A multiple measurement approach for feature alignment

The training process for a pattern recognition system involves developing a statistical model that accurately describes the features of an object. These feature statistics are determined during the training process through an ensemble of measurements of an object. In many cases, however, the starting location of these features varies from measurement to measurement. This problem is overcome by aligning the features from one measurement to the same features from all other measurements. This alignment process allows the pattern recognition system to obtain a proper statistical estimate of the features. Failure to do a good job at aligning these measurements will result in a corrupted statistical estimate of the features of an object; hence, significantly degrading the performance of the pattern recognition system. Previous approaches to this alignment problem have been based on measurement alignment techniques, whose performance is highly dependent on the consistency of the measurements. This paper discusses how individual measurement to measurement alignments can be combined to produce a high quality alignment over the entire set of N measurements.