Classification Of On-board Laser Point Cloud Based On CRF Model

Mobile Laser Scanning(MLS)is one of the most popular platforms for data acquisition,which can quickly,actively and continuously obtain the spatial 3D information and spectral information of the target object.In recent years,MLS has been one of the most effective means for outdoor scene reconstruction.However,in the process of object sampling,MLS treats all kinds of ground objects indiscriminately,resulting in the lack of semantic information in the mass point cloud,which is needed in the application fields such as unmanned driving,forestry investigation,urban planning and robotics.The semantic interpretation typically consists in assigning a semantic label(e.g.building,ground or vegetation)to each point of the considered 3D point cloud,Traditional point cloud classification relies on man-machine interactive interpretation,which has high labor intensity and low efficiency,and cannot meet the requirements of real-time interpretation of object attributes.Therefore,the research on automatic point cloud classification algorithm for MLS is of great significance for quickly obtaining target semantic information.The traditional method of MLS point cloud classification is to calculate some local feature values for all points in the point cloud,then combine these local feature values into a feature vector which called descriptor,and finally use the classical classifier such as random forest or support vector machine to supervise the classification.In the paper Weinmann et al.(2015a),the classification process of laser point cloud based on machine learning algorithm was presented,including four parts: neighborhood selection,feature calculation,model training,and label prediction.The paper proposed to calculate multiple low level geometric 3D and 2D features as feature descriptor of points,which achieved good classification results.However,many features in low level geometric 3D and 2D features are calculated from the three eigenvalues of the tensor matrix,the correlation between features results in redundancy and information loss in feature descriptors.What's more,the classification is only based on the feature vector of the point,and the context information of the point cloud is not considered,which results in a lot of noise in the classification results.By studying the influence of existing point local feature descriptors on the classification results of point cloud,this paper proposes to combine FPFH feature with low level geometric 3D and 2D features as feature vector for point cloud context classification based on CRF algorithm.The specific research is divided into the following two stages:1.Initial classification by combining FPFH feature with low level geometric 3D and2 D features.Firstly we compute an optimal neighborhood for each point of the original point cloud,and then based on the optimal neighborhood,we calculate both FPFH feature and low level geometric 3D and 2D features and combine these two features as the feature vector afterwards,after that we use the random forest algorithm to train a classifier about the training dataset and predict semantic label for the test dataset,finally,for each point of the initial point cloud,we will obtain both a semantic label and probability values for all labels it can take.2.Classification results optimization based on the CRF algorithm.As the classification process of random forest algorithm does not take into account the similarity or correlation between the labels of adjacent points,the spatial smoothness of the initial classification results is not high and there is a lot of noise.For this problem,undirected graph model is established for the original point cloud to express that the label of each point is influenced by labels of its k neighborhood points,then two different random forest classifier are trained to calculate unary potentials and pairwise potentials of the undirected graph model separately,posterior probability p(y?x)refers to the probability of point cloud x taking the labels y,it can be calculated with unary potentials and pairwise potentials based on the undirected graph model.Obviously,among all possible label vectors y,the one making joint probability p(y?x)biggest is the best.By combining FPFH feature with low level geometric 3D and 2D features,experiment result shows that the classification accuracy is 2?6% higher than that of merely using low level geometric 3D and 2D features.