Blind person discrimination device, system and method based on distributed fiber sensing equipment

Abstract

The application provides a blind person discrimination device, system and method based on a distributed optical fiber sensing device.The device comprises an optical time domain reflectometer module, a signal preprocessing module, a signal cutting module, a signal feature extraction module and a signal classification module.The method provides a high recognition rate without video monitoring equipment and wearable sensing equipment, more comprehensively and effectively provides application support for the field of security and blind person travel, enhances the adaptability for different human body recognition, and identifies the source of the human body walking signal accepted by the current device, whether it is a visually impaired user or a normal user, so as to provide related services for the visually impaired user in the future.The long-time and short-time feature joint recognition mode is adopted in the feature extraction algorithm, and the accuracy is further improved.

Description

Technical Field

[0001] This invention belongs to the field of human body recognition technology, and in particular relates to a blind person identification device, system and method based on distributed optical fiber sensing equipment. Background Technology

[0002] With the rapid development of internet and wireless communication technologies, users are placing increasingly higher demands on identity authentication technologies, primarily in terms of accuracy, real-time performance, and contactless operation. Theoretically, biometric authentication is the most reliable method because it uses each person's unique biometric characteristics to represent their digital identity and performs corresponding verification, making it virtually impossible to forge. Common biometric recognition methods include fingerprint scanning, facial recognition, iris recognition, and behavioral recognition. Among these, biometric behavioral recognition technology identifies individuals by observing their walking, typing, or handling of a device. Users of biometric authentication typically experience a smoother and faster experience because it eliminates the need for password recall, and it is also more difficult for attackers to deceive them.

[0003] Gait recognition technology, as an emerging biometric identification technology, aims to achieve automatic identity recognition based on differences in human walking patterns. Gait refers to the way a person walks and its various characteristics, such as the posture of each joint, stride length, stride force, and walking speed. Compared with other biometric technologies, gait recognition has the advantages of being non-contact, observable from a distance, and difficult to fake. The non-contact nature of gait recognition means that the entire recognition process does not require active human intervention; the long-distance observability means that the input information needed for recognition can be acquired, transmitted, and analyzed from a distance; and the difficulty in faking gait recognition stems from the fact that everyone has a different walking posture. As early as the 1960s, medical research by Murray et al. demonstrated that everyone's gait is different, making it extremely difficult to fake a walking posture. This is because subtle differences exist in muscle strength, tendon and bone length, bone density, visual sensitivity, coordination ability, past walking experience, weight, center of gravity, the degree of muscle or bone damage, and individual walking habits.

[0004] Currently, visually impaired individuals in society need to use guide canes, professional guide dogs, or professional guide assistants to walk outdoors on designated tactile paving. Canes and guide dogs can only distinguish obstacles on the ground; they are ineffective against tree branches extending into the air. Furthermore, many public places in cities are unfriendly to guide dogs, and the urban environment also interferes with their work. In complex urban environments, the use of canes is often ineffective due to the lack of maintenance of tactile paving and other infrastructure. The labor costs for guide assistants are high, and the number of professional guide assistants is limited, making this solution impractical for visually impaired users in non-first-tier cities. Indoor walking environments are even more complex. For the vast visually impaired population, a low-cost, low-latency, and high-performance travel solution is still lacking. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art, and to propose a blind person identification device, system and method based on distributed optical fiber sensing equipment.

[0006] The present invention is achieved through the following technical solution: The present invention proposes a device for distinguishing between blind and non-blind people based on distributed optical fiber sensing equipment. The device includes an optical time domain reflectometer module, a signal preprocessing module, a signal segmentation module, a signal feature extraction module, and a signal classification module.

[0007] The optical time-domain reflectometer module is used to generate light signal intensity data generated by vibration signals;

[0008] The signal preprocessing module is used to preprocess the extracted two-dimensional intensity signal, process the collected data, filter out background noise, calculate the number of people identified in a certain time slot, separate the collected different human body signals, and remove still frames without target user signals.

[0009] The signal segmentation module is used to segment the preprocessed signal, divide long signals into frames, and separate the two signals within each frame.

