Towards Robust Visual Pattern Recognition: from Emotion Recognition to Human Activity Monitoring

Prof. Dr. Salim Bouzerdoum, University of Wollongong, Australia

1 Jul 2010, 17:00; Location: S4|10-1

One of the challenges of visual pattern recognition is robustness to photometric and geometric distortions that occur in uncontrolled natural environments. To this end, we have proposed recently a new paradigm for visual pattern recognition which combines feature extraction and classification in one hierarchical image classification architecture. The feature extraction and classification stages are designed concurrently and are optimized with respect to one another. The feature extraction stage comprises directional and adaptive two-dimensional filters, inspired by the early processing of mammalian vision; they are designed to be tolerant to contrast and illumination variations. In this seminar, we will outline a number of applications where the proposed architecture has been successfully used, namely face detection, gender recognition, facial expression recognition, and human motion classification based on radar miro-Doppler signatures. In the latter, the radar Doppler signal is depicted as an image in the time-frequency domain, which is then processed by the proposed image classification architecture. The feature extraction stage implements steps that, in essence, act on revealing the distinctive micro-Doppler features of the human walking and, as such, allows effective discrimination of various types of human motions characterized by the nature of arm swings: free arm-motion, partial arm motion, and No arm-motion. All three categories are considered important for police and law enforcement, especially when humans are inside buildings and enclosed structures, or monitored while moving in city canyons and street corners.

More recently, we have employed Compressed Sensing (CS), also known as compressive sampling, in machine learning. CS is an emerging powerful signal processing paradigm that can perform data acquisition and compression simultaneously. It is an effective approach for significantly reducing the number of data measurements without altering the signal quality and reliability. If time permits, we will also describe the application of compressive sensing to visual pattern recognition, where CS is used for simultaneous feature selection and parameter estimation. The seminar will conclude with a demonstration of gender recognition from real videos.

Category: CE Seminar


Technische Universität Darmstadt

Graduate School CE
Dolivostraße 15
D-64293 Darmstadt

Phone+49 6151/16-24401
Fax -24404

to assistants' office

Open BSc/MSc Theses

Show a list of open BSc/MSc topics at GSC CE.

 Print |  Impressum |  Sitemap |  Search |  Contact |  Privacy Policy
zum Seitenanfangzum Seitenanfang