A road recognition early warning system and method for a blind guiding stick

Abstract

The application discloses a road recognition early warning system and method for a blind guiding stick, belongs to the technical field of blind person's assisted travel, collects the perturbation vibration signals generated when the blind guiding stick contacts with the ground, and constructs a perturbation vibration characteristic vector group; the characteristic vector is input into a multi-layer fusion characteristic analysis model, abnormal sensitive features such as low-amplitude periodic disturbance, heterogeneous feedback response, contact delay gradient change and the like are extracted; a perturbation abnormal risk atlas is constructed based on the above features, and a risk level is divided; a high-risk area is further extracted and a dangerous point set is recognized; the dangerous point is spatially fitted with a blind user path prediction track, and it is judged whether there is a path crossing risk; if there is an intersection, a multi-modal early warning is triggered, including vibration, voice and directionality prompt; the application can realize accurate identification and dynamic early warning of small road abnormalities, improve blind guiding safety and environmental adaptability, and has low power consumption and high real-time performance.

Description

Technical Field

[0001] This invention relates to the field of assistive travel technology for the blind, specifically to a road recognition and warning system and method for guide canes. Background Technology

[0002] Currently, guide canes, as an important assistive tool for visually impaired people's daily travel, mainly rely on physical contact to provide feedback on the ground. However, in actual use, guide canes often fail to identify certain highly dangerous but minute road anomalies, such as partially collapsed but not completely damaged manhole covers, slightly raised curbs due to thermal expansion and contraction, or "soft depressions" formed by the separation of paving materials from the base layer. These anomalies, because they do not have obvious height differences or physical protrusions, are often overlooked by guide canes, causing visually impaired people to suddenly lose their balance during use, resulting in falls, or even serious injuries such as falling into manholes or being injured by vehicles, posing a serious safety hazard.

[0003] Although some assistive devices for the visually impaired have incorporated technologies such as image recognition or ultrasound for road detection, they are generally slow to respond to subtle abnormal features, making it difficult to provide early warnings of such "secondary hazards." Furthermore, due to the complex and varied materials of urban road surfaces and the significant differences in road feedback characteristics across different areas, traditional early warning models lack universality and struggle to dynamically adapt to new scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a road recognition and early warning system and method for guide canes, in order to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a road recognition and early warning method for guide canes, comprising:

[0006] The perturbation contact vibration signal between the guide cane and the ground is obtained, and the perturbation vibration feature vector group V is constructed.

[0007] The feature vector set V is input into the multi-layer fusion feature analysis model F for micro-anomaly identification to extract the anomaly-sensitive feature subset V' corresponding to the target road segment;

[0008] Based on the anomaly-sensitive feature subset V', a micro-disturbance anomaly risk map R for the current road area is constructed, where R classifies the potential danger level of each unit contact point according to the risk level.

[0009] Extract abnormal regions D in the risk map R whose risk level is higher than the preset threshold T, and determine the set of dangerous points D' based on the connectivity of the regions and the degree of structural mutation.

[0010] Spatial fitting is performed between the dangerous point set D' and the predicted walking path trajectory P of the guide user to determine whether there is a risky section that crosses D'.

[0011] If an intersection is detected, the guide cane warning mode is triggered, outputting a multimodal warning signal that includes directional cues, vibration intensity changes, and voice prompts.

[0012] Preferably, the construction of the perturbation vibration feature vector set V specifically includes:

[0013] Collect micro-displacement data and vibration acceleration data after the guide cane makes initial contact with the ground;

[0014] Bandpass filtering was applied to the vibration data;

[0015] Short-time Fourier transform is used to extract time-frequency domain features from the filtered signal, forming an initial feature set including vibration energy density, dominant frequency offset, and periodic rate of change.

[0016] The initial feature set is normalized and dimensionality is reduced by principal component analysis to construct a perturbation vibration feature vector set V.

[0017] Preferably, the construction of the multi-layer fusion feature analysis model F includes:

[0018] Based on the perturbation vibration feature vector group V, a multi-layer fusion feature analysis network structure is constructed, which includes an input layer, a dual-branch feature extraction layer and a fusion discriminant layer. The dual-branch feature extraction layer is used to extract time-domain dynamic features and frequency-domain steady-state features, respectively.

