This paper presents a novel algorithm for pedestrian detection in still images using covariance matrices as object descriptors. The key contribution is a classification method on Riemannian manifolds, which leverages the geometry of the space of symmetric positive definite matrices. The algorithm is tested on the INRIA and DaimlerChrysler pedestrian datasets, achieving superior detection rates compared to previous approaches. The algorithm uses covariance descriptors derived from feature images, which are computed using integral images for efficient covariance calculation. These descriptors are then mapped to a Riemannian manifold, where classification is performed using a LogitBoost algorithm adapted to the manifold structure. The method involves computing the mean of points on the manifold and using it to define tangent spaces for learning weak classifiers. The algorithm employs a cascade of LogitBoost classifiers, where each level is trained to correctly detect a high percentage of positive examples while rejecting a significant portion of negative examples. A greedy feature selection method is used to focus on important subwindows, resulting in a sparse set of classifiers. The method is evaluated on the INRIA dataset, showing significant improvements in detection performance compared to existing methods. The results demonstrate that the proposed approach achieves lower miss rates at various false positive per window (FPPW) thresholds, with the best performance at 10⁻⁵ FPPW. The method is efficient and effective, with the potential for further improvements through the inclusion of more negative examples.This paper presents a novel algorithm for pedestrian detection in still images using covariance matrices as object descriptors. The key contribution is a classification method on Riemannian manifolds, which leverages the geometry of the space of symmetric positive definite matrices. The algorithm is tested on the INRIA and DaimlerChrysler pedestrian datasets, achieving superior detection rates compared to previous approaches. The algorithm uses covariance descriptors derived from feature images, which are computed using integral images for efficient covariance calculation. These descriptors are then mapped to a Riemannian manifold, where classification is performed using a LogitBoost algorithm adapted to the manifold structure. The method involves computing the mean of points on the manifold and using it to define tangent spaces for learning weak classifiers. The algorithm employs a cascade of LogitBoost classifiers, where each level is trained to correctly detect a high percentage of positive examples while rejecting a significant portion of negative examples. A greedy feature selection method is used to focus on important subwindows, resulting in a sparse set of classifiers. The method is evaluated on the INRIA dataset, showing significant improvements in detection performance compared to existing methods. The results demonstrate that the proposed approach achieves lower miss rates at various false positive per window (FPPW) thresholds, with the best performance at 10⁻⁵ FPPW. The method is efficient and effective, with the potential for further improvements through the inclusion of more negative examples.