| Over the past decades, considerable progress has been made in the area of automatic image interpretation using computer vision and pattern recognition methods. However, there is still a large gap between the requirements of most image interpretation applications and the accuracy and reliability achieved by automatic methods. Many attempts to automate these tasks are too fragile, and they require checking by experts before any final decision can be made. For this reason, most successful systems retain a "human in the loop", where a human operator can aid the automatic image interpretation through human-computer interactions (HCI).; In this thesis, we introduce a framework for image interpretation based on HCI. This framework consists of five components, a human-computer interface, a user model, computational image interpretation models, a knowledge transfer scheme, and a performance evaluation scheme. We applied this framework to image feature tracking in cartographic map revision, which is expensive, time-consuming, and currently has to be done manually. We implemented an interface to access, record, and parse human actions. The human data was used to predict user actions (such as view changes) using hidden Markov models and to develop a computer-assisted road tracking system. In this system, the human operator retains complete control over the operations with the computer acting as an apprentice and, later, as an assistant. The apprentice learns simple tasks from the human operator by tracking and modelling all input operations in road tracking. Eventually the apprentice can take over these tasks from the human and execute them, returning control to the human operator whenever problems arise. Two tracking methods were implemented, using Bayesian filtering and profile matching, as well as online learning and novelty detection. Experimental results confirmed that our approach is effective and superior to existing methods. Our approach is computationally efficient, and it can rapidly adapt to dynamic situations where the image feature distributions change. |