[0010] The signal feature extraction module is used to extract features from the segmented signal fragments and output a set of feature vectors;

[0011] The signal classification module is used to classify the feature vectors extracted from each group of signals and output whether the person is blind or not.

[0012] This invention proposes a method for distinguishing between blind and non-blind people based on a distributed optical fiber sensing device. The method is implemented using the aforementioned device, and the specific method is as follows:

[0013] Step 1: Start the optical time domain reflectometer and record the target user's walking data S(z,t), obtain the generated signal strength data and output it in the form of a TXT document;

[0014] Step 2: Preprocess the obtained two-dimensional intensity signal using the signal preprocessing module. First, set a certain threshold X. threshold The filter considers signals with energy below a certain intensity as noise and filters them out directly; still frames are removed, i.e., an energy threshold E is first set. threshold The signal is divided into multiple frames, and the frame energy E is calculated for each frame. i The system determines whether the frame is a still frame, checks for "walking-stopping-walking" patterns, corrects the disruption to the original step interval caused by discarding the stop frame, calculates the start and end points of each frame after correction, splices the frames and outputs them; the two-channel sensor signals of the device are separated to obtain dual-channel signals, and the maximum amplitude value of the two-dimensional signal of each channel of the dual-channel signal is taken to obtain the one-dimensional signal X(t) of time-maximum amplitude.

[0015] Step 3: Use the signal cutting module to complete the signal framing; First, set the number of frames and divide the entire signal length into the number of equal-length modules. By detecting whether there is an incomplete step in the frame, determine whether it is necessary to extend the frame header or frame tail. If there is an incomplete step, extend the frame header or frame tail at the corresponding position to complete the output of each frame.

[0016] Step 4: Extract features from the signal within each frame, extracting a total of 5 sets of features, including the single-frame feature vector [f] of the i-th frame. i,1 f i,2 f i,3 f i,4 f i,5 The elements are defined as follows:

[0017] (1)f i,1 : Step frequency, which is the estimated number of steps per minute;

[0018] (2)f i,2 Stride length, which is the average distance of a single step;

[0019] (3)f i,3 : The average energy of the vibration signal;

[0020] (4)f i,4 The zeroing rate of the time-domain signal X(t) with maximum amplitude;

[0021] (5)f i,5 : Average spatial span of the vibration signal;

[0022] Step 5: Combine all features into a feature vector, standardize the feature vector, randomly select a sample R from the training sample set each time, find the k nearest neighbors of R from the sample set of the same class as R, find the k nearest neighbors from each of the R different class sample sets, update the weight of each feature, use the ReliefF algorithm to filter the feature weights, and input all features into the CNN network for convolution and classification.

[0023] Furthermore, the step frequency is defined as the number of steps per unit time, and its calculation method is as follows:

[0024] 1.1 Set the sliding window length L, single sliding distance D, step point peak time coordinate array T, and observation window length L. w Calculate the number of sliding steps K, initialize the one-dimensional array flag, initialize the dynamic array G storing the autocorrelation peak value, calculate the mean value inside the sliding window, if the mean value inside the window is large, it proves that there is a vibration signal in the window, that is, there may be a step to label the window with the large mean value.

[0025] 1.2 Initialize the coordinate array within the peak time of each step point, search for the peak coordinates, and the corresponding coordinates are the time position coordinates of the step point; set the flag j to 1, and when j < K-6, perform a judgment; if...

[0026] flag(j)+flag(j+1)+flag(j+2+p)=3+p,p≥0

[0027] Then perform the judgment again, if

[0028] flag(j+p+1)+flag(j+p+1+q)=-2-q,q≥0,

[0029] Then store the coordinates of the maximum value of window j+2+p into G;

[0030] 1.3 Calculate the length L of array G G Number of steps N step =L G -1, initialize the peak time coordinate array T of the step points. p L before X(t) W Intra-point search peak coordinate T P (1) For the kth value, it can be obtained through formula T P (k)=T P (k-1)+G[k]-G[k-1] is calculated to obtain;

[0031] 1.4, Output step frequency is: f i,1 =F=60×f s ×N step / [T P(L G )-T P (1)].

