A 3.0$_\mu$W@5fps QQVGA self-controlled wake-up imager with on-chip motion detection, auto-exposure and object recognition

Arnaud Verdant (1) , William Guicquero (1, 2) , Nicolas Royer (1) , Guillaume Moritz (1) , Sébastien Martin (1) , Florent Lepin (1) , Sylvain Choisnet (1) , Fabrice Guellec (1) , Benoît Deschamps (3) , Sylvain Clerc (3) , Jérôme Chossat (3)

1 CEA-LETI - Commissariat à l'énergie atomique et aux énergies alternatives - Laboratoire d'Electronique et de Technologie de l'Information
2 DOPT - Département d'Optronique
3 ST-GRENOBLE - STMicroelectronics [Grenoble]

Analyzing image content usually comes at the expense of a power consumption incompatible with battery-powered systems. Aiming at proposing a solution to this problem, this paper presents an imager with full on-chip object recognition, consuming sub-10$\mu$W using standard 4T pixels in 90nm imaging CMOS technology, opening the path for both wakeup and high-quality imaging. It combines multi-modality event-of-interest detection with self-controlled capabilities, a key for low-power applications. It embeds a log-domain autoexposure algorithm to increase on-chip automation. The power consumption figures range from 3.0 to 5.7$\mu$W at 5fps for a QQVGA resolution while enabling background subtraction and single-scale object recognition. This typically shows a measured 94% accuracy for a face detection use case.

Domains: Micro and nanotechnologies/Microelectronics; Signal and Image processing

Arnaud Verdant, William Guicquero, Nicolas Royer, Guillaume Moritz, Sébastien Martin, et al.. A 3.0$_\mu$W@5fps QQVGA self-controlled wake-up imager with on-chip motion detection, auto-exposure and object recognition. 2020 Symposia on VLSI Technology and Circuits, Jun 2020, Honolulu, HI, United States. ⟨cea-02903508⟩