Department of Image Processing

In recent years, we have witnessed an expansion of road traffic that causes several problems, especially in urban areas. Local authorities and governments invest lot into traffic surveys to keep track of infrastructure bottlenecks. At the same time, researchers have made significant progress in deep learning, allowing for automated, accurate and performant solutions for computer vision tasks. In our work, we aim to create an open-source software tool that enables the automation of traffic surveys. We investigate approaches to multi-object tracking and suggest a framework that analyses a stationary camera recording of road traffic and counts the vehicles that move through the user-defined regions of interest. We evaluate the performance of our framework on a couple of datasets, comment on metrics and hardware usage, and report accuracies. We conclude with findings and requirements placed on the source videos; we suggest a few improvements to the framework. We identify that the framework works well and within the boundaries of allowed error if there are no or just a few traffic signs or automobiles blocking the view.

Thesis: https://dspace.cvut.cz/handle/10467/101104
PPTX: https://docs.google.com/presentation/d/1dJkiO3_c8qPfXYJRnV65lON51RspRhy5...