pets.jpg The tracking system used in the "PeopleVision" project and the IBM Smart Surveillance Solution. privacytarget.png Automatic video understanding technologies can help protect privacy in video surveillance systems. This project seeks to hide privacy-intrusive information in surveillance video. heatmap.jpg Retail is a major application area of automatic video analytics. I organised a special session at AVSS 2007 on the topic. This project created a system for returns fraud prevention. homographytracks.png A multi-camera system to automatically track pedestrians with an active camera. The calibration is learned automatically from observing people walking through the scene. particlefilter.png Several 3D person tracking algorithms were developed as part of the EU CHIL (Computers in the Human Interaction Loop) project, using 2D blob trackers, face trackers and this particle filter tracker. articedgefit.png A system to track the person's limbs with a 3D graphical model using 2 or more cameras. Running in real time and with moving cameras to cover a wide area. lips.jpg Facial feature locations from the face tracker (below) were used to generate visual features in the first audio-visual, large vocabulary continuous speech recognition system. The system was also used to do speech activity detection for "visual push to talk" facialfeatures.png The video face detection and tracking algorithms (below) were extended to a mutliplatform face recognition system that worked on live video, still images or broadcast video. skintone.png A system for real-time face detection in video that also localized the facial features for recognition, expression understanding and speech recognition. fpclassifydecisiontree.png Fingerprint classification using a combination of methods (including HMM and decision trees) that gave the best classification results published at the time on the standard NIST database. distortedfingerprint.png A novel way of removing the distortion in fingerprints that improved recognition performance. durationmodel.png Handwriting recognition for tablet computers. The system became part of the IBM ThinkScribe and TransNote products. offlinehandwriting.png PhD Thesis on "Off-line handwriting recognition using recurrent neural networks" - a neural-network / HMM hybrid. Incorporating forward-backward retraining of the networks, large vocabulary language modelling and out-of-vocabulary word modelling. LIMSIModel.png Conducted at LIMSI, at the Universite de Paris XI, as part of a European project.