Algorithm 1 Training strategy of FPIRST

Input: Given $R$ videos from the HNUFD video dataset, feature parameter image training sample $X=\left\{X_1,X_2,\dots ,X_M\right\}$ after data processing and their type labels $Y=\left\{Y_1,Y_2,\dots ,Y_M\right\}$ . Output: The well-trained model FPIRST. 1: Construct the FPIRST method shown in Figure 1; 2: Initialize the parameters; 3: Repeat 4:   Randomly select a batch of instances $X_b$ from $X$ ; 5:   Forward learn training samples through the FPIST model; 6:   Compute the error $L\left(\theta \right)$ by $L\left(\theta \right)=\Vert y_t-{\widehat{y}}_t{\Vert }_2^2;$ 7:   Propagate the error back through FPIST and update the parameters; 8:   Find $\theta$ by minimizing $L\left(\theta \right)$ with $X_b$ ; 9: Until the end condition is satisfied.