[0032] Furthermore, the stride length is the distance between adjacent step points, and the number of step points N is input. step For a single frame signal and peak coordinates, for the i-th step, when i does not exceed the maximum step limit, calculate P. i (z), and estimate the position z i When i exceeds the maximum number of steps, calculate S1, S2, and S; where,

[0033] P i (z) is calculated by the following formula:

[0034]

[0035] In the formula, t range =200;

[0036] z i Calculated by the following formula:

[0037]

[0038] The formulas for calculating S1, S2, and S are as follows:

[0039]

[0040]

[0041]

[0042] Furthermore, the average energy is calculated using the following formula:

[0043]

[0044] Furthermore, the zeroing rate is calculated using the following formula:

[0045]

[0046] In the formula, Ⅱ(·) is an indicator function. If · is true, it takes the value 1; otherwise, it takes the value 0.

[0047] Furthermore, the spatial span reflects the intensity and propagation capability of the vibration signal. For the input two-dimensional signal, the length of the observation window is set, and the number of windows M is calculated. After calculating the spatial span of this window, the window completes its displacement. The spatial span within the window is defined as:

[0048] W i =max(z) i )-min(z i )

[0049]

[0050] The beneficial effects of this invention are as follows:

[0051] This invention provides a device, system, and method for identifying blind individuals based on distributed fiber optic sensing equipment. This method offers a high recognition rate without requiring video surveillance equipment or wearable sensing devices, providing more comprehensive and effective application support for security and the mobility of blind individuals. It enhances adaptability to different human body types, identifying whether the human walking signals received by the device originate from a visually impaired user or a sighted user with low latency and high accuracy, thus paving the way for future services for visually impaired users. The feature extraction algorithm employs a joint recognition approach combining long-term and short-term features, further improving accuracy. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the process of the device of the present invention.

[0053] Figure 2 This is a schematic diagram of a two-dimensional intensity-time signal acquired by an optical time-domain reflectometer.

[0054] Figure 3 This is a schematic diagram of a one-dimensional intensity-time signal obtained after preprocessing.

[0055] Figure 4 This is a schematic diagram of the channel splitting of a two-dimensional signal.

[0056] Figure 5 This is a physical image of the optical time-domain reflectometer used in the device described in this invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] This invention proposes a device and method for distinguishing between blind and non-blind people based on distributed fiber optic sensing equipment. It can overcome the limitations of the original vision- or sensor-based human body recognition application scenarios and the inconvenience of wearing devices. It enhances the adaptability to different human body recognition, and identifies the source of the human walking signal received by the device as a visually impaired user or a normal user with low latency and high accuracy, so as to provide relevant services for visually impaired users in the future.

[0059] This invention proposes a device for distinguishing between blind and non-blind people based on distributed optical fiber sensing equipment. The device includes an optical time domain reflectometer (OTDR) module, a signal preprocessing module, a signal segmentation module, a signal feature extraction module, and a signal classification module.

[0060] The optical time-domain reflectometer module is used to generate light signal intensity data generated by vibration signals;

[0061] The signal preprocessing module is used to preprocess the extracted two-dimensional intensity signal, process the collected data, filter out background noise, calculate the number of people identified in a certain time slot, separate the collected different human body signals, and remove still frames without target user signals.

[0062] The signal segmentation module is used to segment the preprocessed signal, divide long signals into frames, and separate the two signals within each frame.

[0063] The signal feature extraction module is used to extract features from the segmented signal fragments and output a set of feature vectors;

[0064] The signal classification module is used to classify the feature vectors extracted from each group of signals and output whether the person is blind or not.

