Event-based object detection from spiking neural networks to IoT edge deployment: a systematic review

Event cameras output asynchronous per-pixel brightness changes instead of synchronized frames, enabling always-on, low-power perception for IoT applications such as smart surveillance, autonomous drones, and robotic perception. As research combining event sensors with spiking, conventional, and hybrid neural networks accelerates, gaining a structured understanding of the field’s trajectory becomes essential for guiding future work. This review synthesizes 59 detection studies (2019–2025) spanning three detector families (event-native ANN, pure SNN, and hybrid ANN–SNN) to provide that big-picture map. Progress has been rapid: event-native ANNs now exceed 52% COCO mAP on key automotive benchmarks, SNN inference pipelines have been compressed by orders of magnitude, and a neuromorphic drone detector operates at sub-10 mW on-chip power (idle measured, dynamic modeled) under sparse-detection duty cycling. However, evaluation practices differ widely in benchmarking protocols, metric definitions, and energy reporting: we find that a single evaluation choice—the inference frequency—can shift reported accuracy by up to 42.6 mAP points, far exceeding the few-point spread that separates state-of-the-art architectures, so protocol differences can dominate apparent architectural ones when interpreting published results. To support future research and IoT practitioners, this review offers a unified technical primer on event-camera detection, organized benchmark synthesis tables with protocol annotations, an energy evidence inventory distinguishing hardware measurements from analytical estimates, and an IoT deployment analysis addressing communication, memory, and latency constraints. A research agenda identifies the concrete steps needed to advance event-camera detection from laboratory benchmarks toward field-deployable IoT systems.