The script will load sample point cloud (x,y,z,scaled_time) and perform trajectory estimation and motion and instance segmentation. The installation is only via pip and the script run on CPU.
pip install matplotlib
pip install scikit-learn
pip install open3d
To use it, run the following command:
python icp-flow.py
The script describes the parameters and variable used to plotting for extraction of the output.
| Parameter | Description |
|---|---|
| EPS | DBSCAN parameters for grouping points, tunes over and undersegmentation |
| MAX_ADJACENT_TIMES | Number of consecutive time frames considered for DBSCAN clustering |
| Z_SCALE | Scaling the heights for clustering |
| MIN_NBR_PTS | Minimal number of instance points in all times for reconstruction consideration |
| MIN_INCLUDED_TIMES | The instance must have been observed in this many frames |
The main part of the script outputs the following variable that can be used for storing the results. The mask is mapping results to the loaded point cloud.
| Variable | Description |
|---|---|
| dynamic_mask | 0/1 if the point is static or dynamic |
| dynamic_id | indicates whether the instance is static or dynamic |
| trajectories | Trajectory coordinates per time |
| boxes_dict | Fitted Bounding box (x, y, z, l, w, h, theta) |
The PONE sample sequence can be than visualized as indicated in the script.
For own data, it is recommended to load the points (x,y,z,t) into the pts variable and scale the time accordingly to the clustering (such as the comments in the script) or import the ICP4dSolver function into your framework.