Smart home

Smart Mirror / Human Interaction Interface

In smart home environments, an intelligent mirror is an increasingly popular choice for a central user interface, as it provides a smooth, pleasant and unobtrusive control instance for the entire apartment. Personalized information from the outside world and from the house itself converges in this mirror, and certain aspects can be controlled.

In order to personalize the displays and to provide a natural form of control, the use of neural networks is a popular choice, however, the required computing performance typically also leads to high power consumption and poor energy efficiency. Although the computing performance of embedded solutions for neural networks increases, there is no capability of calculating several large networks simultaneously. LEGaTO addresses this problem by the development of a cognitive edge AI platform, capable of handling such use cases without using a cloud computing solution.

Using available off-the-shelf components and existing toolchains for implementation of the smart mirror or other smart home technologies leads to high power consumption. Although this is feasible for a demonstrator, it is a critical showstopper for using the technology in a real-world scenario.

With the help of the LEGaTO cognitive edge AI platform and the LEGaTO toolchain, the processing can be distributed to energy-efficient microserver modules. These show a high degree of application-specific performance, e.g., Nvidia Xavier-based microservers with integrated tensor cores offer a good solution for most neural networks, while FPGA-based microservers are a perfect fit for preprocessing of the sensor data. This combination of several heterogenous microservers allows reaching an acceptable performance for the application while keeping the power consumption at a reasonable level.

Components

Official repository for a MagicMirror² prototype, developed within LEGaTO, is available here.

Scientific publications

• LEGaTO: first steps towards energy-efficient toolset for heterogeneous computing

Events

News