Neuromorphic CMOS imager for sparse vision data acquisition

The recent interest in modelling the human retina opens the doors to neuromorphic imagers. Neuromorphic engineering succeeds in achieving a biomimetic retina by providing an electrical model as close as possible to neuron architectures involved in the vision process: the event-based dynamic vision sensor (DVS) is designed for low-data and low-power image sensing acquisition. Its particularity resides in asynchronous pixels responding only to relative changes in light intensity. These sensors show a wide dynamic range, low power consumption and good time resolution. The visual neuromorphic field is thus not only promising for robotics, but also for real-time tracking. Dynamic vision sensors seem suitable for detecting sparse data acquisition but raise one question: how to efficiently decrease the power consumption of an asynchronous pixel responding only to relative changes in light intensity? Inspired from a state-of-the-art image sensor, this study proposes a new DVS design in a mature 0.18 um CMOS technology to tackle this challenge. Three different figures of merit are targeted: the dynamic range (to be maximized), the pixel area (to be minimized) and the power consumption (to be minimized). Moreover, compared to the state-of-the-art DVS working at 1.8 V or above, the main constraint added to this study is a supply voltage of 0.75 V to be compatible with the CAMEL image sensor from UCL. Pixel simulations show a detection in light changes of 10% with 3% of contrast matching. Moreover, the reported dynamic range is 140 dB. Finally, this new design provides a decrease of static power consumption from more than one order of magnitude (from 690 nW to 20.54 nW), at the expense of an increase in pixel latency of 42 us.