Educational project
Object reidentification in a multiple surveillance camera system
SUPSI Image Focus
Surveillance cameras are pervasive elements in our cities today: they act as a deterrent against illegal activities or traffic offences, help investigate crimes and can provide useful analysis of traffic and people. The collection and analysis of video streams from multiple sources, possibly in real time, poses a number of ethical, legal and technical challenges. As engineers, we mainly focus on the latter issue, leaving the first two to politicians and lawyers.
In this project, a solution was created that can re-identify the same object when it moves from one camera to another. This approach allows surveillance operators to quickly search the system using morphological or chromatic search criteria, instead of watching hundreds of video streams, and then allows operators to automatically select and track interesting objects within different surveillance cameras located in diverse areas of a city.
In this work, we developed a software algorithm capable of re-identifying pedestrians and vehicles in different video streams. When attempting to re-identify a particular object, the goal of the algorithm is not to re-identify the object with a probability of 100 per cent, but rather to present a list of objects that could match the original object.
Re-identification is done by calculating a distance between the different objects, which corresponds to the overall
difference between these objects. The distance is calculated by running several algorithms on the objects; each algorithm considers different characteristics of the objects, such as shape or colour, to calculate how different one object is from another.
Conclusions
In this work, we developed a software algorithm capable of re-identifying pedestrians and vehicles in different video streams. When attempting to re-identify a particular object, the goal of the algorithm is not to re-identify the object with a probability of 100 per cent, but rather to present a list of objects that could match the original object.
Re-identification is done by calculating a distance between the different objects, which corresponds to the overall
difference between these objects. The distance is calculated by running several algorithms on the objects; each algorithm considers different characteristics of the objects, such as shape or colour, to calculate how different one object is from another.