Gesture Recognition Algorithm of Human Motion Target Based on Deep Neural Network
Algorithm 1
Human motion target gesture recognition algorithm.
(i)
Input: the original data of human motion, as well as the positioning results in the minimum space-time domain of the motion gesture node structure diagram.
(ii)
Output: recognition results of human motion target gesture of .
(iii)
According to the test samples in the gesture database of the human motion target to be recognized and the sample training set , obtain the human gesture feature distribution set as .
(iv)
Where is the number of human motion target gestures in the training set.
(v)
Construct a deep neural network classifier and obtain the weighted value of the deep neural network classifier as .
(vi)
where is the initialized eigenvalues and is binarized fitting results. Through the feature extraction results of motion gesture nodes, the deep neural network is introduced to obtain the input and output iterative equations of the deep neural network classifier: .
(vii)
where is the learning pace length and is the maximum iteration times of the training. Using the structural similarity algorithm, the weighted coefficient of the human motion target gesture classifier is obtained and expressed as .
(viii)
Through the deep neural network classifier, the gesture characteristics of the human motion target are calibrated, and the recognition statistics are obtained as .
(ix)
According to the classification method of the video image, the data are fused and classified and recognized. The image pixels after feature extraction are traversed through the window sliding to traverse the entire image, and the calculation process can be expressed as .
(x)
where is the traverse result, is the position of the output of the feature map of the previous layer of the node; is the value of the feature diagram at line and column , is the value at line and column ; and is the derivative error.
(xi)
As the number of traversal results deepens, a connection and sharing relationship is formed, and the activation function is used to transform the linear transformation into a nonlinear transformation [19]. After introducing the nonlinear activation function, the deep network can simulate any function. The PReLU activation function was selected for this study .
(xii)
where is the input node. This function is a piecewise function. When , the gradient is not 0, which solves the dead zone problem of the disappearance of the gradient. The sliding window is used with the same size and step size to calculate the sliding matrix and feature map. From the perspective of the amount of data and the number of parameters, the amount of calculation is reduced. It can reduce dimension and abstract results at the same time and improve the fault tolerance of the algorithm [20].
(xiii)
The fully connected method is used to connect the network nodes, and the output formula of each neuron is .
(xiv)
where is the input value of the node; is the activated function; is the weight vector, and is the deviation. is the transpose symbol. Through the abovementioned full connection method, the output information characteristics can be gathered.
(xv)
The aggregated human motion target gesture information features are extracted, and the gesture recognition result of the human motion target based on the difference of biological characteristics is obtained: .