[0019] In the temporal feature branch, a one-dimensional convolutional neural network is used to perform convolution operations on the time series structure of V to obtain the temporal feature matrix Vt.

[0020] In the frequency domain feature branch, V is decomposed into a multi-resolution frequency domain based on multi-scale wavelet packet transform to obtain a frequency domain feature matrix Vf containing different frequency band response modes;

[0021] The Vt and Vf are input into the fusion discriminant layer, and a comprehensive feature vector Vc is formed by weighted feature concatenation. A multi-layer fusion feature analysis model F is then generated based on the Softmax classifier for subsequent extraction of anomaly-sensitive features.

[0022] Preferably, the abnormally sensitive feature subset V' includes low-amplitude periodic perturbation features, heterogeneous feedback response features, and contact delay gradient change features.

[0023] Preferably, the construction of the perturbation anomaly risk map of the current road area based on the anomaly-sensitive feature subset V' includes:

[0024] The abnormally sensitive feature subset V' is associated with the spatial coordinates of each cane touch point to form a set of spatiotemporal data units D containing feature and location information;

[0025] Based on dataset D, a density-based spatial clustering algorithm is used to identify areas with abnormally dense local features and to preliminarily divide potential high-risk area clusters C.

[0026] Weighted statistical analysis was performed on the data units within each region cluster C, and a multi-factor risk scoring function R(x, y) was constructed based on the low-amplitude periodic disturbance value, energy non-equilibrium degree, and gradient change amplitude.

[0027] The risk score is mapped to the two-dimensional coordinate system of the road to generate a perturbation anomaly risk map R, where each unit contact point is marked as high, medium and low according to the risk level range.

[0028] Preferably, determining the set of danger points D' includes:

[0029] In the risk map R, each unit contact point is screened using a preset risk threshold T, high-risk points with risk scores not lower than T are extracted, and a preliminary abnormal area set D is formed.

[0030] Based on set D, connectivity analysis is performed on adjacent risk points according to the condition that the spatial distance does not exceed the set connectivity radius, and several continuous abnormal region blocks Dg are formed by merging eight neighboring regions.

[0031] For each anomalous region block Dg, calculate its internal risk score gradient change rate and boundary mutation amplitude, and determine the local point with the most significant structural mutation in the region based on the gradient change peak and mutation threshold.

[0032] Local points in each region block Dg whose gradient change rate exceeds a set mutation threshold are extracted as a dangerous point set D'.

[0033] Preferably, the step of spatially fitting the set of dangerous points D' with the predicted trajectory P of the guide user's walking path to determine whether there is a risky section where the path crosses D' includes:

[0034] Based on the user's current pose, inertial navigation data, and historical path features, a short-term path prediction trajectory P for the user is constructed using the Kalman filter algorithm.

[0035] Spatial interpolation and path segment discretization are performed on the predicted trajectory P to generate a trajectory point set P′ consisting of multiple path nodes.

[0036] Spatial matching is performed between the path point set P′ and the danger point set D′. If the Euclidean distance between any path node and any danger point is less than the set risk proximity radius Rth, it is determined that the potential path crosses the risk section.

[0037] Record the set of all path segments that meet the proximity condition as the risk segment set Prisk.

[0038] The present invention also provides a road recognition and warning system for guide canes, comprising:

[0039] Vibration signal acquisition module: acquires the perturbation contact vibration signal between the guide cane and the ground, and constructs a perturbation vibration feature vector group V;

[0040] The micro-anomaly identification and analysis module: inputs the feature vector group V into the multi-layer fusion feature analysis model F for micro-anomaly identification, and extracts the anomaly sensitive feature subset V' corresponding to the target road segment;

[0041] Risk map construction module: Based on the abnormal sensitive feature subset V', construct the micro-disturbance anomaly risk map R of the current road area, where R classifies the potential danger level of each unit contact point according to the risk level;

[0042] Hazard point extraction module: Extracts abnormal regions D in the risk map R whose risk level is higher than the preset threshold T, and determines the hazard point set D' based on the region connectivity and the degree of structural abrupt change;

[0043] Risk assessment module: Spatial fitting of the dangerous point set D' with the predicted walking path trajectory P of the guide user to determine whether there is a risk section that crosses D';

[0044] Multimodal warning output module: If an intersection is detected, the guide cane warning mode is triggered, and a multimodal warning signal including directional prompts, vibration intensity changes and voice prompts is output.

