Research On Terrain Recognition In Unstructured Environment By Image Features And Active Learning

Ground intelligent robot is a kind of robot device which can run autonomously or semi-autonomously through the cooperation of sensing,processing,decision-making and implementation modules on the basis of information from its sensor deceives.As the huge demand in military,civilian and space exploration areas,the intelligent robot technology is undergoing rapid development.With the relative to the structured and semi-structured environment,the natural surface in unstructured environments such as grassland,land,and sand are more challenging environment to understand more.It because that there are various kinds of light conditions and surface categories in the natural environment,in addition landscape category is uncertain in every image.And coupled with the variability of weather conditions,and seasonal variation of surface and scenery unpredictable situation,the difficulty of environment understanding to intelligent robot on the ground of unstructured environment would be complex and challengeable.This article mainly aims at issues,which conclude that the ground robot in unstructured environments of passable area detection and traffic-ability discriminate based on the active learning make in-depth discussion and research.The main contents are described below.For the problem that the features from single color channel have the disadvantage of having poor recognition effect of the shadow region,a global texture feature is proposed,which contains multiple channels features and spatial information and reflects the global information of a sub-image in original image in order to solve the poor recognition from shadow.First the proposed feature extract multi-channel DCT texture characteristics of image,and then extract the image color covariance characteristics,finally extract the global spatial information,which contains the local information and spatial information of image features.Comparing the proposed features with other features in area detection by multiple classifiers,the experimental results show that the recognition effect of multichannel DCT texture feature for illumination changing region is improved obviously,and the multi-channel color texture feature with spatial information reduces the discontinuity of areas.In this way,the proposed features help to improve the ground robot's understanding ability to adapt to the various environments.Taking into account that the natural surface area in natural environment always has spatial continuity,therefore the proposed color texture feature will be integrated into the global space information.In addition,since the each color component of color image is sensitive to light conditions changing,therefore the proposed color texture feature make an expansion within multi-channel characteristics,and contrives a fusion of spatial information multi-channel color texture feature.Experimental results show that the global space information help the identification process to reduce regional discontinuity and its extended form on multi-channel make a great contribution to the robust ability with obviously light changes,recognize the region has improved,easing the problem that the low recognition rate bring due to the light of changes.For the problem that the traditional active learning algorithms have the disadvantage of sample learning unreasonable for unbalanced dataset,the adaptive active learning algorithm is presented based on unbalanced data set,and the algorithm is used to solve the passable area identification for unstructured roads.On the one hand,the algorithm can build the sample similarity matrix based on the similarity of the sample set between each sample point to find the data sets on global and local information on the classification of the most discriminating subset of samples.On the other hand,the algorithm can adaptively selected sample subset of the regulation based on the number of categories,so that the selected subset of samples that cover all categories.In this way,the learning result avoid the problem that the small sample size category is ignored by mistaken to take into noise data.Experimental results show that the proposed algorithm can correctly identify accessible area with the accurate rate 98.6 percent in sequential images under the situation of the training set consisting of only a small number of samples,which fully demonstrates the effectiveness of the proposed algorithm and that it improves adaptive capacity of ground intelligent robot for each kind of road environment.For the problem that rectangular feature windows at the regional boundaries of unstructured environment are prone to the mixed samples of areas and single kernel function in SVM algorithm is not good at mapping heterogeneous characteristic vector,the recognition algorithm for passable area is put forward,which based on superpixel segmentation and multi-kernel learning algorithm.First,extract the superpixel feature vector on each superpixel based on the SLIC superpixel segmentation and build training sample set with high confidence.And then,learn a multi-core linear combination model on the training sample set by multi-kernel learning algorithm,and identify which area is passable.The proposed algorithm extracts the features from the pixel level windows and learns a multi-kernel linear combination mapping at the same time do the features fusion for kinds of features.In this way,the recognition of passable area has strong robustness for the scenes containing various noise data.Otherwise,the proposed algorithm has strong adaptive ablility for the diversity of scenes by building adaptive unsupervised training set with high confidence.Against the problem of accessible discrimination of a vehicle in unstructured environment,the accessible discrimination algorithm based on active learning and sparse representation is presented for the sequences of images from unstructured environment.This algorithm intends to give discrimination about if the current position is accessible avoiding the abnormal conditions brought by the far deviation from the accessible area but the intelligent root continuing run straightly.from the perspective presents a sequence of images identification based on active learning dictionary passable discriminant algorithm.Experiments show that the proposed accessible discrimination algorithm could initiative identify accessible image from the sequence of image by active dictionary learning and the algorithm can provides location opposite direction which provide a basis for the next direction adjustment by intelligent robot.