MSc Thesis Defence:A Non-Local Approach to Tree Peak Detection in LIDAR Data- Ssushant Joshi
The majority of existing methods for delineating trees from LIDAR point cloud use a region growing approach. Seed points representing the highest point of the trees (tree-peaks) are detected in the point cloud. The remaining points are iteratively assigned to one of the seed points, thus growing the region representing trees. The tree-peak detection methodologies are based on local geometry analysis, identifying locally highest points within some appropriately sized neighborhood as tree-peaks. Such approaches result in false positives over a tree with large crown and false negatives with multiple trees in close proximity. Hence, we advocate anon-local approach to tree-peak detection which analyzes mutual relation between points in the point cloud. Initially, we assume that each point in the point cloud could possibly be a true peak. Then a series of filters applied on the point cloud removes false peaks, eventually, producing superior results over the state of the art.