Automatic Fluid Edge Tracking in a Shaking Aquarium

Simplification of the real world through mathematical or physical modeling is one of the great ambitions of researchers in recent years. Researchers utilize these models to describe real-world events or interpret them. However, mathematical and physical models cannot describe the real-world event completely. Therefore, in this thesis to interpret and analysis real-world events, a physical model has been reproduced. The main objective of this thesis is to create an automatic method that can track and extract the necessary edges of the fluid within a shaking aquarium through computer vision techniques. Two methods have been applied and compared to serve this purpose. The first approach is creating a function to extract features from raw input by using traditional digital image processing. Th e second approach is to develop an algorithm based on deep learning architectures and techniques. The setup consists of a shaker, an aquar-ium tank, which is filled with water, and a flexible blade that is fixed in the middle of the tank. A shaker is used to generate motions in the tank by moving from left to right which conse-quently creates waves and motions on the water surface and causes the blade to bend. A high-speed camera is used to records these interactions and waves and save them as frames. These frames are taken as the raw input data to our function and algorithm to extract features.The traditional image processing operations were utilized to the frames in several steps. Since the input datasets contain a high level of noise, therefore, several methods of denoising and smoothi ng were applied to denoise the input data in order to extract useful edges from them. Nevertheless, the traditional methods were unable to reach the expectation that is to extract the desired edges from the water surface and its interaction with the blade.On the other hand, the deep learning technique does not contain limitations of traditional method. This approach requires more time and resources in terms of labeling and hardware respectively, but it delivers expected results. Deep learning approach was able to extract desired features and edges automatically in this thesis through the input datasets with high precision and accuracy. Despite a few studies had been conducted in this field, this thesis recommends further research in developing methods and approaches to track edges through deep learning object tracking.