A hybrid analytical/intelligent methodology for sensor fusion

In this thesis, a sensor fusion technique for object recognition has been developed. Sensor fusion deals with the problem of the synergistic use of multiple sensors in order to enhance the performance of estimating the information source. The complementary and redundant nature of multiple sensors can reduce the amount of uncertainty so as to enhance results pertaining to the measurement. The objective of this thesis is the development of a sensor fusion algorithm for identifying objects which may be isolated or overlapping. The progress of automated object recognition systems has suffered from the difficulty of interpreting an image which contains overlapping objects. A systematic approach to sensor fusion is presented to address this problem so that it can satisfy the two requirements in object recognition: accuracy and timely assessment. It is assumed that a vision sensor and a tactile sensor are employed in a simulation mode, to identify objects on a test bed. The images from both sensors are also simulated via a computer. Since the two sensors are complementary in the sense that they have a different direction of view and redundant image data are available on some parts of the objects, thus providing information about the overlap and hence the identification of each object is easily achieved. For the recognition of isolated objects, fusion occurs at the symbolic level. Dempster-Shafer theory has been used in pooling evidence from several sources of information. In order to employ DS theory for the purpose of object recognition, the procedure of deriving basic assignments has been developed and a degree of certainty (DOC) associated with a decision is proposed to indicate a reliability index. The concept of active perception has been provided to show that the order of features affects the performance which is a function of time and DOC and to suggest that the method of optimal selection of sensors can be found by examining the measure of specificity. An associative memory has been used to speed up the hypothesis and verification procedure. A new type of associative memory without a varying threshold is developed for the purpose of future implementation.