Examiner: Lennart Svensson
This thesis evaluates the performance of applying machine learning object detection models, developed for lidar point clouds, to 3+1D radar point clouds. The RADIal dataset, a rural, publicly available, 2D annotated, car-centric dataset is used to train and evaluate the models. A method for generating 3D BBs, which are required to train the models, using the available 2D annotations and lidar point clouds is presented. Evaluation is performed on multiple object detection models, including a point-based model, 3DSSD, and three voxel-based models, PointPillars, Voxel-R-CNN and SECOND\@. When training and evaluating on lower resolution lidar data and 3+1D radar data, the performance is shown to drop dramatically for the voxel-based models.
The point-based 3DSSD model is, on the other hand, shown to generalize well to the lower resolution point clouds, both with and without Doppler shift and Return Signal Strength feature information. Other qualities are evaluated, and two model augmentations are chosen to further improve the performance of 3DSSD: multi-frame accumulation and distance-based model selection. Multi-frame accumulation is shown to decrease performance when ego-motion compensation is unavailable. Distance-based model selection, where an ensemble of models are trained to predict objects at different distances, is shown to achieve performance on radar data that matches the 3DSSD model´s performance on lidar data from the same dataset.
Rikard, Samuel and Lennart
Analysen, meeting room, Rännvägen 6B, EDIT trappa D, E och F