Scene Understanding Of Government Service Robot Based On Deep Learning

In recent years,the application robots of specific scenes such as intelligent service robots,public security patrol robots and pension service robots have developed rapidly.Scene understanding is a very important research topic in robot research.This paper takes the government service robot specific scene understanding algorithm as the research focus,and uses the deep learning method to study the algorithm modules such as target detection,image semantic segmentation and image depth estimation in the scene understanding algorithm to form an accurate description of the scene content.The main research contents include:(1)Aiming at the problem of fast and lightweight real-time target detection,a DSC-SSD algorithm based on deep separable convolution is proposed.The deep separable convolution layer is used in the network structure to redesign the model structure.Network model compression and lightweight applications.(2)Aiming at the problem of target detection and image semantic segmentation,this paper proposes a joint target detection and semantic segmentation scene understanding algorithm based on Faster R-CNN and DeepLab,designs the algorithm model,and creates a local dataset in Microsoft COCO format.Training model in Transfer learning,enabling simultaneous target detection and semantic segmentation in the scene.(3)Aiming at the problem of monocular image depth estimation in the scene,a single image is used for depth estimation,which may correspond to multiple real scenes,and there is no auxiliary information that can be used to restore the depth of the scene.A two-dimensional image based on associated multi-frame images is proposed.The depth estimation algorithm and the network model is designed.The first image used as the target frame,and the remaining images are used as reference frames to provide auxiliary information for the target frame depth estimation.The output of the network is the depth map corresponding to the target frame.The algorithm modules are merged to form the robot to accurately judge the content of the government service scene,and the judgment information is used in the actual government service robot system.The experimental results show that the algorithm can obtain better description of the scene content and meet the expected algorithm design requirements.