In November 2018, Geovation start-up FlowX (in partnership with Vivacity Labs) won a £50k contract with the Department for Transport & Digital Greenwich. This was part of the GovTech Catalyst process led by the Government Digital Service (within the SBRI — Innovate UK framework).
The competition focused on “harnessing the power of data to better understand and respond to road congestion”.
The competition is split into two phases.
- Phase 1: £50k feasibility study to develop an early prototype
- Phase 2: £500k R&D contract to commercialise the prototype
Phase 1 was recently completed and we submitted an ‘End of Phase 1 Report’. Below is a high-level overview of that report.
How it works
Step 1 — We integrate with a transport authority’s existing CCTV network. We ensure to have in place extremely robust data security agreements.
Step 2 — We use the latest computer vision techniques to anonymously classify objects (car / pedestrian / cyclist / truck etc).
Step 3 — We track these anonymous, classified objects across the screen — creating anonymous, granular data (counts / occupancy / speed / path etc).
Step 4 — We use this anonymous, granular data as an input into machine learning models to flag when traffic conditions are abnormal.
Outcomes from Phase 1
1 — First of a kind
To the best of our knowledge, this is the UK’s first integration of deep neural networks on existing highways CCTV cameras in production.
We have successfully integrated with a total of 14 cameras across Devon, Leeds, and Exeter Councils.
2 — Repeatable data privacy agreements
Data privacy is rightly a primary concern for stakeholders at local authorities, so we have in place extremely robust data security agreements.
3 — Integrated with analogue cameras
Exeter & Devon Council cameras are analogue. We therefore needed to install Analogue-to-Digital converters for each feed. This provides us with confidence to tackle the large proportion of existing public CCTV feeds which remain analogue.
4 — Solved the Pan-Tilt-Zoom (PTZ) problem
Most existing public CCTV cameras in the UK are not fixed, and instead can be moved by an operator for a better view of the roadspace. This means a simplistic system returns erroneous data after a PTZ camera is moved.
We therefore wrote software which can automatically ‘understand where it is looking’. This method successfully identifies a pre-set view at any time of day, even if the position is 30% out of alignment from the exact pre-set in direction or zoom. The count line and zone positions adjust automatically.
5 — Data for strategic decision making
We are continuously extracting anonymous, classified data from the CCTV feeds. This data itself is useful for longer-term strategic decision making, providing reliable, city-wide, real-time information of vehicle traffic, cyclists and pedestrians.
6 — Prototype system of real-time incident detection
Using unsupervised anomaly detection, we flag when traffic conditions deviate past a configurable standard deviation — based on the norm for that day of week at that time of day. Traffic conditions are defined by their flow (counts / occupancy / speed / path etc).
Real-time incident detection helps operational traffic control centres:
- Respond to incidents in real-time
- Prevent the build-up of congestion
- Be less resource-intensive
Transition into Phase 2
Our key users are traffic operators in traffic control centres. Our principal priority moving forward is to add genuine value to our users, building a robust system of real-time incident detection which seamlessly complements existing workflows.
We will find out soon whether Phase 2 is going ahead.
Watch this space 👀