Multi Person Tracking in Real-time Video Streams


The goal of this project was to demonstrate how it is possible to combine a video based object tracking system with a robotic device which moves, with some degree of precision, according to the movement of the tracked object. More specifically, a person-tracking system was developed which analyses video captured by a Pan-Tilt-Zoom (PTZ) camera to estimate a target person’s movement relative to that camera.

Those detected movements are then translated to commands which are sent to operate the robotic device.

Accurately tracking people, or any object, in real-world scenarios is still an open computer vision problem, and a challenging one. For the scope of this project, the approach was to use the tools provided by the OpenCV computer vision library to be able to track a target person up to some degree of sudden changes in movement direction, occlusions and ambient light conditions. These are the main real-world parameters which affect object tracking.

As a robotic device, a moving light with 2 Degrees of Freedom (Pan and Tilt) was used as a proof of concept. This robotic light can be controlled through its DMX input interface, an industry standard for lighting control. For the integration of this DMX interface with the controlling computer, we have used an Arduino Uno board fitted together with an Arduino DMX Shield which acts as a bridge and converts the commands sent by the computer to DMX signals.

With this system we are able to light the way for a moving person, although not without limitations. The off-the-shelf robotic light was obviously not developed for this purpose, so its movement commands do not allow enough granularity for the accuracy of this type of application and also its reaction time to commands is not very fast.


Author: José Sá

Type: MSc thesis

Partner: Faculdade de Engenharia da Universidade do Porto

Year: 2013