Multiscale Image Analysis for the Visual Servoing Problem

ROUAG, Fadi Elislam (2025) Multiscale Image Analysis for the Visual Servoing Problem. Doctoral thesis, Faculté des Sciences et de la technologie.

Text Thesis.pdf Download (4MB)

Abstract

Visual servoing enables the control of robotic systems directly from visual feedback and has become a fundamental technique in vision-based robotics. Among existing approaches, Photometric Visual Servoing (PVS) relies on raw image intensities as visual features, avoiding explicit feature extraction and matching. While PVS offers high accuracy and a compact formulation, its performance is strongly affected by illumination variations, limiting its robustness in realistic operating conditions. This thesis proposes an enhanced photometric visual servoing framework based on multiresolution image analysis, referred to as Lifting Wavelet Transform-based Photometric Visual Servoing (LWT-PVS). The method integrates the lifting wavelet transform into the PVS formulation to exploit spatial-frequency image representations while preserving the direct nature of photometric control. By selecting appropriate wavelet sub-bands, the proposed approach reduces sensitivity to illumination changes and photometric disturbances, while maintaining stable convergence properties. A unified mathematical formulation of LWT-PVS is developed, including the definition of photometric visual features, error functions, interaction matrices, and discrete-time control laws. The complete control architecture is implemented using the ViSP (Visual Servoing Platform) simulator, ensuring consistency with established visual servoing tools. Extensive simulation studies are conducted for multiple scenarios involving different initial camera poses and both nominal and varying illumination conditions. The performance of PVS and LWT-PVS is evaluated in terms of error convergence, camera velocity behavior, and robustness to lighting variations. The results show that LWT-PVS significantly improves robustness to illumination changes compared to classical PVS, while preserving convergence accuracy and stability. The contributions of this work demonstrate the relevance of multiresolution representations for photometric visual servoing and provide a robust control framework suitable for real-world robotic applications operating under challenging illumination conditions.

Actions (login required)

View Item