Tested on the IADB Infrared Human Behavior dataset. The test selected 120 behavior samples in the data set (consisting of 10 people performing 12 different behaviors), and the behavior types include bending over, kicking with one foot, punching punches, jumping, jumping forward with both feet, standing still. Jump up, run, walk sideways, hop, walk, one-handed and two-handed. After motion cycle detection, t is 15 frames. The DMHI-PC feature generation parameters are set to p=4, q=6, M-N=6, and the cumulative contribution rate of 99% of the feature values is maintained after PCA dimensionality reduction. The multi-class SVM is composed of 12 SVM base classifiers, and the regularization parameters of all base classifiers are C=1 and linear polynomial kernel functions are used.

 

Run 10 consecutive 3-fold cross-validation tests to get the confusion matrix shown in Table 5.1. The values on the diagonal of the matrix reflect the correct recognition rate of various behaviors, and the values on the off-diagonal reflect the misrecognition rate between any two types of behaviors. The test results show that the recognition rate of most of the tested behaviors is quite good.