Real-Time Mapping Using Stereoscopic Vision Optimization

摘要

This research focuses on efficient methods of generating 2D maps from stereo vision in real-time. Instead of attempting to locate edges between objects, we make the assumption that the representative surfaces of objects in a view provide enough information to generate a map while taking less time to locate during processing. Since all real-time vision processing endeavors are extremely computationally intensive, numerous optimization techniques are applied to allow for a real-time application: horizontal spike smoothing for post-disparity noise, masks to focus on close-proximity objects, melding for object synthesis, and rectangular fitting for object extraction under a planar assumption. Additionally, traditional image transformation mechanisms such as rotation, translation, and scaling are integrated. Results from our research are an encouraging 10Hz with no vision post processing and accuracy up to 11 feet. Finally, vision mapping results are compared to simultaneously collected sonar data in three unique experimental settings.