Outdoor Scene Dynamic Feature Point Filtering in SLAM Localization

The extraction and matching of the feature in V- SLAM is important to ensure the accuracy of the location. This paper presents a dynamic feature point filtering algorithm which combines the semantic segmentation and the geometric constraints. The algorithm performs the semantic segmenta- tion on the preprocessed RGB images after highlight/shadow removal to initially obtain the masks of the suspected dynamic objects. Then the motion consistency detection is integrated to determine the motion states of the feature points, preserving the static features while filtering the dynamic ones, while the dealt features are subsequently used for camera motion matrix estimation. Experimental have been performed to validated on the public dataset, i.e. TUM, KITTI and custom-built dataset to demonstrate the effectiveness in V-SLAM systems.