Research On Object Classification And Scene Localization Based On Machine Vision

Object classification and scene localization based on machine vision are challenging research hotspots in machine vision. The former is closely related to object detection, tracking and behavior analysis, and the latter can effectively help and improve the analysis and cognitive ability of robots in structured or unstructured indoor and outdoor environments. Classification of dynamic object in the scene is not easy to obtain robust classification characteristics, because of the object deformation and the interference of external factors, and has poor real-time performance. A few years ago, the bag-of-words(BOW) model from the text field was applied to classify images, called bag-of-visual-words (BOVW) model. BOVW model has a good performance in describing scene image, it is suitable for robot localization in large-scale outdoor scene. However, there are shortages on big data processing. In response to the above problems, the main research contents in the thesis are as follows:Firstly, a approach for moving object real-time classification based on contour features is presented. Vector model combined by a variety of contour features with good discriminant ability is proposed to describe the moving object, in which, these features are robust for translation and rotation of moving object. Firstly, Gaussian mixture model is used to motion detection. More accurate moving object contour image is achieved after image morphology processing, and then contour features extracted by feature vector model are applied as the basis of classifier learning and classifying. Final classification results are obtained. Experiment results demonstrated that this approach not only has higher classification accuracy, better real-time performance, but also improves the analysis and cognitive ability on moving object of robots, especially for its navigation, feasible region detection, motion priority criterion and so on.Secondly, mainly for the limitation of GPS sensors in some cases, an image matching method based on the bag of visual words is presented, and applied to visual self-localization of mobile body. Double-layers indexing structure and voting algorithm are introduced in this method. "From coarse to fine" double-layers indexing structure is established based on the group image set and reduces the searching space of image matching, the memory load and the computational complexity sharply, it is not easily affected by foreground change when localization and has good localization accuracy for complex time-varying environment of campus. The method only relies on monocular vision to locate. It makes up for the deficiency of the other sensor localization capability, providing a good support for visual navigation and motion planning of mobile robot.Finally, an outdoor object localization method based on multiple views of feature points matching is produced to make up for the deficiency of single view template matching. SURF is employed to describe and match the character, and this character is appropriate for diversity and complexity of objects in large-scale outdoor scene. Then, FLANN and ANN algorithm are applied to filter error matching points, which effectively improve the quality of matching and save the computation time. Meanwhile, multiple views matching is more in line with the application scenarios of angle changing when mobile robot wanders in outdoor environment. The experimental results verify the effectiveness of the method.