Study On Deep-learning Based Object Recognition And Multi-robot Formation Application

With the development of computer vision and image processing, object recognition is widely used in industrial, military, aerospace and some other fields. It has been a hot topic of research in the fields of automatic control, computer vision and pattern recognition.Feature extraction of object has two major methods, artificial feature extraction and automatic feature extraction. The traditional neural network is one of automatic feature extraction methods, but it is easy to be over-fit because of the global connection. The Deep Learning methods used in this topic is a kind of hierarchically automatic feature extraction method, it can extract the features from layer to layer and the higher layers of features are composed of lower layer of features. In this thesis, a large number of object samples are obtained firstly by using objects detection, objects segmentation and some basic methods of image processing. Then Convolutional Neural Network(CNN) is used to train a network which is used to recognize the objects. By contrast, CNN can reach a higher right recognition rate to 95.9%, while the traditional three layers of neural network can get a rate of 93.4% only. Furthermore, the double Deep Belief Network(2-DBN) is proposed in the dissertation to solve the ambiguity problem of image recognition. The 2-DBN is a method of information integration. And we can find that the right recognition rate reaches to 97.3%, higher than that of CNN.Meanwhile, the existing algorithms in robots formation are not good enough to solve the problem of adaptive formation in the complex environments. This thesis combined object recognition using deep learning algorithm with the multi-robot formation to solve the problems of obstacle avoidance and formation switching. The robots could learn the environment and form certain formation or switch the formation automatically by analyzing the environment situations. The robots can learn what the obstacles are through a pre-learning network, and then they move to form the formation.Finally, to verify the proposed method for multi-robot formation, we build a platform with Visual Studio2010 and MFC. The size of the simulated environment is set as 600 cm by 400 cm, and obstacles with different shapes can be manually placed in the environment. In the simulation, 4 wheeled robots based on differential driven model are placed in the environments with simple, complex and experimental configurations are tested to verify the effectiveness of the proposed method in this thesis.