A Comparative Study of Segmentation and Classification Methods for 3D Point Clouds

Abstract

Active Safety has become an important part of the current automotive industry due to its proven potential in making driving more joyful and reducing number of accidents and causalities. Different sensors are used in the active safety systems to perceive the environment and implement driver assistance and collision avoidance systems. Light detection and ranging (LIDAR) sensors are among the commonly utilized sensors in these systems; a LIDAR produces a point cloud from the surrounding and can be used to detect and classify objects such as cars, pedestrians, etc. In this thesis, we perform a comparative study where several methods to both segment Region Growing and Euclidian Clustering) and classify (Support Vector Machines, Feed Forward Neural Networks, Random Forests and K-Nearest Neighbors) point clouds from an urban environment are evaluated. Data from the KITTI database is used to validate the methods which are implemented using the PCL and Shark library. We evaluate the performance of the classification methods on two different sets of developed features. Our experiments show that the best accuracy can be obtained using SVMs, which is around 96.3% on the considered data set with 7 different classes of objects.