[0045] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0046] 1. This invention collects the micro-perturbation contact vibration signals between the guide cane and the ground, constructs a multi-dimensional feature vector, and fuses time-domain and frequency-domain features. This enables precise identification of subtle abnormal ground structures, such as soft depressions, localized uplifts, or hollow areas—hidden hazards that are difficult to detect with conventional guide equipment. Combined with risk mapping and spatial clustering analysis, it achieves spatial modeling and risk classification of abnormal areas, ensuring high-sensitivity detection and dynamic perception of micro-scale road surface hazards.

[0047] 2. Compared with the shortcomings of existing guide assist methods that rely on coarse-grained environmental perception or a single early warning mechanism, this invention proposes for the first time a multi-layer fusion analysis method based on perturbation feature perception. Combining path prediction and spatial fitting mechanisms, it effectively judges the intersection relationship between the user's travel path and dangerous areas. Furthermore, through a multi-modal early warning method that combines vibration, voice and direction prompts, it significantly improves the timeliness and adaptability of early warnings. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0049] Figure 1 This is a flowchart of the method of the present invention.

[0050] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0052] Example 1, please refer to Figure 1 As shown in this embodiment, a road recognition and warning method for guide canes includes:

[0053] The perturbation contact vibration signal between the guide cane and the ground is obtained, and the perturbation vibration feature vector group V is constructed.

[0054] The feature vector set V is input into the multi-layer fusion feature analysis model F for micro-anomaly identification to extract the anomaly-sensitive feature subset V' corresponding to the target road segment;

[0055] Based on the anomaly-sensitive feature subset V', a micro-disturbance anomaly risk map R for the current road area is constructed, where R classifies the potential danger level of each unit contact point according to the risk level.

[0056] Extract abnormal regions D in the risk map R whose risk level is higher than the preset threshold T, and determine the set of dangerous points D' based on the connectivity of the regions and the degree of structural mutation.

[0057] Spatial fitting is performed between the dangerous point set D' and the predicted walking path trajectory P of the guide user to determine whether there is a risky section that crosses D'.

[0058] If an intersection is detected, the guide cane warning mode is triggered, outputting a multimodal warning signal that includes directional cues, vibration intensity changes, and voice prompts.

[0059] To effectively identify micro-scale contact anomalies between the guide cane and the ground, this invention specifically includes the following steps when constructing the micro-perturbation vibration feature vector set V:

[0060] This invention integrates a piezoelectric triaxial vibration sensor module and a MEMS displacement sensor at the bottom of the guide cane, for synchronously acquiring minute displacement changes and vibration acceleration data generated at the moment the guide cane head contacts the ground. The sensor sampling frequency is set to 1000 Hz, which can capture minute disturbance changes at the 1-millisecond level to ensure the timeliness and integrity of the acquired signal.

[0061] Given the presence of strong low-frequency ground vibrations (such as vehicle traffic) and high-frequency electronic interference (such as motors, electromagnetic waves, etc.) in urban road environments, this invention performs bandpass filtering on the acquired raw vibration signals to retain effective signal components in the frequency range of 1 to 100 Hz.

[0062] A fifth-order Butterworth filter was used, with cutoff frequencies set to 1 Hz (low-pass) and 100 Hz (high-pass) to suppress interference from irrelevant frequency bands and extract mid-to-low frequency disturbance features that are highly correlated with the contact behavior of the guide cane.

[0063] The filtered vibration signal was analyzed in the time-frequency domain using Short-Time Fourier Transform (STFT). A Hamming window function with a sliding window size of 256 sampling points and an overlap of 128 points was used to segment the signal into frames. The following key feature parameters were extracted through spectral intensity analysis:

[0064] Vibration energy density: The integrated power spectral density function in the 1 to 100 Hz frequency band, which characterizes the strength of the stiffness response of the ground structure;

[0065] Main frequency offset: Calculates the deviation between the maximum amplitude frequency and the fundamental frequency, used to identify feedback differences between different materials;

[0066] Periodic rate of change: Analyzes the trend of spectral peak changes within adjacent windows to detect response fluctuations caused by discontinuous structures.

[0067] The above features constitute the initial feature set F0, which is used to describe the core properties of the vibration response during the contact of the guide cane with the ground.

