Street Scene Segmentation And Recognition Using Super-Segment

Street scene recognition is an important component in automatic driving and city virtualization. Objects in the street, such as house, fence, tree, and so on, need to be recognized during the automatic driving. City virtualization replaces all the real objects in the street with virtual3D models. Therefore, the real objects need to be classified first in such applications.There are two problems in the existed methods to classify objects in street scene: the computation of training and recognition is time consuming, and the commonality of the classifier is not high enough, due to the complex features chosen. This thesis proposes a method based on the Super-Segment, whose features is simple and easy to compute. The method reduces the time of the training and recognizing, and improves the commonality at the same time.Our procedures are as below:all points, which belong to the ground, will be ex-tracted from the points cloud first, and our algorithm performs well even if the ground has some slope; and then the points cloud will be divided into small segments accord-ing to their adjacency using K-D Tree; the plane parameters of each segment will be derived and utilized to merge the segments to Super-Segments; features of all Super-Segments, such as height and projection properties, will be derived and trained partly using ANN and AdaBoost; eventually, objects in the scene will be recognized using the classifier.Method in this thesis results precision of90.9%for ANN and93.7%for Boost, and outperforms other existing methods for the object house, fence and car.