Written evidence submitted by Hazen.ai (RSM0053)

 

Executive Summary

Over recent years, questions have been raised over the safety of “Smart Motorways” as a result of a number of traffic deaths and serious incidents which have occurred as a result of vehicles stopping in an active lane and becoming stranded. With Smart Motorways a key part of the transport strategy to increase capacity on the UK’s road network, a solution to this must be found. This paper looks at how new Artificial Intelligence based systems can offer a cost-effective solution to rapidly detect stranded vehicles and alert emergency services and local road users.

Our key finding is that retrofitting the existing CCTV camera network already deployed on the motorways with new generation video-analytics software is the quickest and most cost- effective way to address Stopped Vehicle Detection. These are the same type of algorithms that are now being extensively used in self-driving cars and can be installed inside the Regional Traffic Control centres to generate automatic alerts for the operators. This strategy avoids the overhead of installing expensive Radar sensors through-out the motorway network and requires no new sensor deployment, making it an attractive option to improve safety on the smart motorways.

 

In this document

 

 

The Problem

The very concept of smart motorways is under intense scrutiny after Sheffield’s Coroner, David Urpeth, concluded in January that the lack of a hard shoulder contributed to the death of Jason Mercer and Alexandru Murgeanu on the M1 in South Yorkshire in June 2019.

 

Mr. Urpeth termed the smart motorways “an ongoing risk of future deaths”. Mercer’s and Murgeanu’s vehicles had collided slightly, but due to lack of a hard shoulder, they were stopped for more than six minutes in an active lane when a lorry crashed into them. Over the past two years, this is the fourth time a coroner’s report has raised concerns about safety of the smart motorways.

At the core of the issue are the “dynamic hard shoulders” introduced as early as 2006, which aim to alleviate congestion by converting the hard shoulder into an active traffic lane during congestion hours. While the availability of an additional lane does increase the capacity of the motorway and reduces congestion, it removes the crucial safety space for broken down vehicles to stop on, hence raising an important safety concern.

A review of crash data does indicate that smart motorways with dynamic shoulders are statistically as safe, if not more, than the regular motorways. However, this does not eliminate the concern that some deaths on motorways could have been avoided if the hard shoulder was available for a broken-down vehicle to maneuver onto. The provision of spaced and monitored refuges - supposed to answer this issue - has not been seen as sufficient, and simply more of them would not solve the problem.

The latest iteration of smart motorways includes the “All Lane Running” or ALR strategy, introduced in 2014, which permanently converts the hard shoulder into a lane. This is presumably done to avoid confusing motorists by its “dynamic” change. In fact, this is what Grant Shapps, the Transportation Secretary, promised in his 18-point action plan last year that aims to make smart motorways safer. Amongst other measures, the 18 points include “abolishing the confusing dynamic hard shoulder”. If anything, this means that the hard shoulder for emergency stopping will be even more scarcely available.

 

Another important measure in the action plan is the proposal of “substantially speeding up the deployment of Stopped Vehicle Detection technology across the entire ALR motorway network” with the goal of reducing response time from an average of 17 minutes to 10 minutes”. The Stopped Vehicle Detection, or SVD, technology implies a network of sensors that can quickly identify an incapacitated vehicle on the motorway and alert the Traffic Control Centres so that help can be dispatched quickly. So far, according to recent reports, only 37 miles of the smart motorway network has some sort of SVD technology deployed. Shapps has promised to speed up the roll-out of SVD technology and complete it by 2023. Safety advocacy groups have expressed frustration with the delay and want the completion to be expedited.

 

Limitations of Current SVD Strategies

As mentioned above, the government's action plan to improve safety on smart motorways identifies the need to improve Stopped Vehicle Detection and response times. There are two existing strategies for SVD. Firstly, there is a network of CCTV cameras in operation which is continuously scanned by human operators to identify any incident. In addition, Radar-based SVD sensors have been trialed on some sections of the motorway. We discuss the limitations of both these solutions below.

The traditional solution of SVD is by human operators in Regional Traffic Control Centres who continuously monitor feeds from the many thousand PTZ cameras that are installed around the motorway network. They scan the camera feeds for any incident as well as for congestion and traffic management parameters. While these types of control rooms serve a very important purpose, they have important limitations in relation to SVD. Research has shown that human operators have limited attention spans – as low as a few minutes only – when monitoring multiple surveillance feeds, because of loss of focus and the mundane nature of the tasks. Moreover, even when an event is correctly identified, it may be after a significant delay. Analysis of past events has shown that the it takes on average seventeen minutes for an RCC operator to detect stopped vehicles and to deploy the necessary resources and alerts.

