Eyes on the Road

The rise in traffic on roadways is a major issue that necessitates effective transportation planning, design, and management. As urban populations continue to grow, the number of vehicles on the roads grows in tandem. Accordingly, congestion becomes a pressing issue that affects both the economy and quality of life. Congested roads lead to longer commute times, wasted fuel, increased pollution levels, and decreased productivity, ultimately putting a strain on the overall transportation infrastructure.

To address this pressing concern, city planners must analyze current traffic patterns, anticipate future growth, and identify areas susceptible to congestion. By forecasting transportation demands accurately, they can develop strategies to accommodate the increasing volume of vehicles and travelers. This involves constructing new roadways, expanding existing ones, and creating alternative transportation options like public transit systems and bicycle lanes.

But to support these activities, raw traffic data is needed. And that data can be hard to come by. It is generally collected by either pneumatic tube technologies, stationary cameras, or human counters. Unfortunately, these methods provide data that is sparse, both spatially and temporally, which makes it difficult for urban planners to utilize effectively.

Camera-based traffic monitoring systems are especially appealing because they can continually and accurately monitor roadways with the help of machine learning object detection algorithms. These algorithms have advanced to the point that they can reliably detect objects of interest, while only requiring very minimal hardware resources. And of course this type of system is fully automated and capable of providing real-time data with no humans in the loop.

But one major roadblock remains that prevents the widespread adoption of such technologies — a lack of sensors. Installing and maintaining a large number of cameras to monitor all of the streets in a big city is a huge, and very expensive, undertaking. Perhaps new cameras do not need to be installed, however, says a team of engineers at The Ohio State University. Recognizing that a modern city is already filled with cameras, whether for safety and security or other purposes, they came up with a plan to piggyback on top of those cameras to extract traffic information.

The team leveraged the cameras that were already installed on campus buses around The Ohio State University. Using this large network of existing cameras, they were able to analyze the video streams using the YOLOv4 object detection neural network to count vehicles. They also demonstrated that they could locate other objects in the road, and even tell the difference between a parked car and one in motion. And by borrowing an existing network of sensors, there was no cost involved in getting this project off the ground.

While traffic planning was the focus of this work, additional use cases for borrowed sensor data were discovered. It was also found, for example, that the precise location of a bus could be determined by analyzing streams of images. This information was sufficient to detect if a bus had departed from its planned route.

One of the researchers involved in this study noted that “if we collect and process more comprehensive high-resolution spatial information about what’s happening on the roads, then planners could better understand changes in demand, effectively improving efficiency in the broader transportation system.”

This initial work proved that there is merit in the idea, but there is still quite a bit of work to be done before a real-world deployment can occur. In the future, the team hopes to explore how their approach performs under more varied conditions than were seen on the campus bus system. They also want to evaluate the use of different edge computing platforms to run the object detection algorithms onboard buses in real time.