Previous work in the theory of laser radar operation has addressed the problems of detection, range and Doppler estimation for multipixel speckle targets. In this paper, the theory of automatic tracking for such targets is developed. A track-while-image approach is adopted in which the laser radar produces a sequence of raster scans, called frames, across the region containing the target. From the present frame, an estimate of the current target location is generated and fed to a tracking filter, which both combines this information with the evolving sequence of location estimates, and determines the radar's line of sight for the next frame. Within this structure, various image-centroid, hot-spot, and template-matching trackers for multipixel speckle targets are described. These estimators all exhibit signal-dependent noise, i.e., their covariances consist of a constant term plus one or more terms that are outer-quadratic in target displacement from frame center. Nevertheless, they have been successfully incorporated into a generalization of the Kalman-linear-least-squares tracking-filter that can use pre-computed gains. Specific system examples are presented to illustrate the transient and steady-state behavior of several of these trackers.