Indoor Scene Recognition For Mobile Robot Based On Embedded Vision

With the development and enhancement of robotic theory and technique,mobile robot is widely employed in our daily life.The intellectualization plays a critical role of the mobile robot as an assistant in personal life.Based on scene recognition and embedded technology,this thesis focuses on the research of indoor scene recognition to promote the automated location and navigation of mobile robot.The research can highly improve the indoor locomotivity and intellectualization of the mobile robot.After deep discussion of commonly used image features,this thesis proposes a more comprehensive feature extraction method,which combines the complementary property with respect to spatial scale of local features and global features.Firstly,we utilize Gist algorithm to detect global features and PHOG to extract local features,respectively.Then,multi-classification comparison experiment on various feature fusion has been conducted.Synthetic experiment and real experiment demonstrate the excellent performance of Gist+i:*PHOG method on feature fusion.This thesis adopts machine learning method to render the high accuracy of feature classification.The SVM method is employed in the feature classification because of its superb advantage in small sample,nonlinear and high dimensional feature classification.Synthetic experiment on the 1-a-1 feature classification method on the indoor scene multi-recognition has been carried.Experiment results indicate that the proposed method achieves high recognition rate up to 80%on indoor scene recognition.Considering the high computation cost and data scale of Gist and PHOG method on indoor scene recognition,this thesis proposes a dimension reduction method based on the feature selection.This thesis modify the F-score on the dimension unity method to reduce the error brought by the feature discrimination.Real experiment has been taken on to decide the parameter of dimension reduction.This thesis puts forward an embedded indoor scene recognition system based on the i.MX6Q multi-core ARM processor and transplants the scene feature extraction,feature dimension reduction and SVM code into C++.We design a friendly human-machine interface Qt GUI and take on indoor scene recognition experiment,and we use the multiple threads to accelerate the system.In this thesis,the research work is summarized and the prospect of further study is proposed.