End-to-end Vehicle Tracking and Counting in Traffic Videos

With the reduced manufacturing cost of cameras and the progress in the computer vision field, intelligent transportation via computer vision has raised much attention. However, there remains a huge gap between academic computer vision research and application. There lacks enough attention to end-to-end computer vision system real-time processing speed. This thesis aims to bridge the gap between the state-of-the-art computer vision research and real-world application. We first address the critical problem of proper initialization and termination in object tracking algorithm and propose a heuristic method for automatic tracking initialization and termination in chapter 2. Then we work on learning the scene-specific semantic knowledge and apply them for other tasks such as vehicle tracking and counting in chapter 3 to 5. Chapter 7 describes our public dataset from real traffic cameras. Chapter 2, 6 and 7 consist of the work before the preliminary exam, which is a complete end-to-end vehicle tracking and counting system running in real time. We demonstrate the performance improvement by the heuristic method and further boost by the semantic knowledge.