- The STM32 AI ecosystem provides essential building blocks for neural networks to run on STM32 MCUs
- It enables a cost-effective and power-efficient solution. Various deep-learning frameworks such as Keras, TensorFlow Lite, and ONNX exchange format are supported natively
STMicroelectronics and Schneider Electric are demonstrating a prototype IoT sensor that enables new building-management services and efficiency gains by understanding building-occupancy levels and usage. They have collaborated to integrate Artificial Intelligence (AI) into a high-performance people-counting sensor, which overcomes the challenge of monitoring attendance in large spaces with multiple entrance point. The advanced IoT sensor has been developed by combining the expertise of ST’s AI group and the deep sensor-application expertise of Schneider Electric to identify and embed a high-performing object-detection neural network in a small microcontroller (MCU).
Flexibility and efficiency in hardware design
Maxime Loidreau, IoT Sensors Program Manager at Schneider Electric said, “This promising technology opens a new solution for attendance monitoring and people counting in numerous applications such as monitoring queues, building usage, and social distancing. Our innovative demonstration, created with STMicroelectronics, finds applications in various segments, from hotels to offices and retail, and more generally any building where knowing attendance and space occupation has a value. This will redefine the building of the future!”
Schneider Electric’s increase in design productivity comes from its use of the STM32Cube.AI toolchain, which has mature capabilities for developing AI applications for the broad portfolio of STM32 MCUs. This allowed Schneider Electric to gain valuable flexibility and efficiency in hardware design from the engineering resources, sophistication, and ease of use provided by the STM32Cube software-development ecosystem. The prototype people-counting sensor combines a LYNRED ThermEyeTM family thermal imager, integrated in a unique ultra-low-power design created by Schneider Electric, with a Yolo-based Neural Network model running on the recently introduced high-performance STM32H723 MCU from ST.
Miguel Castro, AI Solutions Business Line Manager at STMicroelectronics said, “This project demonstrates the power of deep learning to enhance embedded data-processing performance, showing how high-value applications can be hosted on a cost-effective microcontroller-based platform. Our STM32Cube.AI ecosystem empowers users to create flexible solutions within a fast time-to-market window. Customers can enjoy even greater productivity leveraging the support of our technical team to overcome engineering challenges.”
Essential building blocks for neural networks
The STM32 AI ecosystem provides essential building blocks for neural networks to run on STM32 MCUs. It enables a cost-effective and power-efficient solution. Various deep-learning frameworks such as Keras, TensorFlow Lite, and ONNX exchange format are supported natively.
They said, “Included in the ecosystem is the X-CUBE-AI software expansion package, which extends the capabilities of the STM32CubeMX initialization tool to automatically convert pre-trained neural networks, generate optimized libraries for the target MCU, and integrate these into the user’s project. Additional support to automate laborious development tasks includes several ways of validating neural network models and measuring performance on STM32 MCUs without creating the necessary C code by hand.”
The general DNN approach supported by ST’s software-development ecosystem, mapped onto the STM32 portfolio, lets users efficiently replicate development effort to create products for multiple markets. The company said that the STM32H723 MCU powering the demonstration at ST Live Days has excellent credentials for hosting AI applications, including high core performance, up to 1Mbyte Flash, high-speed off-chip memory interfaces, and integrated features for connecting a wide variety of sensor types
