• Supervised and unsupervised learning from imbalanced datasets for assistance in movement of persons with low vision
The Client : Javna agencija za raziskovalno dejavnost RS
Project type: Bilateral projects
Project duration: 2014 - 2015
  • Description
Recently, special attention of machine learning community has been given to realworld classification problems with imbalanced class distribution or with classes where misclassifications cost varies significantly. For example, in medical diagnosis typically a rather small number of patients have some critical diagnoses (illnesses), however an incorrect diagnosis can have fatal consequences in cases where an ill patient is proclaimed healthy. The classical classification approaches (when classes are approximately balanced or when we have equal misclassification costs for all classes) are not applicable in such cases as they tend to maximize the classification accuracy and therefore the instances from minority class are treated by those approaches as noisy instances. Yet, what we want to achieve is to minimize the number of misclassified cases from the minority class while keeping the relatively high overall classification accuracy (or we want to minimize the overall misclassification cost). For that purpose a different approaches were developed which may either include in the algorithm itself the misclassification cost, or the modifications of the algorithm in order to focus more on cases from the minority class, or to use various techniques to modify the training set appropriately, such as oversampling of the minority class cases or undersampling of the majority class cases. Also, hybrid approaches are also becoming popular. One of the civilization responsibilities is the need for increasing the quality of life for blind and weaksighted people using the modern technology. Daily activities which for healthy subjects seem to be trivial, such as finding the required destination, using various objects, safe navigation through complex paths, recognizing the local environment and an appropriate orientation in it, contextual comprehension of the environment etc. are much demanding and often unsolvable tasks for blind or weaksighted subject. Most of the approaches to solve those problems are based on GPS technology combined with inertial sensors and cameras. The most common problems with those approaches are the following. The GPS technology has a limited accuracy in urban areas, especially in the areas close to building due to multiple signal reflections. This may cause that a blind person assumes he or she is located at the sidewalk while standing in the middle of the road. This problem can be reduced using inertia sensors, however using the approach of identifying a small number of deviations of two different systems, which transforms the original problem onto the problem of classification of data with imbalanced classes. Cognitive understanding of environment, recognition of objects and avoiding the obstacles is relatively successfully solved using approaches that combine cameras, visual markers or RFID sensors, and WiFi receivers, however, this is applicable only in controlled environments and outside those the usability is much limited. A basic limitation is information complexity of unmarked environments which can be avoided by providing a small number of useful data with respect to the current position, orientation and activity that a blind or weaksighted subject wants to perform (such as crossing the road). This problem can also be transformed into a problem of dealing with imbalanced data. Therefore, there is a need to try to use the existing methods as well as to develop new algorithms for supervised as well as unsupervised machine learning for this kind of problems. This becomes even more interesting if the system includes the data from GIS. In this context we emphasize the movement through usual paths with pointing to some important objects on the path (such as a restaurant, library, store etc.). However, this is even more important on unknown paths where at the beginning only starting and end point are known. We expect that the use of machine learning methods, i.e. supervised and unsupervised learning on imbalanced datasets with combination of GPS, inertia sensors, cameras and an audiorendering system, will successfully solve the described problems when using the existing approaches for helping during movement of blind and weaksighted people. Therefore, the goal of the proposed project is through the development of appropriate machine learning algorithms prepare the basis for future development of a modular mobile navigation system which will enable the blind and weaksighted people to autonomously, safely and accurate movement in known and unknown environments. The basic requirements for such a system are realtime operation in dynamic environments without the need to provide the external infrastructure in advance, and with an emphasis on high level reliability. This way the preconditions for better socialization and higher lifequality for blind and weaksighted people will be achieved. Due to interdisciplinarity of the project, which shall use the methods of machine learning, cognitive modeling, and biodynamics of movement, there is a need of collaboration of the Laboratory for biomechanics, automatics and systems from Faculty of electrical engineering, machinery and shipbuilding, University of Split, Croatia (where for more than 15 years a research team works on problems, related to Biomechanics of human movements, control and rehabilitation), and our Laboratory for Cognitive Modeling, Faculty of Computer and Information Science, University of Ljubljana, Slovenia.