Automotive

To run more than a few dozen AI models — facial landmarks, detection of worn items, gaze detection and more — on an inexpensive CPU, for the driver monitoring system in Toyota’s advanced driver assistance system “Advanced Drive.”

The request came when the hardware and the Arm-core processor had already been decided. With the processor fixed, the only remaining option was to make the models faster on the side being deployed. Idein contributed optimizing compiler technology that accelerates deep learning inference, along with research and development on lightweight models, and had it running in three to four months. That collaboration began in 2017 and has since extended to a low-speed autonomous-driving path planner and multimodal agents.

Looking ahead to the arrival of the smart city, AISIN wanted to enter the AI camera field and bring a capable, extensible product to market early.

The Raspberry Pi general-purpose computer was paired with the “Hailo-8” AI accelerator chip, and Idein wrote the software that draws out its performance. Actcast then took on the remote management needed for large-scale operation, which meant AISIN did not have to build device control software from scratch. That division of labour is what made launch six months from the start of development possible.

To make an AI agent that converses naturally with visitors on an exhibition floor work without a staff member standing by throughout.

In a noisy hall, speech recognition slips easily, and a mishearing is enough to derail the conversation. Generative AI corrects those misrecognitions automatically from context, which prevents the breakdown. Alongside this, an AI camera judges on the spot when one visitor has been replaced by the next, and resets the conversation. Together these two let the service keep running with no staff intervention.