韩国未上市自动驾驶感知软件公司 StradVision 称 MultiVision（L3/L4）首阶段明年完成、合成数据占比将由 5% 提至 30%，并把与 AWS 合建的道路数据平台单独做成 SaaS

StradVision, an artificial intelligence-based automotive object recognition software developer, will complete the first phase of development for MultiVision, its system aimed at Level 3 and 4 autonomous driving, by next year. The company will also turn its road data processing platform, built in collaboration with Amazon Web Services, into a separate software-as-a-service product.
Two passenger cars stood out upon entering the autonomous driving test center at StradVision's Dongtan office in Hwaseong, Gyeonggi Province, on the 8th — one fitted with a camera facing forward, the other with cameras front, rear and on both sides. They run FrontVision, which uses a single camera to detect vehicles ahead, lane markings, road edges and traffic signs, and SurroundVision, which uses four cameras to monitor blind spots around the vehicle.
Inside a demonstration vehicle, a monitor displayed live road conditions along with nearby pedestrians and obstacles, each marked with a distinct color and tag. With SurroundVision, parking was carried out automatically, with no need to touch the steering wheel, gear selector or brake. FrontVision has already been commercialized in overseas markets including India and China, while SurroundVision has secured global mass-production orders and remains under development.
ADAS Adoption Spreads Worldwide, Advancing Camera-Based Autonomous Driving
StradVision plans to integrate and upgrade these camera-based recognition technologies to complete an initial version of MultiVision, which supports Level 3 and higher autonomous driving, starting next year.
"We are currently running joint development projects with multiple global customers," said Kwon Tae-san, StradVision's chief operating officer. "Our goal is to finish the first phase of development by the middle of next year at the earliest, then discuss mass-production projects with OEMs to reach Level 3 to 4 autonomous driving."
Behind the company's push to advance its technology is a wave of vehicle safety regulations spreading across countries. India will phase in mandatory advanced driver assistance system features, including lane departure warning systems and blind spot information systems, on new buses and freight trucks starting next year. The U.S. has also signaled that automatic emergency braking will be required on most new passenger cars and light vehicles in 2029, and is tightening performance standards for automotive perception software. The company's strategy is to use its road data processing platform to supply software that meets each country's safety standards quickly and capture global demand first.
From Cattle on the Road to Rear-End Collisions, Synthetic and Simulation Data Fill the Gaps
The foundation for rapidly developing software tailored to different road environments around the world is what StradVision calls its data flywheel. It links everything from screening and labeling real-world driving data to training, evaluating and improving AI models in a single automated pipeline.
Rare situations that are hard to capture on actual roads are supplemented with synthetic data. If footage of cattle appearing on an Indian road is needed, a three-dimensional asset of a cow is composited into real road video to create training data. Dangerous scenarios that cannot be deliberately reproduced, such as traffic accidents, are also generated as simulation data for training.
"We will raise the share of synthetic data from about 5% of all training data now to 30% by next year," Kim In-soo, head of StradVision's data innovation center, said that day. The company also plans to begin commercializing its in-house data pipeline, SV DataFlow, for external customers next year, expanding data preprocessing, conversion and labeling technology it has used internally as an autonomous driving software development tool into a separate SaaS business.
"What sets us apart is that from the earliest stage of developing autonomous driving technology, we built synthetic data based on real-world driving data from countries around the world, and went on to build a platform that can process it at scale to repeatedly evaluate and improve models," Kwon said. "We will update our software quickly while verifying safety, and secure more mass-production opportunities."
Large-Scale Data Processing on AWS Cloud: Traffic Signs From 32 Countries Learned in Two Months
The heavy computing load involved in building the data pipeline is handled through a hybrid infrastructure combining the company's own data center with AWS. Routine AI model training runs on its data center in Pohang, while AWS GPU resources are brought in for tasks such as generating large volumes of synthetic data in a short period.
Kwon cited the European Union's traffic sign dataset as a leading example. "By running simulations in parallel on AWS infrastructure, we built data for 1,642 types of traffic signs across 32 countries in two months," he said.
Cooperation with AWS is expected to deepen as AI models such as MultiVision grow in scale. "As AI models get bigger and require more computing resources for training, our use of AWS will grow as well," Kwon said.