Abstract
The recent surge in the development of self-driving cars stems mainly from the breakthrough in the artificial neuron networks, a.k.a. the deep learning technique, for recognizing and tracking objects on the roads. To apply the deep learning technique, a large set of properly annotated videos, which serve as the ground truth, are needed to train the neuron network. Although it is not difficult to collect a large amount of road videos from the dashboard cameras of cars, it is very time-consuming and tedious to annotate the videos. So far, this task mainly relies on human to annotate. In this thesis, we study a semi-automatic approach to video annotation, which annotates objects in the video using existing automatic object trackers and then corrects possible errors by human. Note that existing automatic object trackers are likely developed using the deep learning technique based on a certain set of annotated videos. Unfortunately, no object tracker can guarantee object recognition and tracking as good as human. They will still make mistakes. In our semi-automatic approach, these errors are corrected by human. The key issue in this approach is how to detect the errors made by the object trackers and guide the human annotators through the correction process. In this thesis, we propose to examine motion features in the video that has been annotated by the object trackers and check for conditions that may lead to erroneous annotations. Such problematic frames are identified and presented to the human annotators to check and correct. We also verify the accuracy resulted from this tool based on the object location and the life time period. Overall, this tool can save 60% of works while ensuring data accuracy on recognized objects.