Mitsubishi Electric’s AI camera promises price/performance breakthrough for autonomous cars

Hidetoshi Mishima (left) and his engineers (c) Bertel Schmitt

Today in Tokyo, Mitsubishi Electric announced a possible breakthrough on the rocky road to autonomous driving: An AI-powered camera that more than triples the detection range from 30 meters to more than 100.

Using Mitsubishi’s “Maisart” AI technology (it stands for “Mitsubishi Electric’s AI creates the State-of-the-ART in technology,” we were told with a straight face) the system could solve a conundrum that continues to plague the autonomous drive business:

The car typically uses a camera with a wide-angle lens to take in as much information as possible. To detect earl, it would have to spot them from far away. A tele lens would create tunnel-vision.

Mitsu’s system mimics human behavior, we were told today at the Tokyo HQ of the sprawling concern. The system can rapidly focus on the appropriate regions within the field of view, such as cars ahead, or behind. Even from far-away, the system can distinguish between cars, pedestrians, or two-wheelers.

Happy with frugal hardware: The gearhead senses Renesas inside

What’s more (or rather, less) the system is happy to live on rather frugal hardware. Today, the system was demoed running off a computer the size of two credit cards. Mitsu’s General Manager Hidetoshi Mishima refused to disclose the type of computer used, but if I would have to bet, I would put my chips on a Renesas R-Car M3, an ARM Cortex-based SoC (system-on-chip) development kit that has become popular in the Automotive Grade Linux space for its low cost and ample from LAN to CAN connectivity. With its 1.7 GHz 4core chip, the computerlet won’t break any speed records, which makes Mitsu’s achievement even more respectable. Power consumption is another problem plaguing the field. High-wattage computers drain precious range from the battery. Smart algorithms require lighter-weight, and less costly hardware. A similar ARM CPU is in the ultra-low-cost Raspberry Pi3, and that one retails for $35.

This could solve another big problem: The high cost of sensors. LIDAR radar sensors did cost around $10,000 a piece until recently, when Velodyne famously cut the price to around $4,000 a piece.

Nevertheless, Mitsu doesn’t plan on disrupting the fledgling LIDAR market just yet. “Our system is great at longer-range object detection and classification, but not so great at distance measurement,” Mishima told me. “LIDAR and millimeter RADAR are great for distance measurement, but fail at object detection.”

No ship date for the new system has been announced, but it does not seem to be too far away.