Landmark Segmentation and Selective Feature Extraction in Street-View Image

Visual Based Localization (VBL) can be used as an alternative to GPS-based localization system. When used in vehicle, VBL can be performed by acquiring street-view images to determine vehicle's present location. Landmarks in a city or area are collection of objects present in fix location and has persistent unique visual characteristics. By recognizing landmarks in street-view image, estimation of vehicle's location can be determined. However, numerous objects not relevant for location estimation can also present in street-view images. Therefore, a method to localize landmark and selectively extracts feature in street-view image is required for efficient VBL. This paper proposes a method for landmark segmentation and selective feature extraction in street-view image taken during daytime using simple implementation of semantic segmentation network. By comparing to the feature extraction without segmentation, the proposed method achieves more robust landmark feature extraction result to temporal and occluding objects in street-view images.