Used ros2, gazebo, keras and pangolin to gather lidar data, render it, and classify simple geometries with a custom neural network model that achieves 97.6% accuracy. Tried out some other other architectures and they were ranging from 85% to 93% accuracy.
- .sdf files with a lidar for gazebo simulations
- .npy files with lidar data points
- lidar_listener.py: Extracts gazebo lidar data (ros_ign_bridge needed)
- cube_detector.py: Keras model for detecting cubes from clusters in point clouds
- pangoclouds.py: Process, renders, and detects cubes (from cube_detector) from lidar data in pangolin
pangoclouds.py loads a point cloud, removes noise, clusters the datapoints and with a neural network identifies cubes from different geometries successfully with an accuracy of 97.6%