[0065] This invention proposes a method for distinguishing between blind and non-blind people based on a distributed optical fiber sensing device. The method is implemented using the aforementioned device for distinguishing between blind and non-blind people based on a distributed optical fiber sensing device. Specifically, the method comprises:

[0066] Step 1: Start the optical time domain reflectometer and record the target user's walking data S(z,t), obtain the signal intensity data it generates and output it in the form of a TXT document; use a data processing tool to read the two columns of data in the TXT document and store them in the time and intensity arrays respectively;

[0067] Step 2: Preprocess the obtained two-dimensional intensity signal using the signal preprocessing module. First, set a certain threshold X. threshold The filter considers signals with energy below a certain intensity as noise and filters them out directly; still frames are removed, i.e., an energy threshold E is first set. threshold The signal is divided into multiple frames, and the frame energy E is calculated for each frame. iThe process involves summing the squares of all signals within a frame to determine if it is a stationary frame. If the energy within the frame is higher than a threshold, it is considered a non-stationary frame; otherwise, it is considered a stationary frame. Finally, it checks for "walk-stop-walk" patterns. For a non-stationary frame j, it checks if its K adjacent frames are stationary frames. If a stationary frame appears in the adjacent frames of the non-stationary frame, it indicates that the experimental data exhibits a "walk-stop-walk" phenomenon. The step frequency algorithm of the feature extraction module is used to extract the step interval D of the original non-stationary frame and the step interval Q after the stationary frame. Using the difference between the two, if the DQ interpolation is greater than a threshold, DQ zeros are padded between frame j and frame j+K+1; otherwise, QD zeros are padded between these two frames. Correct the disruption of the original step interval caused by discarding the stop frame; finally, calculate the start and end points of each frame after correction, splice the frames and output them; separate the two sensing signals of the device to obtain dual-channel signals, obtain the two positions of the maximum amplitude per unit time and record them as an array of two channels, take the maximum amplitude value of the two-dimensional signal of each channel of the dual-channel signal per unit time, and obtain the one-dimensional signal X(t) of time-maximum amplitude.

[0068] Step 3: Use the signal cutting module to complete the signal framing; First, set the number of frames and divide the entire signal length into the number of equal-length modules. By detecting whether there is an incomplete step in the frame, determine whether it is necessary to extend the frame header or frame tail. If there is an incomplete step, extend the frame header or frame tail at the corresponding position to complete the output of each frame.

[0069] Step 4: Extract features from the signal within each frame, extracting a total of 5 sets of features, including the single-frame feature vector [f] of the i-th frame. i,1 f i,2 f i,3 f i,4 f i,5 The elements are defined as follows:

[0070] (1)f i,1 : Step frequency, which is the estimated number of steps per minute;

[0071] (2)f i,2 Stride length, which is the average distance of a single step;

[0072] (3)f i,3 : The average energy of the vibration signal;

[0073] (4)f i,4 The zeroing rate of the time-domain signal X(t) with maximum amplitude;

[0074] (5)f i,5 : Average spatial span of the vibration signal;

[0075] Step 5: Combine all features into a feature vector and standardize the feature vector to speed up the convergence of the classification algorithm. Each time, randomly select a sample R from the training sample set, then find the k nearest neighbors of R from the sample set of the same class as R, and find the k nearest neighbors of R from each of the R different class sample sets. Then update the weight of each feature and use the ReliefF algorithm to filter the feature weights. Input all features into the CNN network for convolutional classification.

[0076] The step frequency is defined as the number of steps per unit time, and its calculation method is as follows:

[0077] 1.1 Set the sliding window length L, single sliding distance D, step point peak time coordinate array T, and observation window length L. w Calculate the number of sliding steps K, initialize the one-dimensional array flag, initialize the dynamic array G storing the autocorrelation peak value, calculate the mean value inside the sliding window, if the mean value inside the window is large, it proves that there is a vibration signal in the window, that is, there may be a step to label the window with the large mean value.

[0078] 1.2 Initialize the coordinate array within the peak time of each step point, search for the peak coordinates, and the corresponding coordinates are the time position coordinates of the step point; set the flag j to 1, and when j < K-6, perform a judgment; if...

[0079] flag(j)+flag(j+1)+flag(j+2+p)=3+p,p≥0

[0080] Then perform the judgment again, if

[0081] flag(j+p+1)+flag(j+p+1+q)=-2-q,q≥0,

[0082] Then store the coordinates of the maximum value of window j+2+p into G;

[0083] 1.3 Calculate the length L of array G G Number of steps N step =L G -1, initialize the peak time coordinate array T of the step points. p L before X(t) W Intra-point search peak coordinate T P (1) For the kth value, it can be obtained through formula T P (k)=T P (k-1)+G[k]-G[k-1] is calculated to obtain;

[0084] 1.4, Output step frequency is: f i,1 =F=60×f s×N step / [T P (L G )-T P (1)].

