Robust object detection with efficient features and effective classifiers

Abstract

This thesis contains three main novel contributions that advance the state of the art in object detection. The first contribution focuses on a realtime pedestrian detector using a combination of Haar-like features and covariance features. Unlike the original work of Tuzel et al., where the feature selection and weak classifier training are performed on the Riemannian manifold, weak classifiers are trained in the Euclidean space for faster computation. To this end, a novel approach based on AdaBoost with weighted Fisher Linear Discriminant Analysis (FLDA) based weak classifiers is designed. To further accelerate the detection, a faster strategy, known as a multiple-layer boosting with heterogeneous features, is adopted to exploit the efficiency of Haar-like features and the discriminative power of covariance features. Experimental results show that by combining Haar-like and covariance features, the efficiency of final detectors improves by an order of magnitude with a slight drop in the detection performance.

The second contribution reveals the drawback of commonly used AdaBoost and a more effective approach, termed Boosted Greedy Sparse Linear Discriminant Analysis (BGSLDA), is proposed. BGSLDA exploits a class-separability criterion of LDA and a sample re-weighting property of boosting. Experimental results demonstrate an improvement in the detection performance compared to the original AdaBoost framework. This new finding provides a significant opportunity to argue that AdaBoost and its variants are not the only method that can achieve a high classification accuracy in a high dimensional problem, such as object detection.

The last contribution points out the drawback of offline object detection frameworks and an efficient online framework is proposed. Unlike many existing online boosting algorithms, which apply exponential or logistic loss, the proposed online algorithm makes use of LDA’s learning criterion that not only aims to maximize the class-separation criterion but also incorporates the asymmetrical property of training data distributions. The new approach provides a better alternative to online boosting algorithms in the context of training a visual object detector. Experimental results on handwritten digits and face data sets show that object detection tasks benefit significantly when trained in an online manner.

Finally, this thesis concludes with a discussion and future works, which explore potential advances in the learning of feature descriptors, cascade classifiers as well as online object detectors.