 

A more advanced approach to SVD is to is to deploy Radar-based sensors at intervals along the motorway, to generate an automatic alert as soon as a vehicle is detected to be stopped. While it is understood that trials of these systems have had some success, they have also highlighted some potential issues. Firstly, rolling out these systems for the entire smart motorway network, which could be up to 788 miles long by 2025, will be a very expensive undertaking. Moreover, new installations take time and have already been delayed. Till now, the aggregate coverage of Radar-based systems is just 37-miles. Each Radar-unit provides coverage of about 250m only. More importantly though, this Radar technology is not very accurate, and has particular difficulty in distinguishing smaller, static vehicles against background infrastructure. A recently released review of Highways England’s 2016 trial of Radar technology on a 13km stretch of M25 shows that up to 2,400 stopped vehicle incidents could be missed entirely each year on just this 13km section!

 

The Solution: Retrofitting Existing CCTV Network with AI

The new generation of AI-based video analytics technology, termed deep learning, can solve the basic SVD problem as well as address the disadvantages of radar-based systems. Video analytics technology has matured significantly in the past few years and can now out- perform Radar-based sensing in terms of accuracy. It is the same technology that is being relied upon in self-driving cars extensively.

The most important advantage of using video analytics for SVD on smart motorways is that it does not require the installation of any new sensors. Rather, video-analytics can work on the extensive CCTV camera network that is already deployed by Highways England. The video feeds are already being transmitted to the RCCs and can be utilized by the video analytics software without compromising their original purpose. No field work on the motorways themselves will be required. This makes video analytics a particularly attractive technology from cost and ease of deployment perspective.

 

The technology is highly accurate and generates less false alarms compared to Radar sensors. This is because it relies on “visual cues” of what constitutes a vehicle and does not suffer from the problem of distinguishing between background infrastructure and objects of interest. In our trials of using video analytics for SVD, we were able to reliably establish a stopped vehicle within 5 seconds of its stopping on a road at distances of up to 400m from the camera. Our AI software can also determine the distance of the vehicle from the camera, pinpointing the exact location at which the incident has occurred. Moreover, it can generate alerts when a pedestrian is observed on the motorway, or debris.

 

 

Video analytics technology can handle difficult traffic scenarios, can identify object types, and can provide much more accuracy than Radar-based systems at the fraction of the cost.

 

The whole setup can be installed in the existing control centres where video feeds from cameras are already available and does not require new camera installations. We calculated that with the new generation of GPU machines available, a single server rack in the control centre can conservatively support more than 200 camera feeds. Assuming a camera to have an average road footprint of 200m, this implies that 40km of bi-directional road network can potentially be monitored by the server space of a single rack unit.

An important question in retrofitting existing cameras is whether they can provide uninterrupted coverage if the camera is being moved around. All PTZ cameras can be preset to a few views amongst which they keep cycling. If the time-period of that preset is set to be, say, 15 seconds, then this can ensure that a stopped vehicle can be identified within maximum 20 seconds. Moreover, the specific technology that we have built self-calibrates to each view and identifies it when the camera returns to that view. This means that the distance calibration parameters of each preset view are retrieved instantaneously with the camera returning in that view, providing seamless sensing and coverage. Finally, even if gaps in coverage are identified, installing a new CCTV camera is a lot simpler and cheaper than putting up a new Radar.

 

Conclusion

The smart motorways conundrum needs a quick solution which can bolster the public’s confidence in their safety. Decades old expensive Radar technology is not the solution. Rather, we need to innovate using the latest AI algorithms that can retrofit to the existing infrastructure. This is the quickest, smartest and most cost-effective way forward.

 

About the Authors

Sohaib Khan Ph.D. is the CEO of Hazen.ai, an award- winning startup focused on using AI for road safety. He holds a PhD in computer vision and has over 15 years of academic experience, including leading a premier computer science department, before transitioning from the academia to lead his own startup.

 

Steve Hill has led large engineering firms in the defense sector for over 20 years. He was the Managing Director of SEA, UK for 10 years, and before that, he spent a decade in leadership positions at Thales. As MD of SEA, he has supervised the introduction of several successful road-safety and enforcement products.

 

Meredydd Hughes CBE QPM has served in four of the largest UK police forces and was the Chief Constable of South Yorkshire Police for 7 years. Nationally, he was the Head of Roads Policing, and the Uniformed Operations Business Area. He works as an international consultant with keen interest in the introduction of effective innovative technology.

 

 

April 2021