Research On The Surrounding Environment Vision Perception Methods Of Intelligent Robot Based On Machine Learning

Surrounding environment vision perception is the fundamental method that can enable intelligent robots automatically understand unknown environments and flexibly manipulate complex target objects.Recognition of unclassified surrounding objects and spatial perception of target objects are two essential problems for the surrounding environment vision perception of intelligent robots.3D robot vision techniques and machine learning methods bring promising solution for the robot perception problems.The dissertation focuses on the topic that how to utilize machine learning techniques to boost the performance of the vision perception of intelligent robots.Following issues are summarized as the bottle-necks of the robot perception.Firstly,since current researches lack the analysis about the factors that can affect the learning ability of 3D-data-oriented machine learning models,A mathematical model is built to analyze the information obtaining process and factors that affect 3D data machine learning method.The model is applied to design machine learning methods of the 3D vision-based robot perception.Secondly,the performance of the state-of-the-art 3D shape-based object recognition methods cannot match the requirements of the robot perception.Machine learning based category information obtaining methods are investigated and object recognition models are designed to improve the speed and accuracy of the robot perception.Thirdly,the widely-used approaches are shorting on the ability of obtaining object spatial information from 3D shapes and unable to fit the criterion of robot object manipulation tasks.Machine learning based object spatial information obtaining methods are researched and spatial perception models are constructed to evaluate the real-time and precise performance.The main contributions and innovations of the research are:(1).The virtual pattern assumption based 3D data machine learning process mathematical model.The virtual pattern assumption is proposed for analyzing the machine learning process and then the process is profiled via the random variable theory.The experiment and analysis show that the mathematical model can be used to guide the designing of 3D data machine learning models.(2).The convolution network based environmental object recognition methods.The anisotropic convolution model and the point cloud convolution model are proposed to improve the recognition speed and accuracy of robot perception.Experiments reveal that the proposed methods have promising results on task such as recognizing the object from40 candidate categories.(3).The feeding-forward neural networks for intelligent robot spatial perception of target objects.Unsupervised machine learning is utilized to improve the precise of the object rigid body pose estimation results.Besides,modern machine learning techniques boost the perception speed and the performance on multi-object spatial perception tasks.Experiments show that the proposed method can precisely recognize the object and calculate the object rigid body pose.(4).The robot hand-eye vision perception system and line structured light vision measurement system.The vision systems are used as experimental platforms to test and validate the surrounding object recognition method and the target object spatial perception method proposed in the dissertation.The research is applied to improve the robot hand-eye object detection method and the train wheel radius measurement method.