Mamoru Minami

Evolutionary Scene Recognition and Simultaneous Position/Orientation Detection

This paper presents a new method of scene recognition for manipulator real-time visual servoing, which utilizes a hybrid genetic algorithm in combination with a model shaping a target of known shape, and the unprocessed gray-scale image of a scene, termed here as raw-image. The scene recognition method presented here is concerned with the simultaneous recognition of the shape and detection of the position and orientation in the two-dimensional raw-image, of a three-dimensional target being imaged. This approach of scene recognition is applied for the recognition of either a non-self moving target as well as for the a self-moving target such as a living animal, in the presence of noises like lighting condition variations and other objects in the scene, considered as noises. The raw-image is employed since it does not increase the original noise, thereby being more transportable, and moreover contrary to a binary image processing, does not require any filtering processing time. In fact here, the problem of an object recognition in the raw-image is changed to an optimization problem of a model-based evaluation function, named surface-strips model-based fitness function. This fitness function possesses information about the shape of a target, and consists in the computation of the brightness difference between an internal surface and a contour-strips. In this research, in order to find a target object in every newly input raw-image to the recognition system, the highest peak of the distribution of the surface-strips model-based fitness function, which corresponds to the recognition results of the designated target, is searched by a hybrid genetic algorithm, which employs a population of potential solutions to perform the search of the target. This hybrid genetic algorithm employs the “global” search feature of a two-point crossover genetic algorithm (GA), to search a target, together with a GA-based localized search technique that focuses on the target of interest found so far, in order to perform an intensive localized search. The localized GA search technique relies on mutation of bits on the lower portion of genes in positional and orientational binary strings, in order to improve the GA-based scene recognition performance, in terms of fast and reliable recognition of the target. This generational scene recognition by the hybrid genetic algorithm can be designated as “evolutionary scene recognition and position/orientation detection”. In order to appraise the proposed scene recognition method, experiments by a hand-eye camera of a robot manipulator have been conducted to show its robustness and reliability with respect to various disturbing objects and lighting condition changes, and its effectiveness to recognize a natural fish swimming in a pool. These results have shown the suitableness of the method for manipulator real-time visual servoing.

Construction Robot Systems Using Mobile Manipulators - Piling up Blocks with Real Systems and Positioning Accuracy

This research is concerned with a construction robot system that uses a mobile manipulator in order to make a large fence made up of concrete blocks. One of problems in making such a fence by a mobile manipulator is the existence of stacking errors caused by traveling errors under nonholonomic constraints. In the present paper, we propose a basic construction robot system that uses a mobile manipulator equipped with a system of compensating for stacking errors by detecting traveling errors using landmarks and a hand-eye camera. In addition, the systems accuracy in positioning the blocks is evaluated through experiments with a prototype of experimental equipment.