Smart Systems: Bringing Together Computer Vision, Sensor Networks and Machine Learning (SmartSys) Session 3

Time and Date: 16:30 - 18:10 on 13th June 2019

Room: 2.26

Chair: João Rodrigues

157 Towards Low-Cost Indoor Localisation Using a Multi-camera System [abstract]
Abstract: Indoor localisation is a fundamental problem in robotics, which has been the subject of several research works over the last few years. Indeed, while solutions based on fusion of inertial and global navigation satellite system (GNSS) measurements have proved their efficiency in outdoor environments, indoor localisation remains an open research problem. Although commercial motion tracking systems can offer very accurate position estimation, their high cost cannot be afforded by all research laboratories. This paper presents an indoor localisation solution based on a multi-camera setup. The proposed system rely on low-cost sensors, which makes it very affordable compared to commercial motion-tracking systems. We show through the experiments conducted that the proposed approach, although being cheap, can provide real-time position measurements with an error of less than 2 cm up to a distance of 2m.
Oualid Araar, Bouhired Saadi, Sami Moussiou and Ali Laggoune
169 A New Shape Descriptor and Segmentation Algorithm for Automated Classifying of Multiple-Morphological Filamentous Algae [abstract]
Abstract: In our previous work on automated microalgae classification system we proposed the multi-resolution image segmentation that can handle well with unclear boundary of algae bodies and noisy background, since an image segmentation is the most important preprocessing step in object classification and recognition. The previously proposed approach was able to classify twelve genera of microalgae successfully; however, when we extended it to work with new genera of filamentous algae, new challenging problems were encountered. These difficulties arise due to a variety of the forms of filamentous algae, which complicates both image segmentation and classification processes, resulting in substantial degradation of classification accuracy. Thus, in this work we propose a modified version of our multi-resolution segmentation algorithm by combining them in such a way that the strengths of both algorithms complement each other's weaknesses. We also propose a new skeleton-based shape descriptor to alleviate an ambiguity caused by multiple morphologies of filamentous forms of algae in classification process. Effectiveness of the two proposed approaches are evaluated on five genera of filamentous microalgae. SMO is used as a classifier. Experimental result of 91.30% classification accuracy demonstrates a significant improvement of our proposed approaches.
Saowanee Iamsiri, Nuttha Sanevas, Chakrit Watcharopas and Pakaket Wattuya
393 Application of hierarchical clustering for object tracking with a Dynamic Vision Sensor [abstract]
Abstract: Monitoring public space with imaging sensors to perform a object- or person-tracking is often associated with privacy concerns. We present a Dynamic Vision Sensor (DVS) based approach to achieve this tracking, that does not require the creation of conventional grey- or colorimages. These Dynamic Vision Sensors produce an event-stream of information, which only includes the changes in the scene. The presented approach for tracking consider the scenario of fixed mounted sensors. The method is based on clustering events and tracing the resulting cluster centers to accomplish the object tracking. We show the usability of this approach with a first proof-of-concept test.
Tobias Bolten, Regina Pohle-Fröhlich and Klaus D. Tönnies
473 Binarization of Degraded Document Images with Generalized Gaussian Distribution [abstract]
Abstract: One of the most crucial steps of preprocessing of document images subjected to further text recognition is their binarization, which influences significantly obtained OCR results. Since for degrades images, particularly historical documents, classical global and local thresholding methods may be inappropriate, a challenging task if their binarization is still up-to-date. In the paper a novel approach to the use of Generalized Gaussian Distribution for this purpose is presented. Assuming the presence of distortions, which may be modelled using the Gaussian noise distribution, in historical document images, a significant similarity of their histograms to those obtained for binary images corrupted by Gaussian noise may be observed. Therefore, extracting the parameters of Generalized Gaussian Distribution, distortions may be modelled and removed, enhancing the quality of input data for further thresholding and text recognition. Due to relatively long processing time, its reduction using the Monte Carlo method is proposed as well. The proposed algorithm has been verified using well-known DIBCO datasets leading to very promising results of binarization.
Robert Krupiński, Piotr Lech, Mateusz Tecław and Krzysztof Okarma
256 Nonlinear dimensionality reduction in texture classication: is manifold learning better than PCA? [abstract]
Abstract: This paper presents a comparative analysis of algorithms belonging to manifold learning and linear dimensionality reduction. We first combine classical texture image descriptors, namely Gray-Level Co-occurrence Matrix features, Haralick features, Histogram of Oriented Gradients features and Local Binary Patterns to characterize and discriminate textures. For patches extracted from several texture images, we perform a concatenation of the image descriptors. Using four algorithms to wit PCA, LLE, ISOMAP and Lap. Eig., dimensionality reduction is achieved. The resulting learned features are then used to train four different classifiers: k-nearest neighbors, naive Bayes, decision tree and multilayer perceptron. Finally, the non-parametric statistical hypothesis test, Wilcoxon signed-rank test, is used to figure out whether or not manifold learning algorithms perform better than PCA. Computational experiments were conducted using the Outex and Salzburg datasets and the obtained results show that among twelve comparisons that were carried out, PCA presented better results than ISOMAP, LLE and Lap. Eig. in tree comparisons. The remainder nine comparisons did not presented significant differences, indicating that in the presence of huge collections of texture images (bigger databases) the combination of image feature descriptors or patches extracted directly from raw image data and manifold learning techniques is potentially able to improve texture classification.
Cedrick Bamba and Alexandre Levada