Research On Robot Scene Recognition Based On Depth Learning

In the field of robots,how to recognize the current scene where the robot in is a very important research topic in machine vision field when the robot reaches a inexperienced environment,but also a basic research problems in robot localization and navigation.Research on robot's scene recognition can help to obtain the real-time pose data in the working environment of the robot,and it is the key Step to establish the real-time map of the working environment of robot.Researchers hope that the robot can automatically recogniaze the class of the current scene by the previous experience,which will help the robot to complete the next task.The traditional scene recognition method relies on the most critical Steps of scene recognition to extract the characteristics of the scene image,but it takes a lot of time and effort to extract a good scene image feature,and the process requires researchers with heuristic experience.To solve this problem,A model of deep learning convolutional neural network is applied to research robot scene recognition in this paper,which can automatically find the features hidden inside the data by learning,this helps to reduce the workload of manual extration of images features.In order to make the robot have the ability to identify the working scene,a robot scene recognition system is established in this paper,and a convolution neural network structure with multiple convolution layers,pool layer and full connection layer is established for the system.The system uses a visual sensor to acquire and process the robot working scene image data set which required to train the convolution neural network structure.In the process of acquiring the robotic working scene image data set,the laser sensor is used to collect the distance data of the obstacle in the environment,and combined with the robot's pose system data to complete the robot self-localization and enviroment map building.In addition,this paper uses a set of experiments to determine the type of activation function,the sampling method of the pooling layer and the decline of the learning rate used in the convolution neural network model.According to the final results of these experiments,it is concluded that the network in this paper uses the ReLu activation function,the max pooling way in the pooling layer,the learning rate is reduced by the way of each generation and both the first two full connection layer with Dropout technology.In order to verify the effectiveness of the network structure model developed in this paper,using my own collection of data sets and the network data sets to train and test the network structure model of this paper,and do experiments to verify the effect of using this network structure to recognize different number of scene.The experimental results show that the test accuracy is high,when applied to the actual robot system we can shoot multiple times in robot woking scene at multiple angles and then use the trained network structure model to recognize the scene in turn,which can meet the actual needs of the robot scence recognition.