3D Classification of Power Line Scene Using Airborne Lidar Data

Failure to adequately maintain vegetation within a power line corridor has been identified as a main cause of the August 14, 2003 electric power blackout. Such that, timely and accurate corridor mapping and monitoring are indispensible to mitigate such disaster. Moreover, airborne LiDAR (Light Detection And Ranging) has been recently introduced and widely utilized in industries and academies thanks to its potential to automate the data processing for scene analysis including power line corridor mapping. However, today’s corridor mapping practice using LiDAR in industries still remains an expensive manual process that is not suitable for the large-scale, rapid commercial compilation of corridor maps. Additionally, in academies only few studies have developed algorithms capable of recognizing corridor objects in the power line scene, which are mostly based on 2-dimensional classification. Thus, the objective of this dissertation is to develop a 3-dimensional classification system which is able to automatically identify key objects in the power line corridor from large-scale LiDAR data. This dissertation introduces new features for power structures, especially for the electric pylon, and existing features which are derived through diverse piecewise (i.e., point, line and plane) feature extraction, and then constructs a classification model pool by building individual models according to the piecewise feature sets and diverse voltage training samples using Random Forests. Finally, this dissertation proposes a Multiple Classifier System (MCS) which provides an optimal committee of models from the model pool for classification of new incoming power line scene. The proposed MCS has been tested on a power line corridor where medium voltage transmission lines (115 kV and 230 kV) pass. The classification results based on the MCS applied by optimally selecting the pre-built classification models according to the voltage type of the test corridor demonstrate a good accuracy (89.07%) and computationally effective time cost (approximately 4 hours/km) without additional training fees.