부산대학교 도서관

Relentless power reductions in always-on untethered imagers for distributed vision are required to fit the power budgets available from their tightly constrained energy sources. As an effective approach to reduce power, event detection has been explored to reduce system activity in uninteresting frames or regions [1–6]. In imagers with event detection, one of the main challenges is to simultaneously achieve substantial activity reduction for lower system power (i.e., low event detection false-positives) and low power consumption in the event-detection circuitry (generally higher when targeting lower false positives). Frame difference for motion detection is relatively simple and hence power-inexpensive, although it is well known to offer limited activity reduction due to background motion [1, 2]. Background subtraction is generally more effective in reducing activity, but it comes at the cost of higher power due to its higher complexity [3, 4]. Further opportunities to reduce activity are potentially added by object segmentation, which can suppress readout [5] and potentially restrict subsequent processing (e.g., neural network) to specific regions of interest (ROIs) rather than the entire frame. However, in this case, event detection inaccuracies severely limit potential activity reduction, thus requiring further mitigation of false positives [6]. Accordingly, morphological transformations such as erosion have been shown to remove isolated and noisy areas for lower false positives, although at high power consumption due to the digital implementation outside the pixel array [6]. From the above fundamental tradeoff, new solutions with more favorable false positives-power reduction tradeoff are necessary to simultaneously suppress activity/readout in uninteresting regions via object segmentation and accurate event detection, while keeping the latter energy competitive.