Research On Scene Understanding Technology Of Indoor Service Robot Based On Deep Convolution Neural Networks

Scene understanding is an important research issue in the field of computer vision and artificial intelligence.It has important application values in various application scenarios,such as robot navigation,driverless driving,and environmental detection.Scene understanding includes object recognition,target detection,and finally the semantic segmentation task of the scene itself.For service robots,scene understanding is the core technology to achieve true intelligence.The service robot with the ability of scene understanding has the capability of object recognition,target detection and scene semantic segmentation.After the mobile base and high-precision mechanical arm are matched,it can further realize the advanced tasks of autonomous navigation,object delivery,and indoor security.The difficult point in the study of scene understanding is how to obtain the robust feature of the target object when the target object receives the influence of translation,rotation,illumination or distortion in the scene.The CNN(convolutional neural network),which is based on the human visual cognitive mechanism,can get the robust feature of the target object in the scene under the supervised learning method,which is based on a large number of manual labeled training samples.Convolution neural network is an upsurge of research because of its remarkable effect in computer vision.Aiming at the problem of scene understanding,Around the three major tasks of object recognition,object detection and semantic segmentation in scene understanding,we have done the following researches:Firstly,this paper presents an intermediate-layers-concatenated CNN model(ILC-CNN,intermediate layers connected-CNN).This model aims at the current convolutional neural network as the characteristic of the output of the last layer of the network,and fails to make full use of the shortcomings of the network middle layer: firstly combine the front,middle,and end convolution layers,and connect through deep connection;then through the pooling layers,full-connected layers and other operations are used to obtain the feature vectors of the description images;the training of the auxiliary classifiers ensures the validity of the middle layer features and enables the model to be successfully trained.Experimental results show that the model has a significant effect on image classification and recognition tasks,and its extracted features are more identifiable,with higher recognition accuracy than other models.Secondly,for the object detection task,in order to improve the accuracy and speed of object detection,this paper uses the CNN,RPN,and region proposal networks and Fast R-CNN detection framework to construct a target detection system based on the deep learning method.The system extracts features through the shared CNN;generates candidate regions through the RPN;and implements target detection through the Fast R-CNN framework to achieve end-to-end target detection.Experimental results show that compared with other target detection methods such as SVM detection method based on HOG feature descriptors,the target detection system has greatly improved both in detection accuracy and detection rate.Thirdly,for the semantic segmentation task,this paper is based on the idea of a full convolutional network: Through the full-context convolution of the current mainstream CNN model and adding the upsampling layer,the feature map is restored to the original size;according to the artificial pixels semantic segmentation of level annotations trains the task set,and through end-to-end semantic segmentation through supervised learning.Experimental results show that the FCN-VGG16 model in this paper can achieve the pixel-level prediction of objects in the scene and is an effective solution to semantic segmentation tasks.Finally,based on the self-developed indoor service robot mobile platform,this article carries out experiments on object recognition,target detection and semantic segmentation in scene understanding under real indoor scenarios.The results show that each algorithm in this paper can effectively complete the required requirements for each task.