Ambarella, Inc., a Santa Clara, California-based developer of high-resolution video processing and computer vision semiconductors, and Momenta, an autonomous driving technology company based in Suzhou, China, announced a collaborative HD mapping platform for autonomous vehicles.
The combined solution leverages Ambarella’s CV22AQ CVflow computer vision system-on-chip (SoC) and Momenta’s deep learning algorithms to provide HD map solutions, including mapping, localisation for autonomous vehicles, and map updates through crowdsourcing.
“HD map is essential to autonomous driving systems. Ambarella’s CV22AQ CVflow computing platform has made it easier for Momenta to deploy and upgrade our HD map software and algorithm on embedded systems. Our HD map can automatically build maps, perform localization, and update through crowdsourcing,” said Xudong Cao, CEO of Momenta. “We look forward to continuing our work with Ambarella to deliver a wide range of autonomous driving software.”
“We are pleased to partner with Momenta to provide a powerful and open HD map platform,” said Fermi Wang, CEO of Ambarella. “The CV22AQ’s performance and advanced image processing help enable the full potential of Momenta’s advanced AI algorithms.”
Momenta’s vision-based HD semantic mapping solution is highly scalable and production-ready. Through crowdsourcing, the solution can create a closed feedback loop of big data, AI, and HD map updates. Based on localisation, Momenta discovers changes in the map elements and provides frequent updates to the cloud.
The CV22AQ is manufactured in an advanced 10-nanometre process, providing the ultra-low power consumption required for the design of compact automotive systems. Its CVflow architecture delivers real-time processing of up to 8 megapixel resolution video at 30 frames per second (fps) for high-precision deep learning based object recognition.
The CV22AQ’s high-performance image signal processor (ISP) delivers superior image quality in low-light environments, while high dynamic range (HDR) processing extracts more image detail in high-contrast scenes — further enhancing the system’s computer vision capabilities.
Using CV22AQ, Momenta is able to use a single monocular camera input to generate two separate video outputs, one for vision sensing (perception of lanes, traffic signs, and other objects), and another for feature point extraction for self-localisation and mapping (SLAM) and optical flow algorithms.
Ambarella provides a complete set of tools to help customers easily port their neural networks to the CV22AQ SoC. Based on the tool chain, Momenta is able to quickly migrate deep learning perception models to embedded platforms and achieve an accurate output.