Proposes and tests a theory of the temporal integration of local-motion signals, which postulates that signals from local-motion detectors are made coherent in space and time by a special purpose network. This coherence boosts signals of features moving along non-random trajectories over time. Hence, these signal features can be detected with an outlier-selection procedure. The theory has a Local, a Coherence and an Outlier stage. Simulations with a simple neural-network implementation of the theory show that detection is impaired with increasing eccentricity, an effect that varies inversely with step size; and that detection improves over durations extending to at least 600 msec. The theory can account qualitatively for reported experimental features related to the detection of a signal dot, against the background of noise dots in Brownian motion. Motion-Energy and Elaborated-Reichardt models of local-motion detection are discussed.