost computer chip to eventually control traffic lights automatically, though that physical hardware was not part of this specific test. The researcher wanted to see if this combination of software and vision could serve as a reliable foundation for a smarter, more responsive traffic network.

The system they built works in a straightforward sequence. First, a user uploads a video or a photo of a traffic scene to a web dashboard. The software then breaks the video down into individual frames, or snapshots, and prepares them for analysis. The artificial intelligence model examines each snapshot, drawing boxes around anything it recognizes as a vehicle and assigning a label to it. Based on how many vehicles it finds, the system categorizes the traffic as low, medium, or high density. This information is then sent back to the dashboard, where a human operator can see the current count and the status of the road. The researcher also designed the system to eventually detect accidents or traffic violations, such as running a red light, but they were careful to note that these advanced features were not yet fully tested or validated in this initial version.

When the researcher put their prototype to the test, the results offered a clear, if sobering, picture of where the technology stands today. They ran the system on a standard computer processor, processing nine hundred images that contained one thousand one hundred and fifty distinct object instances. The performance metrics revealed that the current setup was not yet ready for real-world deployment. The system correctly identified vehicles only about two percent of the time when it made a guess, and it detected roughly twenty-one percent of the vehicles that were actually present in the images. A closer look at the data showed why the system struggled: the artificial intelligence model they used had been trained on a massive, general dataset that includes many different types of objects, such as people, bicycles, and even airplanes. Because the researcher did not retrain the model specifically on a dataset of only traffic vehicles, the computer got confused. It spent a significant amount of time trying to identify airplanes and people in the traffic footage, while often failing to spot the cars and buses it was supposed to be counting. The researcher noted that the model was not "fine-tuned" for this specific job, meaning it was trying to apply a broad, general knowledge to a very specific problem without the necessary adjustments.

Despite the low accuracy numbers, the project successfully demonstrated that the basic architecture of a smart traffic system is possible. The software pipeline worked as intended: it accepted media, processed it through the vision model, counted the detections, and displayed the results on a web page. The dashboard showed a vehicle count, a density rating, and placeholders for accident alerts and signal status. However, the researcher was explicit that the accident detection and traffic violation features shown on the screen were not functional in this test; they were merely visual elements to show where those capabilities would go in a future, more complete version. The system did not actually detect accidents or violations, nor did it control any physical traffic lights, as the hardware component remained a theoretical part of the design.

The researcher concluded that while the software framework is sound, the "brain" of the system needs significant work before it can be trusted on real roads. The low performance scores indicate that the model must be retrained on a dataset that contains only the specific types of vehicles found on the road, and the confusing categories like airplanes must be removed. They also emphasized that the current results are a baseline—a starting point for diagnosis—rather than a finished product. The prototype proved that a web-based dashboard can visualize traffic data, but it also highlighted that relying on a generic, pre-made model without customization leads to poor results.

Looking ahead, the researcher outlined a clear path for improvement. They plan to retrain the artificial intelligence model with a specialized dataset of cars, buses, and trucks to improve its ability to recognize traffic. They intend to integrate the small computer chip they designed to actually control traffic lights based on the density readings. Future versions will also aim to add real-time video streaming from live cameras, rather than just uploaded files, and will work on training the system to spot accidents and rule violations with high reliability. For now, this prototype serves as a valuable diagnostic tool, showing exactly where the technology falls short and what steps are necessary to build a truly intelligent traffic management system that can keep cities moving safely and efficiently.