Automotive camera SoC combines AI with low power consumption
The CV22AQ provides the performance necessary to exceed New Car Assessment Program (NCAP) requirements for applications such as lane keeping, Automatic Emergency Braking (AEB), intelligent headlight control, and speed assistance functions, the company claims. Fabricated in advanced 10nm process technology, its low power consumption supports the small form factor and thermal requirements of windshield-mounted forward ADAS cameras.
According to Ambarella CEO Fermi Wang, to date, front ADAS cameras have been performance-constrained due to power consumption limits inherent in the form factor. Despite these restrictions, the new CV22AQ provides a combination of outstanding neural network performance and very low typical power consumption of below 2.5 watts, he added. This allows tier-1 and OEM customers to greatly increase the performance and accuracy of ADAS algorithms.
The CV22AQ’s CVflow architecture provides computer vision processing in 8-Megapixel resolution at 30 frames per second, to enable object recognition over long distances and with high accuracy. CV22AQ supports multiple image sensor inputs for multi-FOV (Field of View) cameras and can also create multiple digital FOVs using a single high-resolution image sensor to reduce system cost. It enables DNNs for object detection, classification (i.e. of pedestrians, vehicles, traffic signs, and traffic lights), tracking, as well as high-resolution semantic segmentation for applications such as free space detection.
The CV22AQ’s Image Signal Processor (ISP) combines low-light capabilities with High Dynamic Range (HDR) processing. It includes 8-Megapixel encoding in both AVC and HEVC video formats, allowing customers to add video recording and streaming capabilities to their automotive cameras. The SoC’s cyber security features, which include secure boot, TrustZone and I/O virtualization, enable over-the-air updates (OTA) and also protect against hacking.
A set of tools is provided to help customers port their own neural networks onto the CV22AQ SoC. The toolkit includes a compiler, debugger, and support for industry-standard machine learning frameworks such as Caffe and TensorFlow, with guidelines for DNN performance optimizations.
CV22AQ is currently sampling to leading tier-1 customers and tier-2 algorithm providers. Chip samples with ASIL-B support are targeted to be available in 2019.
For more information, please visit www.ambarella.com.