Relative Altitude Estimation Using Omnidirectional Imaging and Holistic Descriptors

Abstract

Currently, many tasks can be carried out using mobile robots. These robots must be able to estimate their position in the environment to plan their actions correctly. Omnidirectional vision sensors constitute a robust choice to solve this problem, since they provide the robot with complete information from the environment where it moves. The use of global appearance or holistic methods along with omnidirectional images constitutes a robust approach to estimate the robot position when its movement is restricted to the ground plane. However, in some applications, the robot changes its altitude with respect to this plane, and this altitude must be estimated. This work focuses on this problem. A method based on the use of holistic descriptors is proposed to estimate the relative altitude of the robot when it moves upwards or downwards. This descriptor is constructed from the Radon transform of omnidirectional images captured by a catadioptric vision system. To estimate the altitude, the descriptor of the image captured from the current position is compared with the descriptor of the reference image, previously built. The framework is based on the use of phase correlation to calculate relative orientation and a method based on the compression-expansion of the columns of the holistic descriptor to estimate relative height. Only an omnidirectional vision sensor and image processing techniques are used to solve these problems. This approach has been tested using different sets of images captured both indoors and outdoors under realistic working conditions. The experimental results prove the validity of the method even in the presence of noise or occlusions.

The Role of Global Appearance of Omnidirectional Images in Relative Distance and Orientation Retrieval

“…Initially, it was used in different computer vision applications as a geometric shape descriptor, as in [ 66 , 67 ]. More recently, the Radon transform (RT) has been adapted to describe globally omnidirectional images and its performance was tested in [ 41 ], where descriptors based on the RT were used to solve the image retrieval problem, and in [ 23 ], where these descriptors were used to estimate relative altitude from images. The main advantage of this descriptor is that it can be calculated with raw omnidirectional images, as captured by the vision system (with no panoramic transformation).…”

“…The first family of methods consists in detecting some outstanding landmarks or regions and describing them using any algorithm that provides some invariance against transformations, such as SIFT [ 10 ], SURF [ 11 ], BRIEF [ 12 ], BRISK [ 13 ], ORB [ 14 ], FREAK [ 15 ] and LDB [ 16 ]. The second family consists of working with each scene as a whole, trying to build a unique descriptor per image that collects information on its global structure, using some approaches such as Principal Components Analysis [ 17 ], discrete Fourier transform [ 18 ], banks of Gabor filters [ 19 ], color histograms [ 20 , 21 ], directly subsampled versions of the original image [ 22 ] or Radon transform [ 23 ].…”

Creating Incremental Models of Indoor Environments through Omnidirectional Imaging

“…It was also used by Zhou et al [34] to solve the localization through matching the robot's current view with the best key-frame in the database. In addition, some other works define other global appearance description methods such as the Discrete Fourier Transform [35], the alternative used by Paya et al [30] to perform map creation tasks, or Radon Transform [36], used in [37] to find the nearest neighbour in a dense map previously created. Román et al [38] develop a comparison among these global appearance descriptors performing a mobile robot localization in a real environment under changing lighting conditions [38].…”

Point-Line Visual Stereo SLAM Using EDlines and PL-BoW

“…Berenguer et al proposed a method to estimate the relative attitude by using Holistic descriptors of the omnidirectional image. This method solves the problem that the omnidirectional image can not deal with the height change of mobile robots, but it requires further research to apply in estimating movements with six degrees of freedom [17]. Inspired by the Scale Invariant Feature Transform (SIFT) algorithm, Li et al proposed the scale-invariant mean-standard deviation LSD (SMLSD) to extract line features faster without sacrificing detection accuracy [18].…”