Visual Odometry (VO) is a critical upstream task in the field of mobile robotics. Over the past decades, VO has evolved into a parallel design centered on keyframes, where most existing VO frameworks feature a visual front-end that operates at the camera frame rate and a back-end that updates at the keyframe insertion frequency. In this paper, we propose Pipeline-VO, a novel high-speed stereo visual odometry system that introduces an optimized pipeline architecture into optical-flow-based VO methods. Unlike conventional approaches, Pipeline-VO enables the localization and mapping threads to operate in parallel at the same frequency. The localization thread employs an efficient incremental feature update strategy to maintain a rich and consistent set of feature points for pose estimation. Simultaneously, the mapping thread utilizes a bidirectional cyclic matching check to remove inconsistent observations, ensuring a high-quality feature set for accurate pose estimation. Experimental results on public datasets demonstrate that Pipeline-VO operates in real time at 350 FPS on a common laptop while maintaining localization accuracy comparable to or even surpassing that of traditional stereo visual odometry methods.