Based on the obtained MDI, the intermediate vector can be obtained by using a feature extraction method similar to the D0MHI-PC in the previous section.

 

HL={H1, H2,..., Hq, Hmax}

 

It is worth noting that the dimensionality of HL vectors may be high, and the target may also be polluted by thermal infrared imaging noise and MDI generated noise, which will inevitably affect the classification performance. For this reason, continue to perform dimensionality reduction on HL. In order to find the lowest possible dimension under the premise of keeping the essential characteristics of the data as much as possible, the maximum likelihood estimation (MLE) is used to estimate the eigendimension of H, and then linear discriminant analysis (LDA) is applied to HL is transformed into a new vector H with the dimensionality estimated by MLE, called the MDI-PC eigenvector in this section. Finally, the MDI-PC feature vector is used for behavior classification and recognition. Since overfitting is avoided, computational efficiency can be improved while achieving better separability.