Object tracking deals with movement analysis in which the algorithm tracks an objects throughout an image sequence. It is commonly used in the field of security systems, autonomous vehicles, robotics and many more. Under some circumstances object tracking algorithms fail meaning they no longer track the correct object. In the diploma thesis we use two different approaches for predicting object tracker failure, which could run parallel with a given tracker. The first approach is based on image features and the second on optical flow analysis. For bot approaches we processed data from the VOT2019 competition. We separated the data into failed and successful tracking and trained the failure prediction models on it. We tested the final models on whole videos and the data from the trackers and compared them at the end.
|
