With estimates of 137,013 reported casualties in road collisions – 1,760 of which resulted in fatalities – this past year in Great Britain alone, the latest numbers for road safety are dire. Safety is a major cornerstone for quality of life and despite the scope for technology to drive progress, there remains a need for functional and impactful solutions. If we wish to deliver on better measures, computer vision technology centred around humans should be the priority.
Roads are a crucial and arguably underrated part of economic development. Road survey vehicles can be deployed to monitor damage, alongside the use of sensors and cameras. But poor management, disorganised resource allocation and time-consuming surveying methods result in an expensive problem in need of an inexpensive and rapid solution.
AI computer vision systems offer a possible avenue to address this. However, they require large volumes of training data to ensure detection is not reliant on features of a particular environment, as well as to ensure high levels of detection of deformities like cracks and potholes. Issues arise when real-world training data does not optimise vision systems to identify unusual damage, however infrequent the cases may be.
There were 108,542 motor vehicle thefts in England and Wales this past year. While innovation in the automotive industry continues to grow, crime rates remain a concern. Despite advancements in anti-theft technology, a lack of significant reduction in theft suggests there is an evident need for better solutions to tackle the problem. The demand, opportunity and strategies for stealing vehicles are still prevalent and evolving. Vehicle checking systems, cameras, sensors and AI technology are just some of the methods that can be leveraged to increase vigilance.
However, again these technologies have an endless appetite for data to ensure accuracy and hold the ability to respond, in this case, to car theft. Collecting and labelling this data is expensive, time consuming and, in some cases, unfeasible. This poses a serious problem if we wish to see a reduction in crime rates.
Due to the growing population, the amount of waste produced by city dwellers is on track to reach six million tonnes a day by 2025. This is on top of rising costs of waste disposal, with the World Bank predicting collection expenses could top $375bn in five years.
Urbanisation and a scarcity of resources are a couple of the key issues making waste management an area in need of significant improvement. Currently, waste collection systems tend to be on fixed schedules and routes which can result in unnecessary pick-ups and blockages. The subsequent knock-on effects of congestion and fuel consumption make the process even more costly and unsustainable than it needs to be.
The use of IoT solutions and AI-powered computer vision systems can solve some of these issues by indicating when action is actually required. These solutions can take the form of sensors, gateways, cloud platforms, and web or mobile applications. Through optimising remote diagnostics and route planning, this can reduce congestion, improve air quality and prove to be more cost-efficient. All of which will help meet the increasing demand for sustainable solutions.
Where does synthetic training data fit in?
Effective innovation requires accessible and robust services. Existing AI systems are facing problems with training data supply, in addition to the training data becoming out of date (data drift). The dominant approach to data acquisition in computer vision relies on labour-intensive collection and labelling of real-world images.
A potential solution has emerged in the form of synthetic training data. It is artificially generated from computer systems and provides the opportunity to produce greater volumes of accurate training data quickly and more cost-effectively. But, beyond this, it serves a greater purpose of supplying training data for any edge-case required.
This prepares AI systems to deal with a diverse range of environmental conditions - the ones typically needed in dangerous scenarios. It can contain labels for dimensions that cannot be reliably quantified by humans. In the waste management industry, for example, training with synthetic data can improve the dexterity of sorting technology by providing vectors such as range and velocity. It can also allow sensors to reliably detect car theft, as well as alerting the occurrence of road damage by training vision systems to detect deformations. Overall, synthetic training data can be used to ensure systems are responsive and challenge the boundaries set by real-world data.
Furthermore, synthetic training data is not subject to privacy concerns, meaning there is an opportunity to be transparent with citizens by showing how waste is being managed, how new technologies are beneficial and how operations are being conducted to strive for more sustainable development. Synthetic data trained computer vision systems are undeniably beneficial for promoting safer streets, whilst simultaneously reducing costs and boosting productivity. Hopefully, further implementation of these initiatives will provide the stepping stone to safer and smarter cities.
Steve Harris, CEO of Mindtech