A method for location-independent identity recognition

Abstract

The application discloses a position-independent identity recognition method and belongs to the field of target recognition. The application recognizes target identity through mobile phone motion data, opens another path for target identity recognition. First, inertia sensor data of the mobile phone in the hand, waist, pocket and upper arm is acquired, then the data and corresponding targets are used for training an identity recognition network, and finally the identity recognition network is used for recognizing the target identity of newly acquired data. The application opens a new method for identity recognition, and the recognition preparation rate reaches 87.8%.

Description

Technical Field

[0001] This invention belongs to the field of target recognition, and in particular to location-independent identity recognition methods. Background Technology

[0002] Current identification technology based on inertial sensors involves placing the phone in different parts of the body (hand, waist, pocket, upper arm, etc.). The sensor signals obtained from different phone placements will vary. If the placement of the test data is inconsistent with the placement of the model training data, the accuracy will be poor. Summary of the Invention

[0003] This invention identifies the target's identity by using data from the mobile phone's gyroscope, opening up another path for target identity recognition.

[0004] The present invention provides a location-independent identity recognition method, which includes:

[0005] Step 1: Obtain the 3-axis acceleration of the phone at 4 locations to get the 3-axis data and build a database; the 4 locations are: front of pants, arm, waist, and handheld.

[0006] Step 2: Train the identity recognition network. The identity recognition network structure is as follows:

[0007] The identity recognition network consists of 6 layers cascaded in sequence, with an input vector of size 128*3;

[0008] The first layer consists of a convolutional kernel size of 5, a stride of 1, and 64 output channels, followed by a max pooling layer with a kernel size of 2, and finally an activation function ReLU.

[0009] The second layer consists of a convolutional kernel size of 5, a stride of 1, and 128 output channels, followed by a max pooling layer with a kernel size of 2 and a stride of 2, and finally an activation function ReLU.

[0010] The third layer consists of a convolutional kernel size of 5, a stride of 1, and 256 output channels, followed by a max pooling layer with a kernel size of 2, and finally an activation function ReLU.

[0011] The fourth layer consists of a convolutional kernel size of 5, a stride of 1, and 512 output channels, followed by a max pooling layer with a kernel size of 2, and finally the ReLU activation function.

[0012] The fifth layer consists of a convolutional kernel size of 5, a stride of 1, and 1024 output channels, followed by a max pooling layer with a kernel size of 2, and finally an activation function.

[0013] Fully connected layer: Output size is 44;

[0014] Step 3: Network training; The following encoding rules are used to create training labels during network training;

[0015] The training label encoding rules are as follows:

[0016] Let the training label be an n-dimensional vector, and divide it into two parts, n=<n1,n2>, where n1 represents the position encoding field and n2 represents the identity encoding field. The position encoding field n1 adopts onehot format and contains a total of 4 bits, with the front of the pants / arm / waist / hand corresponding to 0001 / 0010 / 0100 / 1000 respectively.

[0017] The identity encoding field n2 adopts onehot format. The number of bits in the identity encoding field n2 is the same as the number of individuals involved in the identification of the system.

[0018] Step 4: During the testing phase, the data format of the network input and output is the same as that in the training phase, but the identity is determined solely by comparing the identity field n2.

[0019] This invention utilizes data from a mobile phone's inertial sensor, pioneering a new method for identity recognition with high accuracy. Attached Figure Description

[0020] Figure 1 This is a diagram of the location recognition network structure of the present invention;

[0021] Figure 2 This is a comparison chart of experimental results in a specific embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the data acquisition location for this invention;

[0023] Figure 4 This is a schematic diagram of the location and identity encoding method in a specific embodiment of the present invention. Detailed Implementation

[0024] Experimental objective: To train an identity recognition network. Specifically, to identify a pedestrian based on their acceleration signal during walking.