[0085] The stride is the distance between adjacent step points. Due to the limitations of OTDR devices, a joint analysis of the signal's spatial dimension and peak position is used, with the number of input step points N. step For a single frame signal and peak coordinates, for the i-th step, when i does not exceed the maximum step limit, calculate P. i (z), and estimate the position z i When i exceeds the maximum number of steps, calculate S1, S2, and S; where,

[0086] P i (z) is calculated by the following formula:

[0087]

[0088] In the formula, t range =200;

[0089] z i Calculated by the following formula:

[0090]

[0091] The formulas for calculating S1, S2, and S are as follows:

[0092]

[0093]

[0094]

[0095] The average energy value is calculated using the following formula:

[0096]

[0097] The zeroing rate is calculated using the following formula:

[0098]

[0099] In the formula, Ⅱ(·) is an indicator function; it takes the value 1 if · is true, and 0 otherwise. Since the noise reduction process sets signal values ​​not higher than the noise threshold to zero, the zeroing rate is essentially the proportion of time occupied by the noise signal. In addition, intermittent single pulses are considered noise higher than the noise threshold.

[0100] Spatial span reflects the intensity and propagation capability of vibration signals. For an input two-dimensional signal, the length of the observation window is set, and the number of windows M is calculated. After calculating the spatial span of the current window, the window completes its displacement. The spatial span within the window is defined as:

[0101] W i =max(z) i )-min(z i )

[0102]

[0103] The method described in this invention uses a Convolutional Neural Network (CNN) as the classifier. Its structure consists of three layers: a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer uses filters to perform two-dimensional convolution operations on the image's two-dimensional grayscale matrix, obtaining the image's feature map through various convolution functions. Since CNNs are commonly used to process two-dimensional images, the amount of data generated after convolution is too large, which is detrimental to computer processing and storage. Therefore, a pooling layer is used to pool the results of convolution, compressing the data without affecting the recognition result. After processing by the convolutional and pooling layers, the hidden features in the original data are obtained. At this point, a fully connected layer is used to classify these features. The fully connected layer is a fully connected neural network; the classification result is obtained by adjusting the weights and the network.

[0104] The neural network used in the method of this invention contains 6 layers. Since the data generated by the scenario targeted by this invention is one-dimensional, a one-dimensional convolution module is used for convolution. The network first performs one-dimensional convolution on the input data and then performs mean pooling on the resulting one-dimensional feature vector. After repeating this operation, the data is input into a flattening layer. Finally, the flattened feature vector is input into a fully connected network to obtain the final output classification.

[0105] This invention provides a blind person identification device, system, and method based on distributed fiber optic sensing equipment. This method offers a high recognition rate without the need for video surveillance equipment or wearable sensing devices, providing more comprehensive and effective application support for security and blind mobility. It enhances adaptability to different human body recognition methods, identifying the source of human walking signals received by the device as either a visually impaired user or a sighted user with low latency and high accuracy, thus paving the way for future services for visually impaired users. The feature extraction algorithm employs a joint recognition approach combining long-term and short-term features, further improving accuracy.

[0106] Table 1. Symbol Explanation Table

[0107]

[0108]

