Learning-based vision and its application to autonomous indoor navigation

Adaptation is critical to autonomous navigation of mobile robots. Many adaptive mechanisms have been implemented, ranging from simple color thresholding to complicated learning with artificial neural networks (ANN). The major focus of this thesis lies in machine learning for vision-based navigation. Two well known vision-based navigation systems are ALVINN and ROBIN developed by Carnegie-Mellon University and University of Maryland, respectively. ALVINN uses a two-layer feedforward neural network while ROBIN relies on a radial basis function network (RBFN). Although current ANN-based methods have achieved great success in vision-based navigation, they have two major disadvantages: (1) Local minimum problem: The training of either multilayer perceptron or radial basis function network can get stuck at poor local minimums. (2) The flexibility problem: After the system has been trained in certain road conditions, it is hard to make the system adapt to new road conditions while retaining good performance for those road conditions that have already been learned. Sometimes this is termed a “memory loss” problem. As part of our SHOSLIF (Self-organizing Hierarchical Optimal Subspace Learning and Inference Framework) effort, SHOSLIF-N (SHOSLIF for Navigation) treats vision-based navigation as a content-based retrieval problem. Three major components of SHOSLIF-N are: (1) Automatic feature derivation: Instead of starting with random initial weights, the system employs either principle component analysis or linear discriminant analysis to derive features which are best suited for navigation tasks. (2) Nonparametric recursive partitioning regression, which is more flexible than global parametric regression used in either ALVINN or ROBIN, is employed in direct input-to-output mapping. Nonparametric recursive partitioning regression is realized with a recursive partition tree (RPT). (3) Self-organizing mechanism. (4) Low computational complexity: the recursive partition tree has a logarithmic retrieval complexity and can be used to address the complexity issue in learning a large number of scenes. For a binary RPT, only the most dominant eigenvector of principle component analysis or linear discriminant analysis is needed for further partitioning of each inner node. This leads to an efficient online incremental learning algorithm: the system learns or rejects a learning sample “on-the-fly” with real time response. Similar to ALVINN and ROBIN, the basic SHOSLIF-N maps a single-framed retinal input into an output steering signal. The system was successfully tested but exhibited limited capability in handling more complicated situations. When the number of different turns or corners was increased to a certain extent, the system sometimes failed to make the turn. One way to tackle this problem is to incorporate state information; that indicates the relative position between the robot and the oncoming corner or intersection, into the system. Therefore, state-based SHOSLIF-N, a system that incorporates states and utilizes a simple yet efficient visual attention mechanism which is helpful in determining the correct state transitions, is proposed and tested. With a set of fewer than 300 learning samples, state-based SHOSLIF-N has been successfully tested in indoor navigation on the 2nd and 3rd floors of our Engineering Building. Using a SUN Sparc-I and a framegrabber, both online incremental learning and autonomous navigation were done in real-time. Comparative study with two ANN-based methods has shown the advantages of the system: faster learning and better performance for the tasks tested.