Multi-Cue Pedestrian Recognition
Stefan Munder · MADOC (University of Mannheim) · 2007
This thesis addresses the problem of detecting complex, deformable objects in an arbitrary, cluttered environment in sequences of video images. Often, no single best technique exists for such a challenging problem, as different approaches possess different characteristics with regard to detection accuracy, processing speed, or the kind of errors made. Therefore, multi-cue approaches are pursued in this thesis. By combining multiple detection methods, each utilizing a different aspect of the video images, we seek to gain detection accuracy, robustness, and computational efficiency. The first part of this thesis deals with texture classification. In a comparative study, various combinations of feature extraction and classification methods, some of which novel, are examined with respect to classification performance and processing speed, and the relation to the training sample size is analyzed. The integration of shape matching and texture classification is investigated. A pose-specific mixture-of-experts architecture is proposed, where shape matching yields a probabilistic assignment of a texture pattern to a set of distinct pose clusters, each handled by a specialized texture classifier, the local expert. The reduced appearance variability that each local expert needs to cope with leads to improved classification performance. A slight further performance gain could be achieved by shape normalization. The second multi-cue approach deals with cascade systems that employ a sequence of fast-to-complex system modules in order to gain computational efficiency. Three optimization techniques are examined that adjust system parameters so as to optimize the three performance measures detection rate, false positive rate, and processing cost. A combined application of two techniques, a novel fast sequential optimization scheme based on ROC (receiver operating characteristics) frontier following, followed by an iterative gradient descent optimization method, is found to work best. The third method investigated is a Bayesian combination of multiple visual cues. An integrated object detection and tracking framework based on particle filtering is presented. A novel object representation combines mixture models of shape and texture, the former based on a generative point distribution model, the latter on discriminative texture classifiers. The associated observation density function integrates the three visual cues shape, texture, and depth. All methods are extensively evaluated on the problem of detecting pedestrians in urban environment from within a moving vehicle. Large data sets consisting of tens of thousands of video images have been recorded in order to obtain statistically meaningful results.