Vol.4 No.2 2011

Research paper : ARGUS: Adaptive Recognition for General Use System (N. Otsu et al.)−80−Synthesiology - English edition Vol.4 No.2 (2011) Fig. 14 Comparative experiment of “gait” recognition[5][6][16]Fig. 13 Gait video and frame differenceFig. 12 Obtained discriminant feature spaces[11]and the additive properties of HLAC features. Once each of the patterns on the left in Fig.9 are presented to the system, the system instantly responds to the test image (on the right) with the numbers yi of each as y = (F F)−1F x. This is by virtue of additivity, where the feature vector x for the entire right-hand diagram can be decomposed into the linear sum x =Σ6i =1 yifi = [f1, . . . , f6]y = Fy, which has as its coefficient the number of feature (factor) vectors fi for each pattern.5.2 Recognition (/enumeration) of topological characteristicsNext, as an example of recognition that is independent of shape, recognition outcomes for topological characteristics are given in Fig. 10. By learning from examples using multiple regression analysis (MRA), the system correctly answered the number of objects (a) or the number of holes (b)[4]. Interestingly, from the examples, the system learned the Euler number that underlies the basis of topology (number of points − number of lines + number of planes)Note4 and used it for recognition.5.3 Recognition of faces and facial expressionsHLAC is not limited to binary images and can be directly applied to gray-scale images as well. Face recognition was done as such an example[12][13]. By integrating with discriminant analysis (DA), the HLAC features extracted from each layer of a pyramid of images representing multi-resolution, even the simple classification method MDDNote5 achieved a high recognition rate of more than 99 % among 119 people[13]. Furthermore, the method was applied to the difficult task of facial expression recognition for seven facial expressions by nine people (JAFFE Dataset[14], Fig.11). Using the MDD and the discriminant analysis that takes into account the position based weighting of HLAC features, a high recognition rate of more than 80 % was achieved[15].5.4 Recognition of person and motionUsing CHLAC features that are a natural expansion from HLAC features when moving images are considered, both object and movement can be recognized in a moving image. Videos of four motions (walking/running to the left/right) by five persons were converted to binary images based on frame difference and thresholding, and the CHLAC features were extracted. Fig.12 shows the results after applying discriminant analysis to person and motion, respectively[11]. Each cluster (category) is well grouped and separated,demonstrating the effectiveness of CHLAC features. Even with a simple classification method MDD, recognition rate of almost 100 % was obtained.5.5 Recognition of gaitIn recent years, the concept of “gait” has attracted attention as a key in the identification of individuals(terrorists, etc.) by surveillance cameras from a distance. Application of CHLAC and discriminant analysis together with the k-NN decision rule to the Gait Challenge Dataset (Fig.13) of 71 individuals compiled by the NIST in the United States has achieved the best performance in the world thus far, significantly surpassing the top five methods[16](Fig.14).5.6 Abnormality detectionWhen there are multiple objects in an image, CHLAC has the additive property in which the sum of the features of each object becomes the features of the whole; therefore, the feature vector for usual (normal) motion will be person1person5person3person4person2leftward runleftward walkrightward runrightward walk40353025201510501050-5-10-15-20-25-301050-5-10-15-20-25-30-35-40-35-30-25-20-15-10-5010155-60-40-20020406080-15-10-501015520253035b) Person discriminationa) Motion discriminationFrom left:CHLACUMDUSFCMUMITCASProbeABCDEFG1009080706050403020100Identification rate (%)Fig. 15 Example of abnormality detection (Here, “falling down” is abnormal.)TimeSubspace distance00204060801001201401601802000.0010.0020.0030.0040.0050.0060.007normalabnormal


page 9