Abstract
The University of Central Florida invention is a scalable, end-to-end, self-supervised system and method for training object detectors for stationary light detection and ranging (LiDAR) sensors. By not requiring manual labeling and classification, the scalable and cost-effective methodology can accommodate more input data and is a substantial improvement over previous techniques. The system uses the training of a robust deep detector model (a student model) based on labels automatically generated by a self-supervised noisy modeled detector (a teacher model). Together, the detector models train a highly resilient deep detector model by using the collective knowledge of all captured and labeled data. Subsequently, the trained deep detector model generates enhanced labels for additional data inputs, which serve as improved ground truths for retraining. The teacher model includes submodules of a preprocessor, a background filter, clustering, and an object classifier and bounding box regressor.
Partnering Opportunity
The research team is seeking partners for licensing, research collaboration, or both.
Stage of Development
Prototype available.
Benefit
Easily scalableLow costFasterMarket Application
LiDAR sensor manufacturersGovernment departments of transportationAutonomous vehicle industryWarehouse management
Brochure
