Study On Recognition Algorithms For UGVs On Unstructured Roads

Unmanned Ground Vehicle is a vehicle that can make decisions and adjust its moving states according to its environment perception automatically under outdoor environments. After making path planning, it then executes its actuators to move.Along with the need of modern war, the development of new material technology, new energy technology, information and automation technology, UGVs have been researched and developed widely in many developed countries.Environment perception is the key technology of UGV system and its efficiency and accuracy is the most important foundation for UGV's autonomous navigation. As an important component of UGV environment perception system, the research on unstructured roads recognition has a great role of importance and reality.Foreign groups mainly use different sensors such as LADAR, sonar, vibration sensor, color CCD, etc,in their research. They derive information through data fusion technology and build complicated environment models to get a comprehensive perception for UGVs. While domestic study mainly depends on visual sensors and image processing algorithms to recognize unstructured roads. Unstructured roads under outdoor environment have ill-structured situations. And it is impossible to represent them through some general models. It is likely to be admitted that road and non-road parts are different in aspect of color and texture features, so this paper has studied methods of unstructured roads recognition algorithms through above features.The main contents of this paper are introduced below in detail.1,A recognition method based on hue information of color images is proposed. RGB values extent in a wide range and they always have an influence on each other, so it is hard to determine a threshold for the recognition of objects in RGB space, even they have the same color attributes. HSV color model belongs to visual perception color spaces. It uses hue, saturation, value to represent color information, which is a more suitable foundation for recognition. Hue attribute has a stable and narrow numeral range and can be chosen as a fundamental recognition parameter. An image is transformed from RGB space into HSV space firstly. Then we used Otsu method to get a threshold to segment road parts from non-road parts with hue values.2,Based on the grey texture information of an image, the grey level co-occurrence matrix and the fractal dimension methods are studied. An unstructured road image can be regarded as a texture image including different texture parts. Texture analysis has four kinds: statistical analysis, structural analysis, model analysis and space/frequency analysis according to the extraction method of texture features. Grey level co-occurrence matrix describes comprehensive information about different direction, neighbor distance and grey distribution. It's the foundation of local model and array orders of grey. Many statistical values such as entropy, energy, contrast, grey correlation can be calculated from grey level co-occurrence matrixes. After image enhancement and downsizing of grey level, grey level co-occurrence matrixes of small regions are calculated. Then we can get texture features from grey level co-occurrence matrixes. And the last step is using k-means clustering to classify those feature vectors in order to segment objects from backgrounds. Most natural objects have fractal feature of grey level, so their fractal dimensions can be calculated to divide the road parts from non-road parts as well.3,Next, this paper puts forward a recognition method based on BP neural network. BP neural network is a simulation of human brain structure and learning skill, and it is the most popular artificial neural network. The texture feature vector is used as the input vector, the hidden layer has 20 nerve cells and the output layer has one nerve cell. Every small region's attributes can be judged through calculation after getting the weights and threshold matrixes.4,The light value is so low that it is impossible to recognize unstructured roads under shady conditions in hue space. This paper uses the maximum entropy method two times to classify the image into four parts: road areas, non-road areas, dubious road parts and dubious non-road areas. Finally, normalized color components are used to determine the attributes of anguish areas.5,After segmentation of an image, this paper utilizes mathematical morphology and area threshold methods to eliminate existing noises. Then we detect road edges by Prewitt operator. In the end, Least Mean Square curve filtering technology is used to get smooth edges.6,Experiments are implemented on the platform based on VC++6.0 and the results have proved excellent performance of the algorithms.