[...]... Detection and Tracking Beyond the Visible Spectrum Kai Jüngling and Michael Arens Abstract One challenging field in computer vision is the automatic detection and tracking of objects in image sequences Promising performance of local features and local feature based object detection approaches in the visible spectrum encourage the application of the same principles to data beyond the visible spectrum Since these... Fraunhofer IOSB, Gutleuthausstrasse 1 76275 Ettlingen, Germany e-mail: kai.juengling@iosb.fraunhofer.de M Arens e-mail: michael.arens@iosb.fraunhofer.de R Hammoud et al (eds.), Machine Vision Beyond Visible Spectrum, Augmented Vision and Reality, 1, DOI: 10.1007/978-3-642-11568-4_1, © Springer-Verlag Berlin Heidelberg 2011 3 4 K Jüngling and M Arens allows for object component classification, i.e., body... advances in object detection in the visible spectrum [8, 13, 24, 32, 36, 38] encourage the application of these trainable, class-specific object detectors to thermal data Although person detection in infrared has its own advantages as well as disadvantages when compared to detection in the visible spectrum [12], most principles can be transferred from the visible spectrum to infrared While some techniques... color is used for data association in tracking In infrared data, person tracking is a more challenging problem than in the visible spectrum This is due to similar appearance of persons in infrared which makes identity maintenance in tracking much more difficult compared to the visible spectrum where rich texture and color is available to distinguish persons Especially on moving cameras, where the image... people appearing at different scales, visible from different viewpoints, and occluding each other The person tracking is evaluated in these three and two additional image sequences under two main aspects First, we show how tracking increases detection performance in the first three image sequences Second Local Feature Based Person Detection and Tracking Beyond the Visible Spectrum 7 we show how our approach... applications like situation assessment, the person detection results alone are not sufficient since they only provide a snapshot of a single point in Local Feature Based Person Detection and Tracking Beyond the Visible Spectrum 5 time For these higher level interpretation purposes, meaningful person trajectories have to be built by a tracking process To benefit from the advantages of the dedicated object detectors,... strength p(Ci | f k ) of an image feature fk , codebook entry Ci combination is determined by: p(Ci | f k ) = tsim − ρ( f k , Ci ) , tsim (1) Local Feature Based Person Detection and Tracking Beyond the Visible Spectrum 9 where ρ( f k , Ci ) is the euclidean distance in descriptor space Since all features with a distance above or equal tsim have been rejected before, p(Ci | f k ) is in range [0, 1]... representative and the image feature (see Eq 1) This means that a feature that is annotated with a body part and resides in a specific codebook Local Feature Based Person Detection and Tracking Beyond the Visible Spectrum 11 entry could contribute to a person hypothesis because the similarity between an image feature and the codebook representative is high enough (this similarity constraint is rather... its center is located inside the ground truth bounding box Only a single hypothesis is counted per ground truth object, all other hypotheses Local Feature Based Person Detection and Tracking Beyond the Visible Spectrum 13 in the same box are counted as false positive The overlapping criterion assesses object hypotheses using the ground truth and hypotheses bounding boxes The overlap between those is... reaches a recall of 0.5 with more than 5 false positives/image This rapid performance degradation is mainly due to inaccuracies in bounding boxes Local Feature Based Person Detection and Tracking Beyond the Visible Spectrum 15 Fig 5 Recall/false positive curves for a sequence 1, b sequence 2, and c sequence 3 Each chart contains four curves that refer to the different evaluation criteria BBI: inside bounding . genesis of this book on ‘ Machine Vision Beyond the Visible Spectrum ’ is the successful series of seven workshops on Object Tracking and Classification Beyond the Visible Spectrum (OTCBVS) held. Computer Vision and Pattern Recognition (CVPR) from 2004 through 2010. Machine Vision Beyond the Visible Spectrum requires processing data from many different types of sensors, including visible, . Arens e-mail: michael.arens@iosb.fraunhofer.de R. Hammoud et al. (eds.), Machine Vision Beyond Visible Spectrum, 3 Augmented Vision and Reality, 1, DOI: 10.1007/978-3-642-11568-4_1, © Springer-Verlag