About classfication encoding: https://desktop.arcgis.com/en/arcmap/latest/manage-data/las-dataset/lidar-point-classification.htm
Some data structures referenced : https://github.com/luizgh/knn
build:
All files can be downloaded here.
Use all files inside src folder to build this project. Please be aware of the settings of the CMakeLists.txt if the program runs across platforms.
Compile and run:
After compiling all the source files, a binary executable file (Kmeans.exe) would be obtained. Run it directly and no arguments are required.
It automatically reads in all .xyz files inside data folder.
The output files are in the same data folder as the input files.
- dataset_truth.laz -- the ground truth dataset (combining the 500 pointcloud files)
- dataset_result.laz -- the classification dataset(combining the 500 pointcloud files)
- report.txt -- the classification statistics report
Comment:
Manhattan distance is used by default in this program as using it has an relatively ideal result.
Euclidean distance has also been tested, the corresponding result files are:
(1) area : (max_x - min_x) * (max_y - min_y)
(2) height difference : max_z - min_z
(3) density : number of points in one file / area
(4) ratio : x_distance / y_distance (the ratio of width / height)
Numbers:
Correct numbers: 349 out of 500
Wrong numbers: 151 out of 500
Accuracy:
Avg accuracy (mAcc) = 0.776148 -- the average of the accuracy rates for the five categories
Overall accuracy (OA) = 0.698 -- total numbers of correct data / total numbers of data