3D object detection for reconstructed scene using multi-layer growing neural gas

Abstract

In this paper, we tackle problems of indoor dynamic reconstructed scene created using multiple static Kinect sensors; toward this goal, we propose a dynamic object detection algorithm based on multi-layer growing neural gas for reconstructed scenes and creating a dynamic hierarchical structured space; in fact the proposed technique creates a multi-layer structure for representing a point cloud as points, fragments, objects/groups, and environment. Moreover the proposed algorithm uses statistical outlier removal technique and a downsampling algorithm based on growing neural gas in order to remove edge and shadow noises being very common in reconstructed scene created using multiple static Kinect sensors. With the proposed algorithm time complexity of object recognition, object tracking algorithms can be decreased. Experimental results demonstrate that the proposed algorithm achieves substantial improvement over the state-of-the-art.