Accurate detection and tracking of targets in low-light and complex scenarios is essential for the development of intelligent machine vision. However, such capabilities are difficult to achieve using conventional static optoelectronic convolutional processing. Here we show that in-sensor dynamic computing can be used for accurate detection and robust tracking of dim targets. The approach uses multiple-terminal mixed-dimensional graphene–germanium heterostructure device arrays and relies on the dynamic correlation of adjacent optoelectronic devices in the array. The photoresponse of the devices can range from positive to negative depending on the drain–source voltage polarity and can be further tailored using the back-gate and top-gate voltage. The correlation characteristic of the device array can be used to selectively amplify small differences in light intensity and to accurately extract edge features of dim targets. We show that the approach can provide robust tracking of dim targets in complex environments.

弱光、复杂场景下目标的精确检测和跟踪对于智能机器视觉的发展至关重要。然而，使用传统的静态光电卷积处理很难实现这种能力。在这里，我们展示了传感器内动态计算可用于对昏暗目标进行精确检测和鲁棒跟踪。该方法使用多端混合维石墨烯-锗异质结构器件阵列，并依赖于阵列中相邻光电器件的动态相关性。根据漏极-源极电压极性，器件的光响应范围可以从正到负，并且可以使用背栅和顶栅电压进一步定制。利用器件阵列的相关特性，可以选择性地放大光强的微小差异，准确提取昏暗目标的边缘特征。我们证明该方法可以在复杂环境中提供对昏暗目标的稳健跟踪。