[0068] To ensure consistency of feature dimensions and stability of subsequent model inputs, each feature in F0 is normalized according to the following formula: each feature value is subtracted from the mean of the sample set and then divided by the standard deviation to convert it into a zero-mean unit variance form.

[0069] Principal Component Analysis (PCA) is used to reduce the dimensionality of the normalized feature matrix. Let the feature dimension be n. After solving the feature covariance matrix, the first k principal components (where k is the smallest integer that can cumulatively explain more than 85% of the original feature information, typically 3 to 5 dimensions) are retained to form the perturbation vibration feature vector group V.

[0070] This invention, after obtaining the perturbation vibration feature vector set V, constructs a multi-layer fusion feature analysis model F, and extracts an anomaly-sensitive feature subset V' from it for road hazard identification. Specifically, it includes:

[0071] The model F consists of an input layer, a dual-branch feature extraction layer, and a fusion discriminant layer. The model employs a modular design to enhance the independent representation of time-domain and frequency-domain features, and improves overall recognition accuracy through subsequent fusion. Specifically:

[0072] The input layer is used to receive a set of perturbation vibration feature vectors V of dimension d;

[0073] The dual-branch feature extraction layer consists of parallel temporal convolution channels and frequency domain transform channels;

[0074] The fusion discriminant layer is used to integrate features from two channels and output a fusion vector for classification.

[0075] In the time-domain branch, the input vector group V is constructed into a time-series input format of shape (d×1) and fed into a one-dimensional convolutional neural network (1D-CNN) for processing. The number of convolutional kernels is set to 32, the kernel width is 3, the stride is 1, and the activation function is ReLU.

[0076] After two layers of convolution and max pooling operations, the temporal feature matrix Vt is obtained, which is used to describe the dynamic change behavior of the ground feedback signal in the time dimension.

[0077] In the frequency domain branch, the Daubechies wavelet basis is used to perform a three-level wavelet packet decomposition on the feature vector group V to obtain multiple sub-signal components of different frequency bands. Their energy spectrum, frequency band center value and rate of change are calculated respectively to construct the frequency domain feature matrix Vf.

[0078] The multi-scale wavelet packet decomposition has strong local time-frequency analysis capabilities, and can extract weak frequency response changes caused by differences in ground material structure, which helps to identify secondary road anomalies such as "soft depressions" or "micro-bulges".

[0079] After aligning the time-domain feature matrix Vt with the frequency-domain feature matrix Vf, a comprehensive feature vector Vc is constructed using a weighted feature concatenation method. The weight factors are determined through cross-validation. Let the weight of the time-domain features be α and the weight of the frequency-domain features be β, where α plus β equals 1. The recommended initial values ​​are set to α=0.6 and β=0.4.

[0080] The comprehensive vector Vc is input to the fusion discriminant layer. Through a classifier structure containing two fully connected layers and a Softmax output layer, the probability distribution corresponding to each type of micro-anomaly is output as the recognition result of model F.

[0081] Samples whose predicted probability values ​​in the Softmax output are greater than a set threshold τ (the recommended value for τ is 0.65) are considered anomalies. Their corresponding key feature dimensions are extracted to form an anomaly-sensitive feature subset V'.

[0082] The abnormally sensitive feature subset V' mainly includes:

[0083] Low-amplitude periodic perturbation characteristics: These reflect the low-amplitude periodic oscillations that occur on the time axis when the guide cane comes into contact with locally non-uniform materials (such as loose paving layers). The calculation method includes: inputting a pre-processed time-domain vibration signal s(t) at a sampling frequency of fs; calculating the spectrum S(f) using Fast Fourier Transform (FFT), truncating the frequency range of 1–30Hz; calculating the set of frequency components Flow whose amplitude is below a set amplitude threshold Ath (e.g., 10% of the maximum signal amplitude); periodically detecting the components in Flow to determine if there is a pattern in the spacing of the basic frequency components (approximately equidistant distribution); and calculating its periodic perturbation characteristic value LPCF. Where: f0 is the lowest frequency principal component, Δf is the average frequency spacing, and n is the number of low-amplitude principal frequencies.

