MSc dissertation proposal 2016/2017
Scheduled Visual Multi-Target Tracking
-- Project description at fenix:
Objectives:
Pan-tilt-zoom cameras provide very high resolution at the points of interest, but cannot detect immediately all events in its field of view, i.e. lack omni-awareness. Nevertheless, one may reason that target tracking allows predicting motion, and prediction may save attention for detecting other targets. In other words, scheduling points of interest can be a mitigating factor, effectively allowing pan-tilt-zoom cameras to approximate omni-awareness.
Considering an urban environment observed by a number of pan-tilt-zoom cameras, the main objective consists in automating the cameras to work in a cooperative manner to track a number of targets (e.g. vehicles or people). This project is based in the research and development of two fundamental components: (i) pan-tilt-zoom camera modeling and (ii) developing an algorithm for assigning tasks to the cameras such that vision based multi-target tracking has optimal performance.
Requirements (grades, required courses, etc):
-
Place for conducting the work-proposal:
ISR / IST
Observations:
Previous works conducted at ISR/IST provide good starting points for the thesis. In particular a number of software prototypes already exist for target detection, tracking and camera scheduling.
-- More information:
Motivation:
Pan-tilt-zoom cameras provide very high resolution at the points of interest. More precisely, collecting images at maximum zoom with small pan and tilt incremental steps effectively allows observing a panoramic scene as using a gigapixel equivalent camera. However, pan-tilt-zoom cameras do not provide omni-awareness. Nevertheless, one can reason whether an efficient scheduling of points of interest can be a mitigating factor, much as the motion of the human eye allows tracking in high (foveal) resolution on the surface of an egosphere. This is the central aspect of this project.
The scenario for this work is a urban environment observed by a number of pan-tilt-zoom cameras. The cameras work in a cooperative manner to track a number of targets (e.g. vehicles or people).
This project is based in the research and development of two fundamental components: (i) camera modeling and (ii) developing an algorithm for assigning tasks to the cameras such that vision based multi-target tracking has optimal performance.
Detailed Description:
The main tasks of the project are the following:
1. Acquisition of Panoramic Video Datasets - Build video datasets with panoramic views that allow simulating up to some extent the pan and tilt movements of PTZ cameras.
2. Camera Modeling - We consider in essence pan-tilt-zoom and omnidirectional cameras, and develop geometric and radiometric models for these cameras. In a first approach, calibration is achieved using calibration patterns and, in a second instance, in natural features of a scene. Calibration may involve both types of cameras working cooperatively in the same network.
3. Active Camera Management for Multitarget Tracking - Assess the capability of Switched Linear Dynamic Models (SLDM) for the online prediction of tracking targets dynamics, and then develop an active camera management policy based in the information theory framework.
References:
[Micheloni10] "Video Analysis in Pan-Tilt-Zoom Camera Networks: From master-slave to cooperative smart cameras", Christian Micheloni, Bernhard Rinner, and Gian Luca Foresti, IEEE Signal Processing Magazine September 2010
[Galego11] Ricardo Galego, "Geometric and Radiometric Calibration for Pan-Tilt Surveillance Cameras", MSc dissertation, MEEC / IST, 2011
[Leite11] Diogo Leite, "Target Tracking with Pan-Tilt-Zoom Cameras", MSc dissertation, MEEC / IST, 2011
[Castanheira13] Tiago Castanheira, "Multi-tasking Smart Cameras", MSc dissertation, MEEC / IST, 2013
[Silva14] Pedro Silva, "Vision Based Multi-Target Detection and Tracking", MSc dissertation, MEEC / IST, 2014
[Marques15] Tiago Marques, "Cooperating Smart Cameras", MSc dissertation, MEEC / IST, 2015
Expected Results:
The student will develop skills in computer vision and signal filtering. The final result must consist of in the demonstration of an application illustrating the successful achievement of the theoretical and practical aspects of the work.
More MSc dissertation proposals on Computer and Robot Vision in:
http://omni.isr.ist.utl.pt/~jaghttp://omni.isr.ist.utl.pt/~jag