Research On Key Technology Of Deep Interactive Robot In The Man-machine Environment

With the development of science and technology,the robot's task has not been able to meet the requirements of human beings in the human operation,artificial intelligence has been the subject of most international technology conferences in recent years,robot with perceptual and cognitive abilities is the inevitable trend of the future development of the robot.In recent years the stereo vision sensor technology continues to improve,we can get the three-dimensional information with low price and high precision,at the same time,computer vision technology is maturer with the improvement of computer processing ability.Therefore,the related technology of intelligent robot with vision as the main sensor is the research focus of domestic and foreign scholars.This paper takes such robot with deep interactive ability as the research object,the most critical techniques were carefully studied.In order to achieve robot perception,cognitive ability,relevant technologies that need to be studied are:3D map building,human behavior recognition and object recognition.The paper realizes the perception of the environment from the improved visual SLAM system,and presents a method of human behavior representation and designs a deep learning framework to realize the cognition of human behavior.Finally,the recognition of specific objects in the environment is realized by an online matching object recognition method,then we can realize the cognition of the environment.The main research contents and conclusions are as follows:1.The algorithm is improved for the ORB-SLAM2 in many mature visual SLAM systems.When the robot in indoor running process,there will be the following:1.There may be some small obstacles on the ground,such as the wire robot can walk along it,as the robot goes by,the whole robot vibrates,the vibration will last a very small time and will soon get back.2.Because of the camera exposure time,during the operation the camera will collect some blurry frames,but these fuzzy frames are not large continuous.This is called a transient disturbance.Due to transient interference resulting in tracking loss,based on the original two algorithms this paper proposed two tracking algorithm,and redesigned the front tracking process of ORB-SLAM2.The improved algorithm makes the system avoid tracking failure,increases the tracking matching range,and tries not to carry out severe global relocation even if the tracking is lost.Experimental results show that the improved algorithm can improve the tracking loss caused by instantaneous interference.Finally,a dense point cloud map building module was designed for ORB-SLAM2.However,due to the large amount of information on the point cloud map,in order to realize the real-time performance of the map module,the map is calculated and displayed in batches,which reduces the refresh rate of the map.The map is built in the data set and the real environment,the map's accuracy is high and the point cloud density is adjustable,and the map module has the filtering function,the obtained map noise point is very few.2.A method of human behavior recognition based on deep learning is proposed.Firstly the paper proposes a kind of human behavior representation and produces a corresponding data set,the data set contains three actions:drink,light up a cigarette and pick up the phone.The entire dataset was completed by three people,two of them finished the training set,the third finished the test set.Secondly two kinds of data processing methods are designed,and three basic deep learning models are designed based on the representation of data sets.Finally,twelve experiments were designed.The data set in the three types of deep learning model has good performance,the accuracy rate of every experiment is almost above 95%,the usability of the data set is verified.Finally,the advantages and disadvantages of the two data processing methods are analyzed through data comparison.3.The paper presents an online matching object recognition,which replaces the semantic map building with real-time recognition.First of all,the result of using behavior recognition is to label the point cloud that is separated from the object interacting with the person,the point cloud with the tag is put into the point cloud database for recognition.Secondly,a three-dimensional point cloud matching process is designed based on ICP algorithm,the matching process first puts the object point cloud in the point cloud library into the same coordinate system with the matching point cloud,then uses the ICP algorithm to match the point cloud which to match to the point cloud in the point cloud database,the matching error can be obtained at the same time.Finally in order to verify the rationality of the method,a matching verification experiment was designed,a point cloud database of three objects is established,which contains two different cups and a mobile phone point clouds.At the experimental stage,the point cloud information of these three objects was collected from another angle again,and the results are analyzed by matching them with the object point cloud in the point cloud database.The result shows that the error of point cloud matching for the same object is much smaller than that of the point cloud matching which not belong to the same object,so the object can be recognized through the error comparison.