Contents I. Introduction
The application of Visual Servoing (VS) techniques has gained significant attention for various robotic tasks [1]. VS involves controlling robot motion based on visual feedback from a camera, enabling tasks such as object tracking, manipulation, and navigation. In complex environment, such as natural ones, accurate perception is crucial and the choice of an appropriate camera system plays a vital role. Multi-spectral cameras, a type of imaging system, have emerged as promising tools for perception in challenging environments. Unlike traditional RGB cameras that capture images in three color channels (red, green, and blue), multispectral cameras are capable of capturing images across a wider range of wavelengths, composed from 5 to 15 bands. This provides valuable spectral information beyond the visible spectrum, offering significant advantages in natural environments where comprehensive perception is essential. However, deploying VS techniques using multispectral cameras in real-time natural scenarios poses a key challenge: processing the large amount of multispectral data while maintaining high temporal performance. Addressing these processing and interpretation challenges is crucial to effectively utilizing the spectral information provided by multispectral cameras. Dimensionality reduction techniques play a crucial role in this regard, as they enable the extraction of key information while reducing the computational complexity of the analysis [2]. In hyperspectral imaging (HSI), a type of multispectral imaging with extremely high spectral resolution, various approach such as feature extraction [3] and band selection [4] have been explored to enable accurate and efficient classification of HSI images. Feature extraction combines multiple bands to create a compressed but informative subset, though it can be computationally demanding and may hinder interpretability due to the creation of new features. Conversely, band selection techniques choose a subset of bands from the multispectral data, preserving spectral meaning and reducing computational burden [4]. Various unsupervised and supervised methods have been proposed in the literature for band selection in hyperspectral data analysis. In [5], the Multigraph Determinantal Point Process (MDPP) approach efficiently finds the optimal band subset, while Principal Component Analysis (PCA) prioritizes the energy of variances in band images [6]. Another method, Sparse Nonnegative Matrix Factorization (SNFM), is used to solve the band selection problem [7]. Similarly, minimizing the correlation of selected bands has been used to identify the best subset [8]. In [9], an alternative approach involves clustering original bands and selecting representative bands from each cluster, proposed as an automatic band selection method. Furthermore, hybrid methods that combine the advantages of previous techniques are introduced [10]. In the context of Multi-Spectral Visual Servoing (MSVS), researchers have explored integrating VS techniques with RGB cameras to enhance robotic vision and control. One notable approach among these techniques is the colorimetry-based approach [11]. The authors investigated the utilization of visual features derived from the individual R, G, and B components of the image. Their objective was to evaluate the suitability of employing features derived from different linear combinations of two or all three components for visual servoing tasks. The results of their study illustrated the promising potential of this approach for enhancing VS applications. Another significant contribution comes from [12]. In this work, authors presented innovative parametric models and optimization methods for the robust and direct registration of color images. Their methodology involved representing a color image obtained by stacking n channels multiplied by corresponding surfaces to achieve precise alignment with a reference image. However, their methodology focused on the use of only the readily available RGB channels. In a different line of inquiry, researchers delved into frequency domain [13]. Rather than analyzing the image itself in the spatial domain, they explored its transformation into the frequency domain. While these approaches have yielded valuable insights, it’s important to note that they are limited to RGB images and do not leverage the additional spectral information provided by multispectral cameras. The primary objective of this work is to bridge this gap and extend the applicability of visual servoing techniques to multispectral cameras with n bands, thereby harnessing the full potential of multispectral imaging in the context of robotic control. To the authors’ knowledge, this is the first work investigating Visual Servoing for multispectral cameras’ data. This work focuses on using the multispectral information to improve the performance of visual servoing tasks, rather than emphasizing the control aspect.
