HOPBAS10K: A 98 × 98 Pixels CMOS Vision Sensor for Background Subtraction

Background subtraction (BS) is one of the first visual tasks in many video processing applications. In this work, we introduce a hardware adaptation of a top-ranked rule-based algorithm, the pixel-based adaptive segmenter (PBAS), implemented on an integrated circuit with on-focal plane processing. On average, our hardware-oriented PBAS (HO-PBAS) proposal features a similar algorithm performance to that of the original PBAS with the benefit of a reduced number of samples of the background model and linear equations, and thus a simpler overall model. This algorithm was implemented as a
${98} \times {98}$
pixels full-custom mixed-signal 180-nm standard CMOS vision sensor. This solution features in-pixel processing with resource sharing strategies and 47-
$\mu \text{m}$
pixel pitch. The in-pixel processing includes the whole algorithm data path along with image preprocessing. The assessment of our implementation through
${F}$
-measure metrics with images captured by our chip from the public dataset changedetection shown on a PC screen results in a decrease of performance of only 6.7% with respect to the software version of PBAS.