What is Low-Power Edge AI Computer Vision?

When you think of computer vision, you may think of high computer performance for processing high-resolution images and the use of the cloud to achieve it. However, human and object detection, image classification, object tracking, and human pose estimation can be inferred on the Edge side with low power consumption, and there is no need to upload to the cloud, so there are advantages from the viewpoint of privacy. I have.

For example, detection of people and people, fall detection, digital / analog meter reading, etc. can be achieved with a low frame rate of 10fps or less, image resolution of VGA/QVGA, and monochrome / grayscale. In addition, the size of the AI model is 2 MB or less, the processing speed is 10 GOPS or less, and the power consumption during inference is 100 mW or less.

Market demand for Low-Power Edge AI

Batteries can be used when adding Edge AI to an existing environment or when installing in a location where power supply is difficult, such as outdoors. However, considering maintainability, it is desirable that the frequency of battery replacement is low.

Katana was developed based on the following needs.

・Can operate for about 2 years on battery drive

・Want to downsize equipment

・Easy to develop and install

・Flexible and supports multiple AI domains (*)

　*In addition to images, Katana can handle various sensing such as voice, temperature, pressure, and PIR.

Device features

Katana is a device featuring low power consumption, multi-core SoC, and easy system integration.

・ Low power consumption: less than 1mW /MHz when active, less than 800uW when idle, less than 100uW during deep sleep

・Multi-core SoC: NPU specialized for image / audio AI, ARM M33 and Tensilica HiFi3, advanced DSP x2

・ Ease of system integration: Complete software tools (SDK), AI compiler and development platform, Wireless Connectivity solutions, etc.

Katana internal block diagram

Katana inside the 6 There are two cores, specifically: AI is an accelerator for Neural Network Processing Unit (NPU) core x2, advanced for time series analysis DSPs core x2, run the application software ARM M33 core, for audio processing HiFi3 Tensilica DSPs is.

again, 2MB on-chip memory, cameras and microphones, PIR Various digital interfaces that can connect sensors, acceleration sensors, etc. Wi-Fi/BLE Connection interface with chip, numerous GPIOs there is.

Examples of intended applications and use cases

Assumed application

Katana is a good fit for five applications:

 

・ Smart building: people counting, space optimization, detection of facility rule violators

・ Support for independent living: monitoring, well-being, detection of specific sounds (such as the sound of breaking glass), voice communication

・ Industry 4.0: Readout of digital / analog meters, predictive maintenance, tracking / monitoring of objects, event detection by sound and video

・ Smart home / security: Human detection, distinction between pets and humans, parcel detection, specific sound detection

・ Consumer electronics: voice control, voice communication, visual activation, voiceprint authentication

Use Case ~Fall Detection System~

Here, as a specific use case of Katana, we will take up a fall detection system aimed at supporting independent living.

 

Some of the fall detection systems currently available on the market are detected by wearable devices. With this method, there are issues such as difficulty in determining whether a person has fallen or is sleeping, the stress of wearing the device, the need for frequent charging, and concerns about loss. The Edge AI Vision System using Katana does not require special equipment to wear, making life less stressful. Detecting the presence of a person ⇒ Detecting a potential fall by sound ⇒ Detecting a fall by image.

Furthermore, since it is processed by the Edge terminal, privacy can be ensured. When a fall is detected, a notification can be sent to a smartphone, etc., enabling prompt response.

Evaluation Kit (EVK)

Synaptics provides a Katana evaluation kit for various demos and prototype development / debugging.

The Katana evaluation kit consists of:

・ Katana SoC module (red frame on the left): Katana SoC, Flash memory, power supply IC

・ Connectivity module (red frame on the right): Synaptics​ ​Wi-Fi / BLE chip "BCM430132"

・ Base Board: Microphone, acceleration / gyro sensor, illuminance sensor

・Camera module: HiMax"HM01B0 (QVGA)"

The evaluation kit is covered with a plastic housing, and has specifications that can be used for different applications, such as overhead installation assuming people enter and leave the room, and installation on a table assuming a TV, etc. In addition, it is a kit that can be battery driven assuming IoT applications.

demo video

Three types of demos using Katana are also available on youtube, so please take a look.

1) Measurement of people and numbers

https://www.youtube.com/watch?v=KRNLyhVubWE

2) Readout of digital meter

https://www.youtube.com/watch?v=cfP3HVvIZig

3) Fall detection

https://www.youtube.com/watch?v=NT0XIXVXnnc

Contact Us

Katana excels at AI processing of images, and is a battery-powered (low power consumption) device that can perform processing such as room entry/exit detection, number counting, location detection, and fall detection. If you are interested in detailed device specifications, other specific use cases, Katana development procedures, etc., or if you have any other questions, please contact us.