Momenta, together with Texas Instruments (TI) -- a global semiconductor design and manufacturing company -- unveiled the latest front camera perception product deployed on TI’s latest Jacinto ™ TDA4VM SoC (System on Chip) on January 7th, 2020, at the 54th International Consumer Electronics Show (CES 2020), which can help automakers meet the safety requirements for new cars, including Euro NCAP 2022/2024.
Momenta is developing a new generation of ADAS solutions based on TI Jacinto TDA4VM SoCs to meet the ever-increasing consumer demands for driving comfort and safety. Momenta’s industry-leading front camera perception and high precision localization algorithms, combined with TI's Jacinto TDA4 processor for ADAS applications, and wide-angle high-resolution cameras, provides effective object detection at long distances and in complex scenarios.
Momenta’s deep learning based algorithms make full use of the DSP cores and accelerators on theTDA4VM SoC for neural network processing. Designed to achieve market leading computational and power efficiency – 10 Tera-Operations-per-Watt (TOPS/W) – Momenta’s algorithms on Jacinto TDA4x architecture offer an array of pre- and post-imaging processing that simplify the computational load for neural networks. When the TDA4 SoC architecture is combined with Momenta’s algorithms, customers can achieve faster, more accurate detection and recognition for front camera perception, at a fraction of the power previously used. For the driver, this means a more confident, more comfortable assisted driving experience than was possible before.
Momenta's deep learning-based front camera perception algorithms, have industry-leading object detection and classification capabilities, offering high-accuracy detection and recognition in a variety of difficult use cases:
· Accurately identify lane lines of various types and colors with the recognition curvature radius of at least 50 meters;
· In addition to different types of vehicles from different angles (including special-shaped vehicles in China), the algorithms are also capable of recognizing vehicles that are partially occluded, extremely close to the front of the car or suddenly appearing in view;
· To further improve public transportation safety and protect Vulnerable Road Users (VRU), the next generation of front camera perception algorithms is optimized for more intelligent VRU detection, which can accurately recognize and track pedestrians, various types of bicycles and electric scooters that are occluded or suddenly appearing.
Based on one chip and single monocular camera, Momenta can also achieve 10cm localization accuracy and low-cost map updates using low-cost GPS and IMU. The front camera perception algorithms automatically extract the semantic HD map information such as traffic signs, lane lines, pedestrian crossings, road markings and traffic lights. Multi-sensor fusion algorithms allow for trajectory calculation with high precision. Perception results and trajectory information are then uploaded to the cloud in real-time, where it is fused with similar data from other nearby vehicles to improve HD map accuracy and realize low-cost map updates.
The combination of high definition map and high precision localization provides precise time and space information for autonomous driving, enhances safety through redundancy, improves the user experience of ADAS, and helps customers create high-performance, low-cost, and competitive products.
With the advancement of autonomous driving toward mass adoption, requirements for reliability and safety have also been made significantly higher. Momenta proposed an important methodology: closed-loop automation, which is the automated iterative feedback loop between data and data-driven algorithms. Utilizing massive amount of road data and iterative data-driven algorithms, autonomous driving technologies can continuously be improved, and long-tail problems that come with mass utilization can also be addressed.
Momenta accumulates more than 2 billion kilometers of driving data every year, which, combined with deep learning-based algorithms and through iteration, improves vision perception and high precision localization. Momenta’s technologies are adaptable to different processing chips and cameras with varied resolutions and optical parameters. These technologies will be deployed and updated using “crowdsourcing” of data from fleet vehicles enabled by Momenta. This closed-loop automation system automatically collects valuable data for training, which greatly improves algorithm iteration speed and in turn, enables the system to adapt to varied road, weather, and lighting conditions, and cover long-tail corner cases.
“Momenta developed a unique methodology for accumulating traffic data in real-time and advancing their learning algorithms in automotive, which can make an impact in ADAS applications,” said Curt Moore, product line manager for TI Jacinto processors. “Pairing a deep learning algorithm like Momenta’s with our Jacinto TDA4VM SoCs gives our customers an opportunity to achieve more efficient performance in deep learning.”
Momenta CEO Cao Xudong said, "Safe and efficient front camera perception is essential for ADAS. With the powerful performance, complete software offering and delicately designed toolchain of Jacinto TDA4x processors, we can quickly deploy and iterate perception, high-precision mapping, localization, and map updating algorithms to achieve precise and reliable traffic detection and recognition in complex scenarios. We look forward to working with TI to create intelligent driving solutions that cover more scenarios in the future."
By combining deep learning technology with high-performance mass-production ready sensors and processors, Momenta will continue to improve the engineering and productization of core technologies ranging from vision perception, high-precision localization and mapping, and accelerate the development of the ADAS industry.