A 128 × 128 monolithic spike-based hybrid-vision sensor with 0.96 Geps and 117 kfps
Authors: Huanhui Zhang, Chi Zhang, Xu Yang, Zhe Wang, Cong Shi, Runjiang Dou, Shuangming Yu, Jian Liu, Nanjian Wu, Peng Feng, Liyuan Liu
The event-based vision sensor (EVS), which can generate efficient spiking data streams by exclusively detecting motion, exemplifies neuromorphic vision methodologies. Generally, its inherent lack of texture features limits effectiveness in complex vision processing tasks, necessitating supplementary visual information. However, to date, no event-based hybrid vision solution has been developed that preserves the characteristics of complete spike data streams to support synchronous computation architectures based on spiking neural network (SNN). In this paper, we present a novel spike-based sensor with digitized pixels, which integrates the event detection structure with the pulse frequency modulation (PFM) circuit. This design enables the simultaneous output of spiking data that encodes both temporal changes and texture information. Fabricated in 180 nm process, the proposed sensor achieves a resolution of 128 × 128, a maximum event rate of 960 Meps, a grayscale frame rate of 117.1 kfps, and a measured power consumption of 60.1 mW, which is suited for high-speed, low-latency, edge SNN-based vision computing systems.
