Object-oriented semantic mapping for indoor environments based on instance segmentation with deep learning
To enhance the fidelity of the object-oriented semantic maps of domestic service robots,this paper proposed an improved semantic map building method.The method performed instance segmentation on the images with a deep learning method and instructed a geometric relationship-based clustering method to segment objects from the point clouds.The clustering results not only preserved the semantic information identified by the deep learning method,but also described geometric boundaries of objects more precisely.To solve the problem of the lack of reliability for object labels in existing approaches,this paper also proposed a semantic map optimization process.It merged instances which had close distances,high overlapping rates,and similar label scores.Meanwhile,it removed false objects.The experimental results show that the proposed method can improve the quality of the map built by the system in terms of the accuracy of the boundaries and the number of instances.
万方数据
维普
