Object recognition in distributed computing environment

Abstract

This paper focuses on the enhancement of processing speed in three-dimensional object recognition that is attained by means of parallel processing technique. The parallel processing technique discussed in this paper is applied to the recognition of an object extracted from a scene reconstructed by segment-based stereo, especially to matching of the feature points of the object extracted from the scene with a geometric object model generated in advance. The object recognition method using feature points of an object requires much processing time if the target object has a number of feature points because the feature points are handled in various combinations to generate candidates for correspondence with a predefined model, and each candidate is checked for its validity to determine the position/orientation of the object finally. The parallel processing technique, however, can reduce this processing time drastically because the validity checking process for each candidate can be performed separately. This paper introduces a system that performs dynamic load balancing by task pooling between the master and slave processes, which can enhance the processing speed with the increase of the number of workstations incorporated in a workstation cluster, even though the workstations have different performances. This system was evaluated for the contribution of the parallel processing technique to the improvement of performance in the object-model matching process. The evaluation result proved the effectiveness of the parallel processing technique, showing a maximum performance improvement of 3.4 times over the case of serial processing when the system incorporated 12 processors (with the total calculation performance being 6.3 times the performance of a single processor having the highest speed in the workstation cluster). This paper also shows that dynamic load balancing provides an effective distributed computing environment even though the system incorporates a mixture of processors with various performances, such that the fastest processor operates three times as fast as the slowest one. © 2000 Scripta Technica, Syst Comp Jpn, 32(1): 10–19, 2001