Abstract: To overcome the insufficiencies of varying illumination, large displacement estimation, and outlier removal, a motion-adaptive V1-MT (MAV1MT) motion estimation algorithm based on machine learning and a bio-inspired model of sequence image was proposed, starting from the theory of visual cognition. First, a structure-texture decomposition technique based on the Rudin Osher Fatemi (ROF) model was introduced to manage the variation in illumination and color. Then, a pooling stage at the MT level with non-normalization, which combines the afferent V1 responses using the adaptive weights trained by ridge regression, is modeled to obtain the local velocities. Finally, through introducing the coarse-to-fine method and pyramid structure subsampling of the local motion, the MAV1MT model is used on realistic video. Theoretical analysis and experimental results suggest the new algorithm, which is more fitting to information processing features of the human visual system, has universal, effective and robust motion perception performance.