[0025] The data includes: gait data of 34 people collected by the built-in sensors of the mobile phone. The sensors are placed in four positions (front of pants / arm / waist / hand), and collect acceleration data for six movements (going upstairs / going downstairs / walking / running / standing / sitting).

[0026] 1. Feature extraction: First, the data is segmented by gait cycle, divided into segments of 1, 2, and 8. The accelerations of the x, y, and z axes are synthesized as the feature input of the network.

[0027] 2. Division of training and test sets: The sampling frequency is 100Hz, and the sampling time is 60 seconds. Therefore, each person will collect 6000 data points. Since there will be interference at the beginning and end of the data collection, the data between 200-5000 will be used as the training set, and the data between 4800-5800 will be used as the test set. According to our gait cycle division, there are approximately 128 points. The gait cycle division is performed on these data points to form the training set and the test set.

[0028] Implementation of the comparative experiment:

[0029] Experiment 1: Training using only identity IDs as labels.

[0030] Since the goal is to identify individuals, the fully connected layer of the network should output the individual's ID. For 34 individuals, the output of the fully connected layer should also be 34. Simultaneously, the labels for the training and testing data should be the individual's ID, and training and testing should be conducted based on this. Experiment 2: Training using the label encoding method of this invention;

[0031] Figure 2 The figure shows a comparison of the accuracy on the test sets of Experiment 1 and Experiment 2. As can be seen from the figure, the accuracy on the test set using the method of this invention was improved from 78.6% to 87.8%, an improvement of approximately 9.2 percentage points.

[0032] The network parameters for identity recognition in this invention are shown in the table below.

[0033] Parameter name value Number of convolution kernels 64-128-256-512-1024 kernel size 5 Network layers 5 Pooling size 2 Dropout probability 0.2 Batch sample size 128 Number of iterations 100 Optimization methods Adam Regularity coefficient 0.0001 Learning rate 0.0001 beta_1 0.9 beta_2 0.999 epsilon 1E-08

Claims

1. A method for position-independent identity recognition, the method comprising: Step 1: obtaining 3-axis acceleration of a mobile phone at 4 positions to obtain 3-axis data and establish a database; The 4 positions are: front of trousers, arm, waist, and hand holding; Step 2: training an identity recognition network, the identity recognition network structure is: The identity recognition network comprises 6 layers connected in sequence, and the input size is a vector of 128*3; The first layer: the convolution kernel size is 5, the step is 1, the output channel is 64, then a max-pooling layer with a kernel size of 2 is connected, and finally an activation function ReLU is used; The second layer: the convolution kernel size is 5, the step is 1, the output channel is 128, then a max-pooling layer with a kernel size of 2 and a step of 2 is connected, and finally an activation function ReLU is used; The third layer: the convolution kernel size is 5, the step is 1, the output channel is 256, then a max-pooling layer with a kernel size of 2 is connected, and finally an activation function ReLU is used; The fourth layer: the convolution kernel size is 5, the step is 1, the output channel is 512, then a max-pooling layer with a kernel size of 2 is connected, and finally an activation function ReLU is used; The fifth layer: the convolution kernel size is 5, the step is 1, the output channel is 1024, then a max-pooling layer with a kernel size of 2 is connected, and finally an activation function is used; The full connection layer: the output size is 44; Step 3: network training; the following encoding rules are used to prepare training labels during network training; The training label encoding rules are as follows: Let the training label be an n-dimensional vector, which is divided into two parts, n = <n1, n2>, n1 represents the position encoding field, and n2 represents the identity encoding field, the position encoding field n1 adopts onehot format, the position encoding field n1 contains 4 bits in total, and front of trousers / arm / waist / hand holding corresponds to 0001 / 0010 / 0100 / 1000 respectively; The identity encoding field n2 adopts onehot format, and the identity recognition of how many people is involved in the system, the identity encoding field n2 contains how many bits; Step 4: during the test stage, the data format of network input and output is consistent with that during the training stage, but the identity is determined only by comparing the identity field n2.

Images