Research Of Heuristic Unstructured Road Detection Algorithm Based On The Combination Of Feature And Model

The visual and auditory information cognitive computing is the interdisciplinaryof information science, life science and mathematical science. The development ofvisual and auditory information cognitive computing reflects the overall strength ofthe national information services and related industries. Unmanned vehicle is theintegration of basic theoretical research of visual and auditory informationprocessing and the interface of brain-computer system. And it reflects the overallstrength of the national visual and auditory information processing field. The roaddetection algorithm based on computer vision is one of the key technology of theintelligence navigation system of unmanned vehicle. The real-time capability,robustness to shadows, illumination variations and anti-interference ability tocomplex and changing environment of a road detection algorithm have a directimpact on unmanned vehicles. Thus, the road detection algorithm became the focusof experts and scholars.The road can be divided into the structural road and the unstructured road. Dueto the complexity of environment, the surface of road are more sensitive to weather,illumination variations, shadows, currently, the accurate and real-time unstructuredroad detection is still a challenging problem. In the study of existing road detectionalgorithm, the feature-based algorithms are robust in various environment, but it stillhas the defect that these algorithm is computational expensive. The model-basedalgorithms are better in real-time capability, but these methods are less robustagainst shadows, illumination variations. The purpose of this study is to improve therobustness, anti-interference ability and execution speed, to make up for theshortcomings of traditional algorithms. The main work of this paper is as follows:Based on existing unstructured road detection algorithms, to deal with theinability to changing environments, noise, shadows and illumination variations, wepresent an improved road segmentation method based on multi-layer neural network.The method learns the color feature from the samples of both the road area andoff-road area, and then use the feature to classify a new pixel. To suppress theinterference of noise, we used a block segment method, and designed a error fixmethod based on the membership probability. To deal with the shortcoming that the road boundary fitting results aresusceptible to noise, we propose a heuristic fitting method. The conditional densitypropagation algorithm is used to track and forecast the vanishing point of the road.The result of forecast are used in the fitting process. Such a method can achieve thepurpose that enhance the stability and robustness of the road detection algorithm.To meet the real-time requirements for unmanned vehicles, we proposed a GPUand CPU cooperative acceleration technique, such an implementation enhanced thereal-time capability of the algorithm significantly.In order to verify the robustness of the algorithm and the effectiveness of thestrategy proposed in this article, we combines subjective judgment and quantitativeanalysis together. Compare with traditional algorithm, the experiments show that theproposed method remains robust in complex and changing environments, and withstrong real-time capability to meet the requirements of the intelligence navigationsystems. This paper has a certain theoretical and practical value.