LIDAR data classification and compression
Metadata[+] Show full item record
Airborne Laser Detection and Ranging (LIDAR) data has a wide range of applications in agriculture, archaeology, biology, geology, meteorology, military and transportation, etc. LIDAR data consumes hundreds of gigabytes in a typical day of acquisition, and the amount of data collected will continue to grow as sensors improve in resolution and functionality. LIDAR data classification and compression are therefore very important for managing, visualizing, analyzing and using this huge amount of data. Among the existing LIDAR data classification schemes, supervised learning has been used and can obtain up to 96% of accuracy. However some of the features used are not readily available, and the training data is also not always available in practice. In existing LIDAR data compression schemes, the compressed size can be 5%-23% of the original size, but still could be in the order of gigabyte, which is impractical for many applications. The objectives of this dissertation are (1) to develop LIDAR classification schemes that can classify airborne LIDAR data more accurately without some features or training data that existing work requires; (2) to explore lossy compression schemes that can compress LIDAR data at a much higher compression rate than is currently available. We first investigate two independent ways to classify LIDAR data depending on the availability of training data: when training data is available, we use supervised machine learning techniques such as support vector machine (SVM); when training data is not readily available, we develop an unsupervised classification method that can classify LIDAR data as good as supervised classification methods. Experimental results show that the accuracy of our classification results are over 99%. We then present two new lossy LIDAR data compression methods and compare their performance. The first one is a wavelet based compression scheme while the second one is geometry based. Our new geometry based compression is a geometry and statistics driven LIDAR point-cloud compression method which combines both application knowledge and scene content to enable fast transmission from the sensor platform while preserving the geometric properties of objects within a scene. The new algorithm is based on the idea of compression by classification. It utilizes the unique height function simplicity as well as the local spatial coherence and linearity of the aerial LIDAR data and can automatically compress the data to the desired level-of-details defined by the user. Either of the two developed classification methods can be used to automatically detect regions that are not locally linear such as vegetations or trees. In those regions, the local statistics descriptions, such as mean, variance, expectation, etc., are stored to efficiently represent the region and restore the geometry in the decompression phase. The new geometry-based compression schemes for building and ground data can compress efficiently and significantly reduce the file size, while retaining a good fit for the scalable "zoom in" requirements. Experimental results show that compared with existing LIDAR lossy compression work, our proposed approach achieves two orders of magnitude lower bit rate with the same quality, making it feasible for applications that were not practical before. The ability to store information into a database and query them efficiently becomes possible with the proposed highly efficient compression scheme.