Outdoor Natural Scene Understanding For Vision-Based Mobile Robot

With the fast development of new technology in pattern recognition, image processing and software engineering, machine vision has become the focus in the field of robotics and artificial intelligence. Understanding of visual sensor information extracted from surrounding environment is an essential task for mobile robot autonomous navigation and environment exploration, especially when mobile robots have to accomplish environment identification and understanding in outdoor unstructured scenes. The intension of this paper is to introduce our research on mobile robot outdoor natural scene understanding based on vision system.In order to solve the image recognition and understanding problem in complex outdoor scenes, a stable and efficient image feature extraction algorithm is proposed for each sample in image library, which is constructed using a series of images extracted in outdoor environment. In our word, every image is convolved by the high-dimensional filter bank to obtain quantified image, and the basic structure of the image feature (Texton descriptor) is extracted by a clustering algorithm.Obstacles avoiding is very important for a mobile robot running in high speed mode. To avoid the false distance measurement with laser range finder, an obstacle detection method based on Texton global image histogram is introduced to describe the outdoor scenes in dynamic field environment. In addition, a global image histogram pyramid matching approach is also used to measure the similarity between two image scenes. Since there are many types of objects existed in complex unstructured environment, support vector machine (SVM) is adopted to build multi-categories identification models for multi-class classification, which can realize the real-time obstacle detection in dynamic outdoor environment.To meet the requirement of the autonomous navigation and environment exploring, autonomous system should have the ability of natural scene understanding and cognition. According to the classification and recognition result of Boost algorithm, the Texton-based semantic weak classifier and Texton-based pixel weak classifier are extracted, respectively. In our study, the corresponding strong classifier recognition model is constructed to achieve efficient image recognition. For the image sub-blocks with a large number of similar structural information, this paper adopts clustering algorithm to construct different categories of image sub-block models library. In identification stage, this paper uses model library matching algorithm to accomplish the image sub-block identification. In order to solve the problem of ambiguity of Texton descriptor, this paper presents the image recognition model based on semantics sub-block, integrates the global categories information and image sub-block categories information and improves average recognition rate of image sub-block markedly. Experiment results and further experiment data analysis show the model and method's validity and robustness.