EMBEDDED DEVICES
Figure 1: Smart camera application flow
Flexible AI solutions for smarter cameras and edge AI appliances
NXP explores how by applying AI to the video streams of high-resolution cameras can enable them to be deployed in many new applications
A
pplying AI to the video streams of high- resolution cameras can enable them to be deployed in many new applications. However, that means that AI solutions at the edge must perform neural-network inferencing on streaming video at different frame rates and varying degrees of complexity. NXP Semiconductors’
i.MX applications processors provide AI capability for many levels of performance through combinations of CPUs, GPUs, and sometimes even neural accelerators. However, it’s important to remember that an increasing number of edge AI applications - particularly in retail, healthcare, security, and industrial control must handle much higher video frame rates (30fps or more); resolutions at 1080p and beyond; and be able to run in parallel with other models.
One way to achieve this is by combining NXP’s
i.MX application processors with Kinara Ara-1 edge AI processors, resulting in a low-cost, low- power system that delivers the very high level
20 NOVEMBER 2022 | ELECTRONICS TODAY
of AI compute power required by increasingly demanding applications.
Smart cameras critical in deployment of AI
Smart camera vision processing and neural- network inferencing support are integrated into the main camera board using NXP’s
i.MX application processors combined with Kinara’s Ara-1 edge AI processors. That enables cameras to perform complex AI tasks before sending the application’s metadata to an edge server, or to the cloud for non-real-time analytics. When host processors like the
i.MX 8M Nano and
i.MX 8M Plus SoCs are combined with the Ara-1 processor they instantly become extremely well suited for low-power smart- camera design. You can see one such design with an image sensor,
i.MX 8M Plus applications processor and one or more Ara-1 chips integrated on a single camera board in Figure 1(a). Since the
i.MX 8M Plus SoC includes
dual image signal processors, it can interface directly to a camera sensor and convert raw sensor data to the RGB format needed for display, storage, and inferencing. In case of other NXP SoCs such as
i.MX 8M Nano that don’t include an ISP - an image sensor with an integrated ISP can be used for the smart camera design.
AI inference flow for smart camera The host processor pre-processes video frames from the sensor to transform the sensor data into a format that matches the input requirements of the AI inferencing performed on the Ara-1 (Figure 1). The pre-processing tasks including normalisation, mean-subtraction, scaling, and quantisation to the int8 data format are performed on the host processor using CPU cores or the GPU.
Quantised input is then sent to Ara-1 over a USB or PCIe interface for the inference. The inferencing is followed by post-processing
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44
