Dynamic Background Subtraction for High-speed Sequences

The presence of intelligent robots and machines is fundamental for our daily and professional lives. They often perform tasks that we find tedious or too difficult to do ourselves. A common such task is to understand and recognize events in images or video. One might want to track how many cars cross an intersection in a day, or if a certain person can be recognized in a crowd of people. A scene from a camera generally has regions of interest (such as moving cars and people), and other regions that do not provide any useful information (background). Separating these regions is essential when doing the tasks of tracking and recognition. A common way of doing this is the so called background subtraction. Conventionally, background subtraction is done without considering the characteristics of different regions of a scene. This thesis aims to find ways of exploiting high-speed cameras to acquire and use the hidden information in between the frames from a camera of regular frame rate. It especially focuses on developing an improved background subtraction method with these ideas in mind. The method proposed in this thesis introduces a technique called “Targeted learning”. This technique targets regions in the background with high dynamics (such as leaves in a tree or moving shadows), and learns them at high speed. This enables high-speed modeling of highly dynamic regions in the background while keeping foreground objects in the foreground. A set of experiments with high-speed image sequences was conducted to test and develop the proposed method. The results from these experiments show that the method is a significant improvement over conventional background subtraction methods. In particular the ability to correctly classify foreground objects is greatly improved.