With the development of the laser scanning systems, and the increasing interest on three-dimensional city scene understanding and reconstruction, more and more efforts have been put on the research of object detection and recognition from laser scanner data. Pole-like objects such as traffic signs, street lighting poles, traffic lights are the key components of the road infrastructure. They play an important role in the modern traffic systems and are highly related to our daily life.
In this thesis, a pipeline for recognizing the pole-like objects from mobile LiDAR (Light Detection and Ranging) data is introduced. In the segmentation step, a local elevation context based region growing method is proposed. Comparing with traditional methods such as elevation histogram based method and curb detection based method, the introduced method is more robust to fluctuations and non-curb area. Further, by observing the key difference of object's penetrability between the vegetation and above ground objects, a local roughness based method is proposed. Based on the calculated local roughness of each point, points are classified into two categories, the vegetation points and the non-vegetation points, through a graph-cuts based method. In the recognition step, a Gaussian Mixture Model (GMM) based modeling method is introduced. Instead of using feature based recognition method which is sensitive to the noise background, in this research, a statistical model. based method is introduced. First, the points' distribution of each segmented cluster is modeled and described by the Mixture of Gaussians. Then, the candidate models are matched with the pre-trained sample models. Finally, the labels of the sample models are assigned to the best matched candidate to obtain the recognition results. Experimental results show that, the introduced recognition method achieves 90\% of the true positive rate on recognizing pole-like objects.