Research On Image Interpretation Algorithms For Unstructured Scene

Ground intelligent robot is a kind of robot which can use its own sensor devices to understand and judge complex ground environment, and on this basis, carry out planning and decision-making to make it run autonomously and continuously. With the promotion of urgent demand on military, civilian and exploration of the universe, etc, ground intelligent robot is developing with each passing day. Ground scene interpretation is the core technology of the autonomous navigation control system of the intelligent robot. Comparing with the structured scene interpretation, such as indoor environment, highways, etc, it is much more challenging for the ground intelligent robot to research on some problems of unstructured scene interpretation. Because the unstructured scene interpretation is effected more seriously by the complex situations in the unstructured environment, for example, illumination, scene, weather and ground irregularity, etc. In this paper, some key technologies on unstructured scene interpretation are studied, and some results of research have been achieved. The main contents of this dissertation include the following aspects:A terrain classification algorithm based on Gaussian Mixture Model for unstructured complicated environment is proposed, considering the illumination variation. First, color features and improved texture features are extracted and integrated. Terrain features in different illumination conditions are computed as training data. Then the Gaussian Mixture Model is trained on these training data. Terrain features with a variety of performance can be well made statistics by the GMM. The number of mixture models of the GMM is attained by the Bayesian information criterion. In addition, one classification strategy is proposed. The principle of this classification strategy is that terrain class of current feature window is determined by the average probability of neighbor windows. Terrain classification in complex unstructured environment can be carried out using the above method.The general scale and rotation invariant features are independently extracted in every frequency sub-band, the relationship between different frequency sub-bands is not considered, and the existing algorithms are seldom applied to outdoor scene recognition. In order to extract better features, two improved algorithms are proposed. The first algorithm is that on the basis of the existing features, the relationship features between high frequency sub-band and low frequency sub-band in the same scale are extracted by constructing histogram. The scale invariance operation isn't applied in this feature, so it is only a rotation invariant feature. The second algorithm is that on the basis of the existing features, the scale invariance operation is applied in Radon transform coefficient matrixes by using logarithmic function, then the relationship is extracted between all the frequency sub-bands under different color planes using linear regression model. Experimental results demonstrate that the performance of two proposed algorithms is good, especially the latter shows well in outdoor scene recognition.Aiming at the problems and limitations in the process of applying general active learning algorithm for the traversability region classification, two strategies are used to improve general SVM active learning algorithm. The first is to a dynamic clustering for selecting the best representative samples, and the other is to tune the SVM hyperplane location according to the difference between expert labeling and SVM classification results currently. On this basis, a new SVM active learning algorithm is proposed, that is KSVMactive. Aiming at difficulties in labeling caused by a large number of the samples, as well as uneven distribution of the samples in the traversability region classification, a nonlinear active learning algorithm based on AUC optimization is proposed. All the classified samples are scored using the sample selection function based on AUC optimization, then the best representative samples are selected according to the scores. This algorithm can well slove the sub-optimal solution problem caused by using loss function and error rate minimization. Experimental results demonstrate that the proposed two algorithms can significantly reduce the workload of labeling the samples, and the classification results are more or less with manual way.The majority of the available classification systems for the traversability region classification focus on the minimization of the classification error rate. This is not always suitable, especially solving the problems with skewed samples distributions and different classification cost. In order to obtain better classifier, two AUC optimization algorithms for training the classifier are proposed. The first algorithm is using dynamic clustering to select the best representative samples and label them by the expert firstly, then add these labeled samples in training set. Finally, a linear classifier is trained using a new AUC maximization method on training set. This algorithm can effectively solve the difficulties of labeling and ranking caused by excessive samples and uneven distribution of samples in traversability region classification. The second algorithm is to introduce the particle swarm optimization into the optimize the AUC objective function, and the particle swarm optimization algorithm is improved by using the Butterworth curves to adjust the change of the particle weight and particles with the lower fitness value being mutated to improve the global search capability of particle swarm algorithm. This algorithm can well solve some problems, such as the precision of the parameter estimation decreasing caused by using gradient descent in the process of the AUC optimization, etc, so that the performance of training the classifier by using the AUC maximization is improved.In the end, the whole paper is summarized and the prospective researching fields are discussed.