Research subject: Robust face tracking for incremental learning in video surveillance
Research Master's degree. Bachelor in computer engineering, software engineering, computer science, or equivalent. A good mastering of programming languages is required. In particular, C++ and MATLAB. Experience in image processing or computer vision is required. Experience with tracking algorithms or in facial recognition is an asset.
The LITIV laboratory is looking for a student, having an education in computer and software engineering, or in computer science, to participate in a research project at the Doctorate level in parternership with professors Éric Granger and Robert Sabourin of the LIVIA laboratory of École de Technologie Supérieure, and with Dr. Gorodnichy of the Canada border services agency.
The project aims at designing methods to identify individuals from a watch list using facial recognition using tracking and a system with multiple classifiers. Indeed, the rapid and covert recognition of individuals of interest from video streams collected over distributed video surveillance cameras remains a challenging problem, especially in dense and moving crowds. In practice, the performance of state-of-the-art systems applied to video-based face recognition typically declines because of the complex environments that change during operations. Thus, robust tracking methods are required. In addition, faces captured in video streams are matched against the facial model of individuals, each one designed a priori using a limited number of reference samples collected during enrollment. To solve this problem, systems for face recognition in video surveillance can exploit spatio-temporal information and learn from new reference video streams to sustain a high level of performance.
This project focuses on the localisation and tracking of a region of interest corresponding to a face using an adaptive model to provide data for the facial recognition. The objectives of this project are to:
- Detect a face;
- Design an appearance model using multiple features to follow a face robustly;
- Develop a strategy to evaluate tracking quality and the quality of the various features to adapt dynamically the model for the tracking conditions (lighting, occlusions, etc.);
- Develop a method to estimate the quality of the region of interest;
- Use tracking quality indicators to dynamically select the most suitable subset of detectors to form an ensemble of detectors during operation.
September 2012 or january 2013