A $64\times64$ Pixel Vision Sensor for Local Binary Pattern Computation

We report on a 64 × 64 pixel vision sensor embedding local binary pattern (LBP) computation implemented at pixel-level. The proposed technique is applied on four neighbours within a 3 × 3 pixel kernel, while the operations are executed on-the-fly, during the sensor exposure time. Thanks to the pixel-level autoexposure control, processing over extended dynamic range is achieved. The number of transistors per pixel has been minimized in order to preserve the fill factor while keeping a reduced pixel size. The resulting binary information is stored on a dynamic memory and read out in a standard raster-scan mode. The proposed technique for the LBP estimation is made in the time domain and uses two voltage comparators per pixel. While the chronological order of comparators toggling identifies the brighter pixel, the voltage difference between the two thresholds sets the LBP sensitivity. The sensor is implemented in a CMOS 0.35 μm, featuring a 17-μm pitch with 15% fill factor. The chip consumes 35 μW at 15 frames/s, powered at 3.3 V for the analog part and 1.5 V for the digital part.