3D Environmental Object Perception Of Robot Based On Deep Learning

With the rapid development of visual sensing technology and artificial intelligence technology,environmental perception has become one of the important research directions in the field of robotics.The 3D modeling of the environment object in which the robot is located and its recognition and detection are the basis for the robot to perceive the objective world.Based on deep learning theory and algorithms,this paper researches the robot's three-dimensional environment object perception,and mainly completes the following:First,the principle of deep learning perception model and representative network structure are analyzed,and the feature extraction process of deep learning perception model is described.Choosing the right deep learning network for target recognition and detection of 3D environmental objects.Secondly,a binocular vision platform for robots was built.Analyzed the ranging principle of the binocular vision system,completed the hardware selection of the camera and lens of the binocular vision platform,and the design of the structure of the acquisition system.Binocular vision image acquisition software was developed based on multithread technology.Then,a deep learning algorithm is used to perform target recognition on threedimensional environmental objects.Preprocess the 3D environmental object raw data collected by the robot's binocular vision platform to complete the point cloudification,point cloud filtering,centralization and normalization of the 3D data of the environmental object.According to the input rules of deep learning networks,a symmetric function is used to deal with point cloud out-of-order input,and an input / feature transformation network is used to handle rotation and translation invariance.Design a deep learning algorithm for 3D environmental object target recognition directly using point cloud data,and perform experiments and result analysis.The experimental results show that the target recognition deep learning algorithm proposed in this chapter can classify large-scale complex 3D point cloud data on the ModlNet40 benchmark with a classification accuracy rate of 89.7%.The algorithm can still accurately identify objects with a small number of point clouds.Finally,the deep residual network is used to complete the target detection of 3D environmental objects.The convolutional neural network is extracted by using the deep residual network with jump connections as the feature of environment objects.Jump connections effectively solve the problem of gradient disappearance and gradient explosion in the back-propagation process,and accelerate the convergence speed of deep learning networks;The multi-level feature pyramid is used to scale the features of different levels of objects,and then information fusion is performed to extract more lowlevel feature information;The region classification network is used to achieve accurate regression of region classification and location.Compared with mainstream algorithms such as Faster R-CNN,the average detection accuracy of the deep learning algorithm proposed in this chapter for deep object detection using deep residual networks has been improved by 2.3%,5.35%,4.60%,and 9.74%,respectively,and this algorithm has faster detection speed.