الفهرس 
Main ContributionOnly 14 pages are availabe for public view
 |  | from | 99 |  |  |
| | | from | 99 | | |
Abstract
3D Object detection is one of the most important perception tasks needed by autonomous vehicles to detect different road agents like other vehicles, cyclists, and pedestrians which is required for driving tasks like collision avoidance and path planning. State of the art 3D Object detection deep learning models take as input point clouds from very expensive LIDAR sensors because it provides accurate depth information. Another alternative is using cameras which are far cheaper but lack accurate depth information which causes a huge gap in average precision from LIDAR models vs camera models. In this thesis, our work is focused on 3D Object Detection for car class from stereo images without LIDAR supervision neither during training nor during inference and the challenging task of running 3D Object Detection on an embedded target Nvidia Jetson TX2 by modifying Stereo R-CNN model and reducing the model size to approximately one third the size of the original model to be more suitable for embedded targets. Experiments on KITTI dataset showed that our model’s inference time is 1.8 seconds and its’ average precision for moderate car class is 17% on the test set. Our model decreases training and inference time by approximately 60% with a 13% DROP on the test set which is an expected trade-off when decreasing the number of parameters inside the model.