Real Time Human Activity Analysis in Wide Environments Using Multiple PTZ Camera Feeds

Visual object detection and distance estimation are tasks frequently invoked in many applications. When the objects are human, their mutual separation and possible interaction are interesting for problems such as sports analysis and surveillance. We look at the difficult case when the individuals are so far apart that they cannot both simultaneously register in a single camera due to FOV limitations, and therefore need to be captured individually in separate cameras. This work addresses the distance estimation problem between two persons by developing a novel end-to-end learning-assisted model that processes synced videos shot from 2 independently steered and zoomed co-located PTZ cameras. By a triangulation strategy, we detect the persons, track them in real time, and measure both their mutual distance as well as their respective distances from the (common) camera location. Furthermore, we employ the YOLOv8 posture estimation approach to detect hand raise events in order to ascertain the likelihood of interaction between the two individuals.