Person recognition - that is, the detection of a person, not its identification - is probably one of the best-known applications for machine learning: neural networks are ideally suited for recognizing objects, people, animals, etc. in images in order to automatically classify and categorize image material. But often these algorithms have two key drawbacks: Person recognition algorithms are very resource-intensive and often take place online in the cloud, precisely because they have a high resource overhead. Associated with this are problems in protecting personal and business data. These disadvantages massively limit the use of such technology if privacy constraints play a role.
Fraunhofer IMS has therefore set itself the goal of researching and developing a person recognition system that is capable of recognizing persons with camera systems in a privacy-compliant manner on low-cost embedded microcontroller systems without the use of any online services. Based on the specially developed software framework "Artificial Intelligence for Embedded Systems" (AIfES), Fraunhofer IMS researches the integration of image processing algorithms and neural networks on embedded systems.
Instead of using huge neural networks, our approach is to first formulate the detailed requirements for the respective application in order to develop specifically adapted algorithms from them. In this way, resource-intensive deep neural networks can be avoided and upstream algorithms can be used to create the smallest neural networks which can even be used on a microcontroller without any problems. This means that the entire image processing and recognition in real time can be implemented locally on an edge device, since the microcontroller used can be integrated here without any problems. No additional servers, services, clouds or other interfaces need to be connected.
The fields of application for person recognition are extremely diverse: For example, cameras in security systems can be equipped with this technology to give the camera the ability to make its own decisions about the relevance of the recorded data. This can reduce the required data bandwidth, for example, by forwarding only relevant data (people in the picture and not just a cat, for example). Due to the extremely low system costs applications in the field of smart homes, vacancy monitoring, traffic engineering, evacuation systems, etc. are obvious examples of areas of application.
Person recognition itself is only one example of a large number of possible recognition algorithms. For example, the recognition of workpieces, vehicles or other moving objects are also conceivable applications that can be implemented on miniature systems with the help of specialized neural networks. The AIfES framework of Fraunhofer IMS offers the appropriate platform to develop a customized solution for your application without much effort.