Algorithm 1 Pseudo-code of the proposed ConvLSTM architecture

1:X = [] 2:Y = [] 3:ListClasses = os.listdirinputDir 4:i = 0 5:fork in ListClasses do 6:    $Yt=np.zeros(shape=(numClasses\left)\right)$ 7:    $Yt\left[i\right]=1;$ 8:    $print\left(k\right)$ 9:    $ListFiles=os.listdir(os.path.join(inputDir,c\left)\right);$ 10:    for l in ListFiles do 11:        $frames=framesExtraction(os.path.join(os.path.join(inputDir,c),l\left)\right)$ 12:    end for 13:end for 14:$X=np.asarray\left(X\right)$ 15:$Y=np.asarray\left(Y\right)$ 16:$X\_train,X\_test,y\_train,y\_test=train\_test\_split(X,Y)$ 17:$X\_val,x\_test,y\_val,Y\_test=train\_test\_split(X\_test,y\_test)$ 18:$seq.add\left(TimeDistributed\right(Conv2D(filters=16,kernel\_size=(3,3\left)\right))$ 19:$seq.add\left(ConvLSTM2D\right(filters=64,kernel\_size=(3,3)\left)\right)$ 20:$seq.add\left(BatchNormalization\right(\left)\right)$ 21:$seq.add\left(Conv2D\right(filters=16,kernel\_size=(3,3)\left)\right)$ 22:$seq.add\left(Dropout\right(0.5\left)\right)$ 23:$seq.add\left(Flatten\right(\left)\right)$ 24:$seq.add\left(Dense\right(256\left)\right)$ 25:$seq.add\left(Dropout\right(0.5\left)\right)$ 26:$seq.compile{(}^{‘}ada{m}^{’},loss={}^{‘}categorical\_crossentrop{y}^{’},metrics={[}^{‘}accurac{y}^{’}])$