Miguel Tavares Coimbra

Contact Details

Title: Research Student
Tel: Internal: [13] 5356
National: 020 7882 5356
International: +44 20 7882 5356
Fax: National: 020 7882 7997
International: +44 20 7882 7997
Email: miguel.coimbra@elec.qmul.ac.uk
Room: Eng. 402

Research Group:DSP & Multimedia

Research Area: Surveillance, Pedestrian Detection

There is a growing necessity for automatic surveillance systems since current ones are not efficient enough in crime prevention. Advances in technology are finally giving us a realistic possibility of creating such a system. We are developing a solution that takes advantages of specific properties of current technology to create an automatic pedestrian detection system that is both computationally inexpensive and efficient. The work is focused on surveillance cameras of metro stations and uses mpeg-2 video technology, more specifically motion vectors and DCT coefficients. These are specifically designed for compression and not motion estimation but our work proves that, with computationally fast filtering operations, an efficient detection system can be designed

Research Contribution

Currently, a smooth motion field is obtained from the MPEG-2 motion vectors using specific rules and filters. Our initial studies show that motion vectors alone are not enough for successful segmentation of pedestrians given their low resolution and noise. A confidence measure is required to generate motion fields that actually resemble the apparent flow of a video scene. Studying a popular optical flow method, namely Lucas Kanade, gave us insights on how to estimate confidence maps for our system. Average gradient magnitude (or edge strength), present in the AC[1] and AC[8] coefficients of the 8x8 DCT transform, is a very efficient measure of motion vector confidence and allows vast improvements in a pedestrian detection system by both decreasing noise and increasing field resolution. Our current work is studying the possibility of using DC coefficients (DC images) for low-resolution background subtraction, giving us extra information for a posterior segmentation step. This final segmentation will conclude our proposed research.

Publications

M. Coimbra, and M. Davies, "A Numerical Comparison of Compressed Domain Approximations to Optical Flow", accepted at WIAMIS 2004, Lisbon.

M. Coimbra, and M. Davies, "Segmentation of Moving Pedestrians within the Compressed Domain”, accepted at IEEE ICASSP 2004, Montreal.

M. Coimbra, and M. Davies, "A New Pedestrian Detection system using MPEG-2 Compressed Domain Information", in Proc. of IASTED International Conference on Visualization, Imaging, and Image Processing (VIIP 2003), Spain, 2003, pp. 598-602.

M. Coimbra, M. Davies and S. Velastin, "Pedestrian Detection using MPEG-2 Motion Vectors", in Proc. of WIAMIS, London, 2003, pp. 164-169.

M. Coimbra. "Crowd Monitoring using MPEG II Motion Vectors". First Year Report, Department of Electronic Engineering, King's College London, 2000

Exhibition

M. Coimbra and M. Davies, "Fast Optical Flow Estimation within the MPEG-2 Compressed Domain", in Challenge of Convergence 2003 Exhibition, University of Surrey, 7th July 2003.