A 2.97μm-Pitch Event-Based Vision Sensor with Shared Pixel Front-End Circuitry and Low-Noise Intensity Readout Mode
Atsumi Niwa, Futa Mochizuki, Raphael Berner, Takuya Maruyarma, T. Terano, Kenichi Takamiya, Yasutaka Kimura, Kyoji Mizoguchi, Takahiro Miyazaki, Shun Kaizu, Hirotsugu Takahashi, Atsushi Suzuki, Christian Brändli, Hayato Wakabayashi, Yusuke Oike
Abstract
Event-based vision sensors (EVS) detect temporal contrast changes with high dynamic range using pixel-parallel comparison with a relative threshold. To achieve this detection feature, an EVS pixel comprises a logarithmic amplifier, a sampling circuit of the initial voltage, at least one comparator, and in-pixel logic circuitry [1]. The complexity of pixel circuitry constrains the pixel size. A back-illuminated (BI) stacked EVS reduces the pixel size to around <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$5\mu \mathsf{m}[2,3]$</tex> , but many applications looking to take advantage of EVS require smaller pixel sizes and higher resolutions. To meet speed requirements in terms of event detection, an asynchronous frame-free readout based on an arbiter circuit is often applied to EVS. Though asynchronous readout achieves response times of several <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$10\mu \mathbf{s}$</tex> in low-activity scenes, the unpredictable arbitration time degrades time accuracy and affects system-level features such as recognition accuracy in high-activity scenes. There have been several efforts to increase event throughput with innovative circuit techniques [4], [5], but the additional processing increases the post-processing cost to reconstruct a frame from asynchronous event data. A frame format synchronized with a stable time accuracy is suitable for recent post-processing approaches performed using a convolutional neural network. Moreover, conventional image acquisition with very low noise is still helpful for the complex tasks combined with event-based vision sensing. Recent proposals combining event detection and intensity acquisition have serious limitations due to random noise or propagation delays from arbitration [6], [7].