Claims

1. A method for distinguishing between blind and non-blind people based on distributed fiber optic sensing devices, characterized in that, The method employs a device based on distributed optical fiber sensing equipment to distinguish between blind and non-blind individuals. The device includes an optical time domain reflectometer module, a signal preprocessing module, a signal segmentation module, a signal feature extraction module, and a signal classification module. The optical time-domain reflectometer module is used to generate light signal intensity data generated by vibration signals; The signal preprocessing module is used to preprocess the extracted two-dimensional intensity signal, process the collected data, filter out background noise, calculate the number of people identified in a certain time slot, separate the collected different human body signals, and remove still frames without target user signals. The signal segmentation module is used to segment the preprocessed signal, divide long signals into frames, and separate the two signals within each frame. The signal feature extraction module is used to extract features from the segmented signal fragments and output a set of feature vectors; The signal classification module is used to classify the feature vectors extracted from each group of signals and output whether the person is blind or not. The method is specifically as follows: Step 1: Start the optical time domain reflectometer and record the target user's walking data. It acquires the generated signal strength data and outputs it in the form of a TXT document; Step 2: Preprocess the obtained two-dimensional intensity signal using the signal preprocessing module. First, set a certain threshold. The filter considers signals with energy below a certain intensity as noise and filters them out directly. Remove still frames, i.e., first set the energy threshold. The signal is divided into multiple frames, and the frame energy is calculated for each frame. The process involves determining whether a frame is a still frame, checking for "walk-stop-walk" patterns, correcting the disruption to the original step interval caused by discarding stop frames, calculating the start and end points of each corrected frame, splicing the frames together, and outputting the result. The two sensor signals from the device are separated to obtain dual-channel signals. The maximum amplitude value per unit time is taken from the two-dimensional signal of each channel to obtain a one-dimensional signal with time-maximum amplitude. ; Step 3: Use the signal cutting module to complete the signal framing; First, set the number of frames and divide the entire signal length into the number of equal-length modules. By detecting whether there is an incomplete step in the frame, determine whether it is necessary to extend the frame header or frame tail. If there is an incomplete step, extend the frame header or frame tail at the corresponding position to complete the output of each frame. Step 4: Extract features from the signal within each frame, extracting a total of 5 sets of features as follows, the first... Single-frame feature vector The elements are defined as follows: (1) : Step frequency, which is the estimated number of steps per minute; (2) Stride length, which is the average distance of a single step; (3) : The average energy of the vibration signal; (4) Maximum amplitude time domain signal The rate of zeroing; (5) : Average spatial span of the vibration signal; Step 5: Combine all features into a feature vector, standardize the feature vector, and randomly select one sample from the training sample set each time. Then from and Find the same sample set of Each nearest neighbor sample, from each Finding in each different class of sample sets The system calculates the nearest neighbor samples, updates the weight of each feature, uses the ReliefF algorithm to filter the feature weights, and then inputs all features into the CNN network for convolutional classification.

2. The method according to claim 1, characterized in that, The step frequency is defined as the number of steps per unit time, and its calculation method is as follows: 1.1 Setting the length of the sliding window single sliding distance Step point peak time coordinate array , Observation window length Calculate the number of slides Initialize a one-dimensional array of flag bits. Initialize a dynamic array to store autocorrelation peaks. Calculate the mean value inside the sliding window. If the mean value inside the window is large, it proves that there is a vibration signal inside the window, that is, there may be a step to label the window with the large mean value. 1.2 Initialize the coordinate array within the peak time of each step point, search for the peak coordinates, and the corresponding coordinates are the time position coordinates of the step point; set the flag. When it is 1, Make a judgment in time, if ， Then perform the judgment again, if ， ， Then the window Store the coordinates of the maximum value ; 1.3 Calculate the array length Steps Initialize the peak time coordinate array of the step points ,exist The former Peak coordinates in point search For the A value, which can be obtained through the formula Calculated; 1.

4. Output step frequency is: .

3. The method according to claim 1, characterized in that, The stride is the distance between adjacent step points; the number of step points is input. Single-frame signal and peak coordinates, for the first Step, when Calculate when the maximum number of steps is not exceeded. And estimate the location , When the maximum number of steps is exceeded, calculate , , ;in, Calculated by the following formula: In the formula, ; Calculated by the following formula: , , The calculation formula is as follows: Unit: cm Unit: cm 。 4. The method according to claim 1, characterized in that, The average energy value is calculated using the following formula: 。 5. The method according to claim 1, characterized in that, The zeroing rate is calculated using the following formula: In the formula, It is an indicator function, if If true, the value is 1; otherwise, the value is 0.

6. The method according to claim 1, characterized in that, Spatial span reflects the intensity and propagation capability of vibration signals. For the input two-dimensional signal, the length of the observation window is set, and the number of windows is calculated. After calculating the spatial span of this window, the window completes its displacement; the spatial span within the window is defined as: 。

Images