Unusual Trajectory Detection: Advances

There has been a growing interest in unusual behavior detection in computer vision and image processing because of its wide range of applications such as traffic surveillance, human/animal behavior understanding, and elderly surveillance. The most traditional methods for unusual behavior detection are trajectory based, which rely on different trajectory representations, feature extraction, and learning methods. In this work, a comprehensive review including different trajectory representations and learning methods for unusual trajectory detection is presented. Additionally, a comparative analysis using different computational methods applied to real-world datasets such as fish and pedestrian trajectory was performed. To the best of our knowledge for the first time in this work, active learning with feature selection is applied for unusual trajectory detection which presents sufficiently good results even with much less training data.