Background
This research aims to answer two open questions within the space of creating lidar maps using vehicle mounted sensors: how to automatically incorporate expert human knowledge into machine learning models to remove objects in the scene that don’t belong in the map; and how to execute those models in real time on vehicle-relevant hardware rather than the typical high-powered desktop hardware found in the literature. The central research question is aimed at answering if removing short term static objects from the map increases overall robustness. The team also hopes to discover and document best practices for the practical implementation of the results in diverse applications.
Approach
The first step to addressing these challenges is to design a structured representation of the data which can be fed into a neural network in such a way to take advantage of current neural network research. By unravelling lidar data into an image-like representation we can effectively train a neural network.
The second step is to use these labeled point clouds to effectively remove objects that result in non-robust features for simultaneous localization and mapping (SLAM) techniques. This is done in a two-stage process. Some features are persisted through the SLAM algorithm and removed in a post processing step as they have low temporal variance. Objects that have a high temporal variance are removed before the SLAM step.
The third step is re-localizing on maps with objects removed. This was done with a Lidar processing pipeline that computes correspondences between lidar clouds based on a scale invariant feature transform (SIFT) feature extractor.
Accomplishments
The team has designed, trained, and tested a neural network architecture to perform the map labelling task. This utilized open-source datasets and classical computer vision data preprocessing. The team also created a SLAM based pipeline for initial mapping and post processing of the labeled data. This allows for objects to be pre-processed before the SLAM stage and have a human operator efficiently clean the map in the post processing step. Finally, the team implemented a lidar scan matching pipeline for testing the work. The team was able to effectively demonstrate that removing long-term dynamic obstacles improved the robustness of the map.