A map construction method based on semantic information in dynamic scenes

Abstract:
The map construction of static scenes can be well achieved by traditional visual odometry. The preconditions for such static scenes are not suitable for complex real-world environments with moving objects, which are scenes for robotics or autonomous driving. These moving objects can increase the outliers and drift of the constructed map. To solve the problem of increased outliers and larger drift when constructing maps in dynamic scenes, this paper proposes a map construction method that combines visual odometry technology to perform explicit outlier removal and drift correction with semantic information in dynamic scenes. With the aid of the pixel-level semantic information, this method can not only eliminate the outliers of the point cloud in the process of polar line search but also perform depth filtering to eliminate some outliers that do not conform to the scene scale. In addition, to correct the drift of the scene point cloud, this method constructs an accurate global map through loop closure detection in dynamic scenes. Experimental results show that our method can effectively reduce the outliers and drift of point cloud maps in dynamic scenes.