The front ADAS smart camera products are currently in mass production inside the mainstream models of leading China car companies, including passenger car models that entered mass production in 2022.
The third generation of Neusoft Reach’s X-Cube 3.0 front view smart camera is based on Ambarella’s CV22 AI vision system-on-chip (SoC), which runs the camera’s L2 AI perception algorithm. The X-Cube 3.0 is currently being implemented in mass production, providing safe and efficient passenger car technologies and solutions for OEMs based on Ambarella’s CV22 platform.
Building on Neusoft’s nearly 20 years of experience in developing AI vision algorithms for autonomous driving, Neusoft Reach’s in-house neural network algorithms fully utilize the high AI performance and powerful image processing capabilities of Ambarella’s CV22 SoC to achieve the simultaneous deployment of multiple models, including detection, segmentation and semantic classification. Optimized training is carried out for the driving environment with localized characteristics in China, including structural road conditions and complex traffic scenarios, such as non-standard vehicles, pedestrians with abnormal posture, obscured pedestrians, fences, roadblocks and special lane lines. Neusoft Reach’s AI models achieved significant performance improvements through the unique unstructured-sparsity capability of Ambarella’s CVflow® AI architecture, along with the support of tools from their CVflow platform and local technical team. These improvements marked a breakthrough in optimization efficiency and enabled high algorithm accuracy, which provides users with a safer and more comfortable driving experience.
Ambarella’s CV22 SoC is based on its proprietary and highly efficient CVflow AI processing architecture, which enables the industry’s best AI performance per watt. Additionally, this chipset uses a 10nm, automotive-grade process technology to help achieve low power consumption. Having supported mass production for many passenger vehicles in recent years, Ambarella’s CVflow platform toolchain offers proven support for a wide variety of neural networks. Additionally, the platform optimizes model importing, including quantization, sparsification, compilation and operation. In particular, it supports unstructured sparsity, as well as retraining to compensate for quantization accuracy loss, which provides advantages such as high execution efficiency and low memory bandwidth usage when running the same network.
Ambarella and Neusoft Reach plan to continue their cooperation in China, using Ambarella’s next-generation SoCs to continue creating intelligent driving systems with distinct technology advantages. The companies also plan to continue their joint development of L2 ADAS products for passenger cars in China, helping OEMs bring additional models into mass production that provide consumers with a safer and more comfortable driving experience.
