Automating traffic incident detection with deep learning

Guest authors

Jean Marie Guyon, CitilogThis post is written by Jean Marie Guyon, Sales and Marketing Manager at Citilog. Read more about Jean Marie at the end of this post.

Where traffic exists, incidents will happen, and hold-ups will occur. For those tasked with keeping highways or urban roads clear and traffic flowing, the ability to minimize disruption caused by incidents is critical in meeting this objective.

Traffic Management Centers (TMCs) have the role to monitor and manage traffic, inform traffic users and control traffic flows in real-time, every day, all year. Video surveillance is a key tool in monitoring road networks, intersections and critical infrastructure such as tunnels and bridges. It provides a real-time view of traffic flow and incidents – including crashes, queues and slow-moving traffic – that might disrupt free-flowing movement on the road. However, given human limitations in attention, concentration and, in the most obvious physical sense of each person only having two eyes, TMC operators cannot monitor all cameras at all times. Therefore, video analytics which makes increasing use of artificial intelligence (AI) and deep learning technologies is needed to actively scout for and identify traffic issues.

Number of false alarms an important factor in creating trust

When a suspicious event occurs, the intelligent cameras will send an alert event to the TMCs along with the video from the location so that a traffic center operator can make further analysis and take appropriate actions. The number of false alarms a system like this report is of course an important factor in creating trust in the system. If operators need to re-verify many reported alarms which do not ultimately turn out to be actionable is not only a loss of productivity but degrades trust in the system. Equally important as a limited number of false alarms, of course, is that the system should not miss any real incidents and accidents that require action, as this would undermine the whole purpose of the system.

The time taken to detect an incident is also an important factor. If an accident is not detected quick enough, the risks of secondary accidents caused by the first grows with the amount of time passing before action is taken (for instance variable message warning signs ahead of the scene of the accident). And as we’ve all experienced, the traffic flow usually slows down significantly not only in the direction of the accident itself, but also in the opposite carriageway due to the phenomenon of “rubbernecking”. People are curious about what has happened on the other side of the road, which often results in further incidents.

Break-through in artificial intelligence for image processing

In the field of computer science, artificial intelligence (AI), has seen several new break-throughs in the recent years. Particularly in the field of image processing, the concepts of Machine Learning and Deep Learning have dramatically improved the possibilities to classify and make computable objects of real events, actions, people and vehicles.

For instance, our latest innovations in Deep Learning video analytics methods have helped us to drastically reduce the number of potential false alarms, compared to traditional pixels-based detection technologies, while maintaining a high accuracy at detecting actual meaningful situations.

One of the main challenges of traditional video-based analytics technology for traffic surveillance (and the main source of false alarms) are shadows, either vehicle shadows or more static shadows of trees, clouds or structures. The very concept of the Deep Learning technology is to ‘teach’ a computer to identify and classify objects. In the context of traffic surveillance, the computer is trained to identify vehicles (cars, trucks, motorbike, etc.) and therefore to effectively ‘ignore’ the shadows.

Once detected the ‘objects’ become data. Citilog’s know-how and expertise are to manage these types of data and create the appropriate output for the operators, in other words: does the data identified through the Deep Learning process represent an actual situation that the operators need to be aware of?

Of course, the accuracy of the Deep Learning engine is highly correlated to the quality and definition of the image to start with. This is where the expertise of Axis in producing high-quality cameras, delivering high definition images in the visible spectrum, becomes a key component to the process.

Video surveillance, with Axis being the leader, has long ago entered the age of digitization and by doing so has generated a huge amount of data. The next challenge is to turn this vast amount of digital data into information that is both impactful and actionable, therefore fulfilling the key needs and objectives of the video surveillance users.

In the field of traffic surveillance, Citilog and Axis are combining their decades of expertise to help drive towards Citilog’s VisionZero strategy for zero false alarms, zero missed incidents and zero lost time, which of course supports the Axis mission to create a smarter, safer world.

Learn more about how Citilog and Axis can support smarter roads and minimize traffic incidents:

Solutions for smarter roads

 

Jean Marie Guyon is the Sales and Marketing Manager at Citilog. With a background in engineering and 18 years of field experience, Jean-Marie has a deep understanding of both the technology and market requirements for the Intelligent Transport System (ITS) industry. He joined Citilog in the early days of development and helped establish the brand as a global name within the industry. Citilog analytics-based solutions address both the current and future needs within modern transportation and mobility management.