A Low-Power VGA Vision Sensor with Event Detection through Motion Computation based on Pixel-Wise Double-Threshold Background Subtraction and Local Binary Pattern Coding

Abstract

This paper presents a micropower VGA event-based vision sensor for IoT applications, achieving a low rate of false positives also in noisy outdoor scenarios. The embedded programmable pixel-wise, double-threshold, motion detection algorithm, applied on a under-sampled array of 160 × 120 pixels, allows to address a wide spectrum of applications. The sensor typically works in detection mode, consuming 344µW at 8fps, analysing the image, suppressing noise in the scene and looking for moving targets to generate an alert which triggers a processor for high-level vision tasks. After the alert is asserted, the sensor switches to the imaging mode delivering full resolution gray-scale images together with their motion bitmaps at 1.35mW. The 4µm pixel sensor is fabricated in a 110nm CMOS and occupies an area of 25 mm2.