Volatile self-selective memristive neuron for millisecond-latency neuromorphic object detection at the edge

The deployment of artificial intelligence for real-time object detection in edge applications is constrained by the power and latency limitations of conventional computing architectures. Here we show a bio-inspired neuromorphic system built around a self-selective GaOx/ZnO memristor to address this challenge. The device exhibits a selection ratio and nonlinearity (both of ~10⁷), picoampere-level leakage currents, and microsecond-scale volatile dynamics. We integrate these memristors into a 32×32 array emulating the first-spike-time-coding mechanism of the frog visual system, enabling millisecond-scale pulse responses. When applied to aerial drone object detection, our hardware system achieves reliable recognition for pedestrians and vehicles, with only a 2.5% accuracy drop compared to software simulations. Furthermore, the array demonstrates a parallel processing scale of ~8.36×10¹² computational nodes under a 10% read margin. This work provides a tangible hardware solution for constructing fast-response neuromorphic computing systems at the edge, suitable for intelligent transportation and real-time monitoring. Real-time object detection in edge applications is constrained by power and latency. Zhang et al. report a self-selective memristor with high selection ratio and fast decay speed to realize high-performance artificial neurons. High recognition accuracy is validated for drone object detection using such a neuron array.