DATE: 2019-2-18
一、參考
https://www.ambarella.com/uploads/docs/CV22-product-brief-consumer.pdf算法
http://www.ambarella.com/about/news-events/press-images/CV22-press-imageapi
https://www.asmag.com/showpost/26902.aspx架構
https://cloud.tencent.com/developer/news/341774app
https://blog.csdn.net/weixin_42229404/article/details/81481728ide
二、CV2二、CV22AQ簡介
Ambarella, a developer of low-power, HD and Ultra HD video processing semiconductors, introduced the CV22 camera SoC, combining image processing, 4Kp60 video encoding and CVflow computer vision processing in a single, low power design.
The CV22’s CVflow architecture provides the DNN (Deep Neural Network) processing required for the next generation of intelligent home monitoring, automotive, drone and wearable cameras. Fabricated in advanced 10nm process technology, it achieves an industry-leading combination of low-power and high-performance in both human vision and computer vision applications.
「CV22 enables customers to deploy cameras with high-performance deep-learning capabilities,」 said Fermi Wang, CEO of Ambarella. Compared with inefficient CPU and GPU solutions, CV22 provides powerful computer vision performance combined with high-quality image processing, efficient video encoding and the low-power operation required to be deployed in mass production. 「It will enable a new level of intelligence in cameras, ranging from person recognition in home monitoring cameras to advanced ADAS capabilities in automobiles,」 Wang said.
The CV22’s CVflow architecture provides computer vision processing at full 4K or 8-Megapixel resolution at 30 frames per second, to enable image recognition over long distances and with high accuracy. It includes efficient 4K encoding in both AVC and HEVC video formats, delivering high-resolution video streaming with very low bitrates to minimize cloud storage costs.
The CV22’s next-generation ISP (Image Signal Processor) provides outstanding imaging in low light conditions while HDR (High Dynamic Range) processing extracts maximum image detail in high contrast scenes, further enhancing the computer vision capabilities of the chip. It includes a suite of advanced security features to prevent hacking, including secure boot, TrustZone and I/O virtualization.
A complete set of tools is provided to help customers to easily port their own neural networks onto the CV22 SoC. This includes compiler, debugger and support for industry standard training tools including Caffe and TensorFlow, with extensive guidelines for CNN (Convolutional Neural Network) performance optimizations.post
安霸半導體公司(Ambarella)推出了一款新型汽車攝像頭芯片，安霸CV22AQ，可爲汽車高級駕駛員輔助(ADAS)系統中的深度神經網絡(DNN)提供計算機視覺。性能
該芯片使用安霸CVflow計算機視覺架構，目標應用包括前置ADAS攝像頭、帶有盲點檢測(BSD)功能的電子鏡、車內駕駛員和駕駛艙監控攝像頭以及具停車輔助功能的全景式監控系統(AVM)。該新型片上系統在車道保持、自動緊急制動(AEB)、智能大燈控制和速度輔助功能方面的應用，超出了歐洲新車評測系統(NCAP)的性能要求。此外，其採用先進的10納米工藝技術打造，功耗低，支持擋風玻璃前置ADAS攝像頭在尺寸和散熱方面的要求。ui
安霸半導體首席執行官Fermi Wang表示：「到目前爲止，因外形尺寸形成的功耗限制，致使前置ADAS攝像頭的性能受到限制。CV22AQ提供業界領先的卓越神經網絡，並且功耗低於2.5瓦，且具有一流的圖像處理能力，可以讓一級供應商和OEM 客戶大大提升ADAS算法的性能和準確性。」編碼
CV22AQ的CVflow架構以每秒30幀的速度提供800萬像素分辨率的計算機視覺處理能力，可精確識別遠距離物體。此外，CV22AQ支持多視場攝像頭的多個圖像傳感器輸入，還可以使用單個高分辨率圖像傳感器建立多個數字視場，以下降系統成本;其支持DNNs用於物體檢測、分類(即行人、車輛、交通標誌、交通訊號燈)、跟蹤以及自由空間檢測等應用。
CV22AQ的高性能圖像信號處理器(ISP)在低光條件下，能夠提供出色的成像。而高動態範圍成像(HDR)處理可從高對比度環境中最大限度地提取圖像的細節，從而進一步提升了該芯片的計算機視覺能力。此外，該芯片包含AVC和HEVC視頻格式的800萬像素編碼，容許客戶爲其汽車攝像頭添加視頻錄製和流媒體功能。