Research On Robot Environment Perception Based On Deep Learning

In the past few years,the rapid development of artificial intelligence has benefited from deep learning technology.especially in the field of computer vision.Combining geometry and images in computer vision to solve the problem of robot environment perception has proved to be a very promising solution.Intelligent robots need to understand the geometric and semantic characteristics of the surrounding scenes in order to make meaningful interactions with the surrounding environment,and also a prerequisite for the robot to perform purposeful actions in the environment.Environmental awareness technology is a key part of robotic intelligence.The environment perception of a robot is that it acquires surrounding environmental information through its own sensors and can understand the information in the surrounding environment.Deep learning technology makes it possible for robots to understand environmental information.The robot can restore the geometric space of the three-dimensional environment well by using Simultaneous Localization and Mapping(SLAM)technology.However,it cannot understand the specific object information in the environment and cannot judge the mutual logical relationship between the objects.Therefore,a method proposed in this paper applies the deep learning method to the traditional SLAM method to generate a three-dimensional map with semantics,and the method can effectively generate real-time three-dimensional semantic maps to realize the perception of the environment.The main research content includes:Firstly,the depth vision based SLAM system is described in detail,including depth camera model,coordinate system,depth camera parameters,visual SLAM basic composition framework,and finally the SLAM process is described from the mathematical point of view.Secondly,this paper deduces the forward propagation and back propagation algorithms of neural networks from the perspective of mathematics,and expounds the basic composition of convolutional neural networks.In order to ensure the real-time information of the environment,this paper needs to mark the RGB image of the key frame in real time.Therefore,this paper introduces the target detection algorithm based on R-CNN series and the fast target recognition algorithm based on regression prediction.After comparison,YOLOv3 is finally adopted.Algorithm and experimental verification in an office environment,the accuracy and real-time performance of the algorithm are verified.Finally,in this paper,the real-time 3D reconstruction of spatial information is performed by ORB-SLAM2 algorithm,and the algorithm is improved.The target detection algorithm YOLOv3 is merged,and the key frames in ORB-SLAM2 are semantically labeled using YOLOv3.Both of them are in ORB-SLAM2.On the three threads,another thread is opened,The Point Cloud Library(PCL)is used to process the key frames with annotations and their corresponding point clouds,and generate 3D semantic maps with semantic information to realize the perception of the environment and the verification is carried out on the TUM dataset,and finally the accuracy and real-time performance of the algorithm are verified on the robot.