Detecting incidents on CityLink
Optimising incident detection performance
The incident
All drivers are familiar with the frustration of peak hour traffic — the rhythm of your regular route is broken, an area that normally freely flows is at a standstill. You creep forward following the bumper to bumper brake lights ahead. As you get closer, you see the cause. Vehicles try to merge as their drivers look around curious and frustrated. The clock on the dashboard indicates the delay.
What’s happened? Hazard lights, a broken down car and a confused driver looking back at the disproportionate chaos and checking their phone waiting for help to arrive. The consequence is clear: a small section of a single lane out of action has ripple effects that slow down an entire freeway and impacts the road network throughout the city.
The challenge
For road operators, considerable effort goes into minimising the risk, but incidents happen — approximately 300 per week on CityLink. How these incidents are detected and managed is crucial. We know:
1. The faster an incident is detected, the sooner its impact can be minimised. The Traffic Control Room can adjust speeds to reduce risk factors like lane speed differences, and close lanes early to allow smooth merging. This reduces traffic flow breakdown in response to an incident.
2. The faster the incident is cleared, the lower the impact on the rest of the traffic. When we can clear an incident from the road and open the lane quickly, the network as a whole benefits. The longer the incident obstructs traffic, the greater the flow on effects.
3. Early intervention in a fire or medical emergency is essential for survival. In the event of a road user being injured or vehicle fire, early intervention is essential for road user safety.
In each of these scenarios: minutes, even seconds, matter.
These priorities were front of mind when Transurban implemented the safe clearance incident response model on CityLink, which resulted in incidents being cleared faster and safer, delivering a 55% reduction in clearance times.
CityLink also uses video incident detection technology to bring the incident to the attention of Operators in less than a minute, often within 15 seconds to enable early intervention as well as clearance. This type of incident detection outperforms traditional incident detection technologies such as traffic loops and studs which relies on traffic flow breakdown to happen prior to detecting the cause.
Our roads are safer and the network runs better as a result of these approaches.
The shortfall
Delivering the safety and network performance benefit using fast, high performing detection technologies such as video analytics has its challenges. In short:
1. Video Analytics can’t detect everything — CityLink video incident detection works very well — a detection rate of 95% of incidents. But 1 in 20 incidents rely on the eyes of our Traffic Control Room Operators watching over 400 cameras to identify what the system has missed. Unlike systems, this is a difficult task for a person to do reliably over long periods.
2. High Detection performance means false positives — The higher your detection rate and the faster you detect, the more false positives an Operator must manage. Under some conditions this can mean 9 in every 10 alarms are false, which due to human factors essentially “trains” the Operator to expect any given alarm will be false.
With current technology, we are unable to cut down the number false alarms without impacting how well we detect incidents. Operators are tasked with manually reviewing each alarm triggered, which at times can be incredibly difficult.
Given this shortfall, solutions that can deliver false alarm reduction are a high priority for CityLink, closely followed by how we can close this gap towards 100% detection performance.
The solution
As part of our Second Horizon Innovation projects, Transurban tested the hypothesis: can combining multiple detection solutions outperform a single detection solution? Using traffic speed as a second source, data and video was processed, labelled, and business rules developed to predict if each alarm was likely to be true. For example, a stopped vehicle alarm caused by a shadow is a false alarm, and we can predict this because the traffic speed is unchanged in response, unlike when a vehicle actually stops as a result of an incident.
We tested this and the results showed it had the potential to cut our false alarms in half, so we gave it a name — the “Road Event Analyser” — and implemented it across a subset of CityLink. Leveraging Genetec’s Transportation Sensor Suite module to interface with current systems and deliver the logic, around 40% of false alarms can be eliminated from these cameras during the day.
With the concept proven, we can use this principle to optimise incident detection. From here we will refine business rules (through machine learning), add more data sources (such as AI video analytics), optimise the architecture and expand the solution over the entire road.
The conclusion
Improving detection systems to overcome the operational constraints like false alarms has tangible impacts on how our control rooms will detect and clear the 300 incidents that might slow you down each week.
Much of the technology is invisible as you drive along CityLink, but you can observe the impact when the lane signs prompt you to safely merge ahead of an incident, or an incident response vehicle is helping someone after a breakdown or accident.
As the technology continues to improve, so too will your travel time savings as we continue to refine these functions to achieve ever safer roads and greater network performance, which is critical to keeping our growing cities moving.