LiDAR data analysis for automatic region segmentation and object classification

Light Detection and Ranging, (LiDAR) presents a series of unique challenges, the foremost of these being object identification. Because of the ease of aerial collection and high range resolution, analysts are often faced with the challenge of sorting through large datasets and making informed decisions across multiple square miles of data. This problem has made automatic target detection in LiDAR a priority. A novel algorithm is proposed with the overall goal of automatic identification of five object classes within aerially collected LiDAR data: ground, buildings, vehicles, vegetation and power lines. The objective is divided into two parts: region segmentation and object classification. The segmentation portion of the algorithm uses a progressive morphological filter to separate the ground points from the object points. Next, the object points are examined and a Normal Octree Region Merging (NORM) segmentation takes place. This segmentation technique, based on surface normal similarities, subdivides the object points into clusters. Next, for each cluster of object points, a Shape-based Eigen Local Feature (SELF) is computed. Finally, the features are used as the input to a cascade of classifiers, where four individual support vector machines (SVM) are trained to distinguish the object points into the remaining four classes.The ability of the algorithm to segment points into complete objects and also classify each point into its correct class is evaluated. Both the segmentation and classification results are compared to datasets which have been manually ground-truthed. The evaluation demonstrates the success of the proposed algorithm in segmenting and distinguishing between five classes of objects in a LiDAR point cloud. Future work in this direction includes developing a method to identify the volume changes in a scene over time in an effort to provide further contextual information about a given area