Multi-Modula Object Recognition And Localization Technologies Aiming Intelligent Service Robot

Intelligent service robots,which have already made their debut in people's daily life,will play more important roles in the future.In their service,robots face various object recognition and object localization challenges,which demand cutting-edge re-search towards practical algorithms and technologies in robot vision.Aiming typical tasks of intelligent service robots,a convolutional neural network based object classifier and a multiple stereo cameras based vision system are proposed.This work features the following two main innovations.Compared to object recognition competitions such as ImageNet,service robots need to train their object classifier from very limited training images.Conventional hand-crafted features based methods lack generalizability,while training deep neural networks on limited training data will result in severe over-fitting.To overcome such difficulties and apply deep learning into service robot systems,transfer learning,data augmentation,and learning rate scheduling methods are used to successfully train the neural network.The object classifier based on this network can run on service robots in real-time,and can generalize to unprecedented environments.Point clouds from stereo cameras are used to segment the images of objects from the background.Experiments illustrate that this classifier exceeds its predecessors in classification accuracy.Stereo cameras can collect more information than 2D cameras,which is potentially benefiting to service robots.Yet obtaining integrated and precise 3D models of given objects at relatively small scale has been proven to be a non-trivial job.Several pre-aligned multiple stereo cameras are used to observe the object at different viewpoints,and a complete 3D model can be constructed.Then the object's location and 3D features can be calculated.A experiment platform is built to test the efficiency of this method.Analysis of the results shows that the errors of stereo cameras are compensated in cer-tain degree.An arm is used to operate several types of objects according to calculated location and features.The success of such operation proves that this method can meet the requirements of service robots.