Influence of image compression on cascade classifier components

Abstract

Bandwidth restrictions and increasing data volumes in the transmission path of automotive driver assistance systems make video compression unavoidable for future applications. Conventional image compression algorithms are solely tuned for optimal human perception. This paper studies the effect on features used in discriminative cascade classifiers for nighttime pedestrian desection, namely Haar wavelet features, Edge Orientation Histogram features and Standard deviation features. The induced error is modeled and evaluated for these feature classes. By approximating the noise on specific image feature instances, a re-adaption of the decision boundaries is possible. Knowing about the sensitivity of specific feature classes allows selecting a robust set of features prior to classifier training.