Image recognition applications in the era of 5G
Ericsson Research and the start-up company SarvAI recently showed how a customized AI-based image recognition application can be deployed within the network, turning cheap webcams into sophisticated image analysis systems. Read below to find out how we did this and what it means for society.
The Urban ICT Arena (UICTA) is an Ericsson test network, based in Stockholm, where industrial and academic partners can explore new application ideas which leverage the full capabilities and performance of 5G. In the K5 project – the Kista 5G Transport Lab – Ericsson Research, together with partners RISE-Acreo and the Swedish Royal Institute of Technology (KTH), work with innovators to help showcase new ideas in the testbed.
5G, distributed intelligence and future image analysis
The 5G system integrates distributed cloud in the network, making it possible to deploy custom applications locally, closer to the end user. In the last months, we have worked together with SarvAI to demonstrate how distributed AI can be deployed in the 5G network. SarvAI has a flexible image analysis solution based on AI, which can be adapted for a variety of use cases taking advantage of simple cameras connected to a 4G/5G network. Most current solutions either rely on complex cameras with dedicated software leading to expensive solutions, or fully-centralized processing leading to longer analysis time and huge bandwidth requirements from connected cameras. Estimates indicate 1 billion cameras connected by 2020 which would result in heavy load on both networks and centralized datacenters for the actual processing of the video feeds.
Distributing the image analysis to local processing near the user can reduce the load on the network as only results from the image analysis is transmitted further in the network. Transfer of images or video feed to an operator center would be enabled once an alert is raised by the system. In addition to the obvious saving in network bandwidth this also means that the time needed for the analysis can be reduced. In some applications this is critical, for example traffic safety – camera solutions can be put in place to detect hazards in road traffic. This can be either vehicles misbehaving or people present in non-safe areas such as red-light pedestrian crossings or railroad tracks. By deploying the image analysis locally, reliability can also be improved which is important in e.g. industrial processes. When image analysis is included in the production flow, any disruption or delay would impact the throughput of a factory and have a direct economic impact.
In the demonstration we set up a few different use cases which show the flexibility of the platform and how simple web cameras can be utilized for advanced image recognition applications.
Use cases for image recognition applications
The first case relates to advertising, where we implemented a platform for measuring audience engagement for advertisement billboards in the UICTA testbed in Kista. For better planning and production values, advertisement companies need information regarding the demographic that engages with their billboards at different locations and times. Currently, this information is mainly collected through street surveys after major campaigns. In this demo, we demonstrated how using AI and computer vision makes it possible to collect this information and present statistics like the diagram below.
Find out more about this case study on the SarvAI YouTube channel.
Similar techniques can be used for security and monitoring purposes, where a common task would be to detect people in a video feed and extract information related to persons present. In this example, a construction site is shown, where a group of people is identified. In the analysis, persons are detected in the image and the system is trained to detect whether they are wearing the required security helmet. In the example below, one person is identified without a helmet and an alert is raised.
This can also be applied to traffic supervision where it would be possible to analyze the traffic situation for different purposes such as safety, traffic condition monitoring, smart traffic routing or traffic information systems. The street scene below shows some of the items which could be extracted, including pedestrians, vehicles (civilian car, taxi, etc.) and bicycles.
Privacy in image recognition systems
An important question with any camera supervision is privacy, especially as computer vision systems become more capable and can identify increasingly specific details. Here the distributed cloud in 5G would be a tool. If we analyze the video feed locally, close to the camera, and only transmit the results of the analysis to a control center, some aspects of privacy can be addressed.
For example, an operator in a control center could see a video feed like in the following image (below). The red rectangle represents a person who has been detected in the video analysis, but only the meta data representing the number of detected persons and their location is sent to the control center. In the control center, the results of the video analysis is overlayed on top of an empty scene from the camera resulting in an anonymized video feed. Thus, the solution here still indicates how many persons are in the scene and where, but individual privacy is kept in a way that the sensitive data captured does not even leave the geographical local area.
The work we did with SarvAI demonstrates how innovative solutions can be built by embedding smart software into the 5G network. This is one example of how we engage with startups, established companies and universities to create new applications which build value on top of 5G capabilities.
Find out more
Visit our Ericsson ONE page to learn more about how Ericsson works with innovation.
Learn more about how AI can be applied to smart traffic management,