6533b85afe1ef96bd12b9318

RESEARCH PRODUCT

Multi-Scale Feature Extraction for Vehicle Detection Using Phis-Lbp

subject

description

International audience; Multi-resolutionobjectdetectionfacesseveraldrawbacksincludingitshighdimensionalityproducedby a richer image representation in different channels or scales. In this paper, we propose a robust and lightweight multi-resolution method for vehicle detection using local binary patterns (LBP) as channel feature. Algorithm acceleration is done using LBP histograms instead of multi-scale feature maps and by extrapolating nearby scales to avoid computing each scale. We produce a feature descriptor capable of reaching a similar precision to other computationally more complex algorithms but reducing its size from 10 to 800 times. Finally, experiments show that our method can obtain accurate and considerably faster performance than state-of-the-art methods on vehicles datasets.