Abnormal motion from video

In this application:

• We train a model which translates images of “normal” events to their corresponding optical flows and detect anomalies at inference based on the reconstruction error.
• Independently, an explainability of the abnormal events may be achieved by analyzing the entropy of the distribution of vessel positions and velocities, obtained by object detection and optical flow calculation.
• By projecting the vessel detections and regions of abnormalities to the globe, it is easier to localize the events and notify them.

For more details, check out our paper published IEEE ITSC 2021.