3D Target Recognition Of Mobile Robots Based On Deep Learning

Target recognition has broad application prospects in industrial production,detection security and daily life,and has become one of the important research directions in robotics,unmanned driving and other fields.The rapid development of artificial intelligence technologies such as mobile robot technology,machine vision,deep learning,and supercomputing,as well as the continuous improvement of sensor technologies such as industrial cameras and lidar,have laid a good foundation for target recognition.This thesis is based on the deep learning method to study the target recognition of 3D objects in the environment by mobile robots.First,the use of Basler industrial cameras and Jackal UGV mobile robots completed the parallel optical axis binocular vision mobile robot platform.Through the binocular vision system model,the working principle of the binocular camera to obtain the target depth information is analyzed,the binocular vision mobile robot platform is completed.An automatic adjustment bracket that can precisely adjust the baseline distance of the two cameras is designed.A binocular synchronous image acquisition program based on multithreading technology is designed.The binocular vision mobile robot is used to collect image pairs,and 3D reconstruction is performed to obtain the original point cloud data of environmental objects.Then,use 3D modeling software and multi-line lidar to assist binocular vision to obtain point cloud data;perform centralization and normalization of the obtained point cloud data;use a combination of multiple filtering algorithms to reduce the noise of the data;use batch automatic rotation and add Gaussian noise to expand and enhance the point cloud data.Through this part of the research,a data gradual growth mechanism is formed,heterogeneous and heterogeneous data are continuously integrated,and the data set is continuously enriched,which is convenient for training the network,and improves the generalization ability and recognition accuracy.Finally,according to the characteristics of point cloud data,the deep learning network architecture using global features and the deep learning network architecture using local scale features are designed.The network can directly input point cloud data.Compared with other methods,it reduces the amount of data calculation and the time complexity of calculation.Deep learning networks based on global features and deep learning networks based on local scale features use progressively increasing data sets for training and testing,and the acquired point cloud data is processed and input into the trained network for recognition.The experimental results show that the classification accuracy of the network using global features,single scale features and multi-scale features in the training set is 98.8%,96.8% and 96.8% respectively,and the classification accuracy in the test set is 93.6%,87.6% and 86.8% respectively.Using the obtained data for recognition can better identify the categories and achieve the expected effect.