Research On Image Understanding Methods Of Outdoor Scene Based On Machine Vision

Outdoor scene perception and understanding are indispensable abilities for mobile robot to navigate and explore outdoor environment automatically. However, one main nature of outdoor scene is that it is unstructured, which is always random, diversified and complex. Moreover, the priori vision information of outdoor scene is quit often poor and object recognition technology is thus still immature. So in recent years research on unstructured environment understanding has become a hot research topic in machine vision. In this research for the purpose of navigating automatically, we get the priori information from the aerial image to recognize symbolic objects in unstructured environments. After achieving vision navigation, the robot then locates itself by binocular visual information, and then recognizes multiclass objects by conditional random field model.We design outdoor scene understanding prototype system based on HSP Electric remote control climb car. According to the requirements on visual-based outdoor navigation and environment exploration, we emphasize on three areas of research:the shadow-based building detection from aerial image, a3D graph model based on depth information and graph model, a multi-class objects recognition process based on conditional random fields.The main research and achievements of the thesis are as follows:Firstly, we review state-of-the-art research on scene understanding in robot vision, focusing on three key technologies particularly relevant to this paper. Noticing the shortcoming of current research on outdoor scene understanding and multi-class objects recognition, we formulate the whole plan of the thesis on building detection from aerial image and on an improved outdoor scene objects recognition methods.Secondly, for shadow-based building detection from aerial image, because the current research has weakness on location detection and boundary extraction, a novel building detection algorithm based on a simplified building-shadow model is proposed. After extracting the shadow of the building, we can quickly search for buildings site and building-shadow boundary by checking shooting time based on the building-shadow model. This process therefore can eliminate straight-line approaching and speed up boundary extraction. After getting the initial building site, we can get the final building areas by comparing the histogram of the seed and the surrounding area.Thirdly, state-of-the-art research on setting the number of segmentation and threshold of region merging are arbitrary. Therefore we propose a3D clustering graph model which combines depth information and graph model to optimize the segmentation number and the threshold of region merging. We then construct accuracy evaluation function of3D clustering graph model based on segmentation number and threshold of region merging. By obtaining maximum accuracy evaluation function, we get the optimal combination of the segmentation number and the clustering threshold. By editing a two-dimensional graphic model according to the three-dimensional depth information, we can finally get3D clustering graph model of the whole scene.Fourthly, for the purpose of understanding multi-class objects, the method on classifier-based object recognition is proposed. We analyze, extract, and merge the different features of the objects. We then propose a variable-length sample selection method to select the training data. Finally we design different classifiers to achieve preliminary recognition of the main objects.Fifthly, because low level feature ignores the relationship among objects, we propose the conditional random fields based on multi-classifier object preliminary recognition. To achieve better segmentation and recognition, our work elaborate on how to model the conditional random fields based on preliminary recognition, how to get single node function and pairwise node function, how to train and learn the model.Finally, we design a prototype system of outdoor scene understanding based on HSP Electric remote control climb car. The system consists of building detection from aerial image module and outdoor scene objects recognition module. The building detection from aerial image module can get priori global information by detecting buildings from aerial image. The outdoor scene objects recognition module can recognize objects by extracting depth information, by getting3D clustering graph model of the whole scene, and by building the conditional random field based on multi-classifier objects preliminary recognition.