<p>In-sensor computing platforms eliminate redundant data transmission, providing an efficient paradigm for motion perception, fundamental to intelligent systems. However, as a core sensory element, existing neuromorphic devices with a single dominant physical mechanism limit the temporal response to a narrow and fixed timescale, thereby constraining adaptability to real-world motion dynamics spanning a broad range of speeds. Here we present a temporally programmable in-sensor motion perception system based on a neuromorphic phototransistor with a coupled ferroelectric–charge trapping mechanism. The design enables continuous tuning of photocarrier lifetime from milliseconds to seconds without external circuitry or bias. The system integrates 3,072 transistors with high uniformity and a dual-addressing architecture to decouple write and read operations, enabling position-decoupled readout with strong suppression of pixel-to-pixel crosstalk and array-level noise, and supports parallel detection of a wide range of motion speed (spanning ~0.34 to 320 km h⁻¹) in real-world traffic scenarios. These results establish a scalable hardware approach for adaptive, energy-efficient motion perception, advancing in-sensor computing for autonomous systems and dynamic scene analysis.</p>