In this paper we study the problem of local motion analysis and apply it to swimming style recognition in broadcast sports video. Local motion analysis is challenging for two reasons: 1) local motion is usually buried in clutters involving complex motion from multiple objects; and 2) the process is more sensitive to noises compared to the recovery of global motion. However, an effective approach to local motion analysis is significant for understanding human activity from image sequences. In this work, we firstly extract the object-induced local motion by utilizing robust motion estimation and salient color. The object motion is accordingly characterized by compensated motion vectors and confidence measurement. Beyond a single image, we attempt to capture the motion periodicity over the local motion sequence. For each period, we locate a so-called salient frame within which we derive a compact representation to distinctly characterize an image sequence with repeated actions. Finally, we employ a hierarchical classifier to distinguish local motion based on periodicity and salient frames. Promising results have been achieved on swimming style recognition in broadcast sports video.