Pedestrian tracking systems are commonly used for visual surveillance, but are also a key component in systems with other objectives such as activity recognition and behaviour understanding. A variety of methods have been developed for tracking single or multiple pedestrians in static or moving cameras by exploiting different types of image information.
Background separation methods track foreground objects by measuring the differences between video frames and a learned background model. These methods tend to be fast and often have the additional advantage of recovering foreground object silhouettes.
- University of Cranfield: Tracking using a combination of background separation and optical flow
- University of Edinburgh: Long term tracking from an overhead camera
- University of Leeds: Tracking system making use of Active Shape Models to separated foreground regions (EPSRC IMV project)
- University of Loughborough: Background separation with foreground silhouette extraction
- University of Reading: Further development of the Active Shape Model tracking system (EU ADVISOR project)
Appearance and motion
Appearance based pedestrian tracking methods work by identifying tracks of image regions that look like people. These methods can be advantageous in dense crowds where occlusions can cause other methods to fail, but are usually more computationally expensive than background separation based methods.
- University of Leeds: System for identifying pedestrians carrying objects
A number of recent projects have aimed to provide automatic interpretations of scenes by tracking pedestrians and analysing their interactions with one another and their environment. The result is a high-level description of how the people in the scene are behaving and of interactions with other objects such as cars.
- University of Leeds: Explaining human behaviour using psychological models of intentionality
Anomalous Trajectory Detection
A task closely related to scene description is that of anomalous trajectory detection, in which the aim is to identify unusual behaviour. The output from these systems can be used to notify security staff of increase the quality of a recording. The most common methods involve modelling ordinary trajectories and recognising when an individual does something that is not well represented by the model.
- University of Cambridge: Tracking system for counting people in dense crowds
- University of Leeds: Bicycle theft detection system
- University of Leeds: Player tracking in sports videos for the analysis of player movement
- University of Oxford: System for measuring the gaze directions of pedestrians in surveillance video
[Information collected by Ben Benfold; now maintained by Adrian F. Clark.]