LISA is a physics based augmentation method that models effects of adverse weather conditions on lidar data. It treats effects of small scatterers on average using the Beer-Lambert law where extinction coefficients are calculated thorugh Mie theory and particle size distributions. Rain and snow models use a hybrid Monte-Carlo method to augment clean lidar point clouds for a given rain rate. LISA can be easily extended to other scatterers with different droplet size models or materials (i.e dust).
One of these effects is reduced range due to attenuation of the laser signal due to scatterers. Note sparsity
of the rainy scenes at large distances from the sensor:
Particles near the sensor can backscatter enough light and cause false detections (blue box). Also note "fuzziness"
of the scan lines under adverse weather arising from reduced SNR (signal to noise ratio) which leads to range uncertainty.
Use the environment.yml file to install the required libraries for point cloud processing and open3d jupyter visualization gen_noisy_point_cloud.ipynb inside python_old has jupyter visulaization code of the point clouds lisa_simulation.py inside python_old has code to simulate on the entire kitti velodyne data and generate simulated data in the path specified.
For original paper code look here which has implementations for fog and snow as well. It is tested better than the current project but slower which is why we switched from a pure python implementation to C++ with python bindings.
For developers: Documentation for the c++ code can be generated using Doxygen.
- Write unit tests for the c++ implementation
- Implement fog and snow models
Cite as