We propose a novel method to detect human poses in videos by concurrently optimizing body part matching and object segmentation. With a single exemplar image, the proposed method detects the poses of a specific human subject in long video sequences. Matching and segmentation support each other and therefore the simultaneous optimization enables more reliable results. However, efficient concurrent optimization is a great challenge due to its huge search space. We propose an efficient linear method that solves the problem. In this method, the optimal body part matching conforms to local appearances and a human body plan, and the body part configuration is consistent with the object foreground estimated by simultaneous superpixel labeling. Our experiments on a variety of videos show that the proposed method is efficient and more reliable than previous locally constrained approaches.