A Bio-Inspired Visual Sensor With UV-to-NIR Fusion for Enhanced Recognition in Autonomous Driving Systems

Conventional vision systems for autonomous driving are hindered by energy-inefficient and latency-prone architectures due to physically segregated sensing, memory, and processing modules. Here, we report a bio-inspired vision sensor that emulates the spectral adaptation mechanism of the Pacific salmon retina leveraging a van der Waals heterojunction of NbNiTe₅ and black phosphorus (BP). Operating at zero bias, the device exhibits intrinsic wavelength-dependent antagonistic photoresponses: negative photoconductance under 365 nm ultraviolet illumination (mimicking visual suppression in bright environments) and positive photoconductance under 820 nm near-infrared light (emulating visual enhancement in dim conditions). This built-in adaptability enables robust environmental perception across extreme illumination scenarios, from high-glare daylight to low-light nights. Furthermore, we integrated this sensor into a full functional system that unifies image perception, non-volatile storage, and in-sensor processing. When deployed for traffic scenario analysis, a convolutional neural network trained on features extracted by the sensor achieved 96% classification accuracy, with performance scaling proportionally to the system's noise suppression capability. Our work establishes a practical strategy for high-contrast bidirectional photonic synapses and highlights the potential of biomimetic systems in neuromorphic vision - particularly for autonomous driving and intelligent surveillance, where reliable operation across diverse spectral environments is critical.