Based On Surface Roads And Unstructured Visual Recognition

Vision-based navigation system is one of the hot research areas of the pattern recognition and artificial intelligence, and mainly used in autonomous navigation of mobile platforms, such as robots, intelligent vehicles, etc. The understanding of the road environment is one of the key technologies of the navigation system, including road direction and road position. Currently, the study of visual-based structured road is relatively mature. This is because the structured road has relatively fixed characteristics, such as traffic lanes, traffic signs, etc., which can help establish a unified model. While for the unstructured road in a complex environment, the study is still not enough. Due to the fact that the lighting, pavement and scene in the complex environment have not laws to follow, the study of it become more difficult. This paper mainly includes three parts:1) Using vanishing point detection algorithm based on texture to split the accessible area. According to the multi-scale and multi-direction characteristics of the Gabor filter, the orientation of the road texture is extracted. This paper presents a fast local adaptive soft voting (FLASV) algorithm to calculate the road vanishing point. Then the accessible area of the road is obtained based on the Orientation Consistency Ratio (OCR). The experiments and analysis show that the proposed algorithm is well adapted to the complex and changing environment, and achieves lower time complexity and higher accuracy.2) Research on the texture image color space representation, especially focused on the tensor discriminant color space (TDCS) model, and we use third-order tensor to describe the texture image, searching for the optimal color space. Also, we research some descriptors of local texture feature, including LBP, HOG and SIFT feature descriptor, and we do some experiments based on multi-feature fusion.3) Research on the classification method based on Multi-surface Proximal Support Vector Machine via Eigenvalues, especially focused on the Regularized General Eigenvalue Classification (ReGEC) and Multi-weight-vector Projection Support Vector Machine (MVSVM). Theoretically, some equivalent formations of MVSVM are given. The experiments indicate that MVSVM obtains better recognition performance than others. Finally, we conducted experiments to indentify the accessible area of unstructured road.