In-sensor computing paradigms for neuromorphic motion vision

Motion perception is fundamental to autonomous systems and requires the extraction of position, direction, velocity, and acceleration from dynamic visual scenes. Conventional frame-based sensors, silicon event cameras, and position-sensitive detectors provide established solutions, but data transfer and peripheral processing can impose substantial costs in resource-constrained applications. In-sensor computing offers a complementary approach by embedding selected encoding and processing functions in the physical dynamics of optoelectronic devices. This review examines general-purpose motion encoding and task-specific motion processing. Within general-purpose encoding, optoelectronic event motion encoding and fading-memory motion encoding address data redundancy through distinct mechanisms. The former mimics retinal ON/OFF signaling. Programmable bipolar photoresponses provide signals of opposite polarity, while temporal dynamics enable the generation of transient events suitable for processing by spiking neural networks. The latter uses fading-memory devices to retain recent optical inputs as decaying electrical states. When assembled into arrays, these devices act as physical reservoir units that map spatiotemporal inputs into high-dimensional states, from which a trained readout can extract motion information such as trajectories and velocity. For task-specific applications, specialized devices generate collision-sensitive and direction-selective responses at the sensory front end, potentially reducing downstream processing requirements. Finally, we discuss challenges in array uniformity, adaptive operation, peripheral integration, and consistent performance evaluation.