Summary: | 碩士 === 國立臺灣科技大學 === 電子工程系 === 106 === As autonomous vehicles have received attention in the recent years, more and more researchers have investigated this active issue. Specifically, object detection in road scenes is one of major tasks for autonomous vehicles, which has showed impressive progress. In this thesis, after analyzing self-driving car datasets, we make the following observations: 1) small and medium-size object samples dominate the datasets, and major portion of objects are occluded, 2) the class-imbalance problem among minority classe degrades the performance of object detection. To address these problems, we use RoIAlign to improve the accuracy of small object detection, and use the focal loss to address class imbalance problem. With these methods, we successful improve the performance of object detection with the Faster R-CNN on KITTI dataset. On object detection from videos, we incorporate temporal and contextual information of videos with T-CNN framework. We also analyze the limitations of T-CNN based on the driving-scene datasets.
|