[0084] Heterogeneous feedback response characteristics: These measure the unevenness of energy distribution across frequency bands in the frequency domain, revealing spectral anomalies caused by abrupt changes in ground material structure or voids. Calculation methods include:

[0085] The vibration signal s(t) is subjected to three-level wavelet packet decomposition to obtain eight frequency band sub-signals. ;

[0086] Calculate the energy of each sub-signal ;

[0087] The normalized energy vector is ;

[0088] Calculate the eigenvalues ​​of the heterogeneous feedback response .

[0089] Contact delay gradient change characteristics: This reflects a time delay or abrupt change in the signal response after the guide cane contacts the ground. It is typically used to detect localized road subsidence, hollow areas, or soft spots. Calculation methods include:

[0090] Set the starting point of contact with the guide cane as... The vibration signal s(t) with a time window length of T=200 milliseconds after contact is extracted; the first derivative of each sampling point within the time window is calculated. That is, the vibration change gradient;

[0091] Sliding calculation of the absolute mean of gradient values The changes within the small window yield the characteristic value of the contact delay gradient change, expressed as: In the formula, The sliding window step size is recommended to be 10 milliseconds, and CDGV is the contact delay gradient change characteristic value.

[0092] The abnormally sensitive feature V' serves as the basic input for constructing the subsequent risk map, used to dynamically label potentially dangerous areas and provide risk warning support for guide path planning.

[0093] This invention constructs a perturbation anomaly risk map R based on an anomaly-sensitive feature subset V'. This map reflects the risk level corresponding to the micro-vibration feedback of the guide cane at different locations, specifically including:

[0094] At each point where the guide cane touches the ground, its GPS coordinates (or relative coordinates based on inertial navigation) are collected and recorded as (x, y). These coordinates are then bound to the abnormal sensitive feature V' extracted at the corresponding time to form a spatiotemporal data unit. ;in: This represents the characteristic value of low-amplitude periodic disturbances. This represents the characteristic value of a non-homogeneous feedback response. This represents the characteristic value of the contact delay gradient change. This represents the corresponding geographic coordinates. All data units form set D.

[0095] Density-based clustering (DBSCAN) was performed on set D using the DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise). The spatial radius parameter ε and the minimum number of samples (minPts) were set. For example, ε ranged from 0.5 meters to 1 meter, representing the largest spatial neighborhood within the same cluster; minPts was set to 5, indicating that at least 5 adjacent data units were required to form a valid cluster. The clustering results formed multiple region clusters. Each cluster represents a local area with frequent anomalies. Points that are not classified in the clustering results are considered noise points and can be temporarily ignored.

[0096] Within each cluster C, its contained data units The following processing is performed: Calculate the mean or maximum value of the three sensitive features within the cluster as representative values; construct a multi-factor risk scoring function R(x,y) in the following form: Where α, β, γ are the weight coefficients of each feature, satisfying... Weights can be set using training data or expert experience; for example, a recommended value can be set to... ; This is the risk score for the current location; the higher the value, the higher the potential risk.

[0097] All calculated The values ​​are mapped to the corresponding two-dimensional coordinate system of the road, forming a continuous spatial risk distribution map R. To achieve visualization and risk warning, the risk score is divided into three levels: high-risk area: (e.g., 0.7); Medium-risk areas: (e.g., 0.4 ≤ R < 0.7); Low-risk area: (e.g., <0.4).

[0098] In this invention, the process of forming a set of danger points D' includes:

[0099] First, risk screening is performed on all unit contact points (x, y) in the perturbation anomaly risk map R. A preset risk threshold T is set (e.g., T=0.6), and points that meet the following condition are extracted: R(x,y)≥T. The screening results constitute a preliminary set of anomaly regions D, which contains all spatial points with potential hazard characteristics, and serve as the basic input for subsequent connectivity analysis.

[0100] Perform spatial adjacency calculations on all points in set D, and set a connectivity radius rc (e.g., 1 meter) to determine whether two points are adjacent.

[0101] The eight-neighborhood region merging algorithm is used to divide the region blocks: if the distance between any two points in the eight directions (up, down, left, right, and diagonal) does not exceed rc, they are considered to be connected and belong to the same abnormal region block Dg.

[0102] The merged result is a set of multiple consecutive abnormal region blocks: For each anomalous region block Dg, construct its local two-dimensional risk scoring function. And calculate the gradient rate of change of the score over the spatial range: In the formula, This indicates how quickly the risk score changes in the horizontal direction, i.e., the rate of change of the score in the x-direction. This indicates how quickly the risk score changes in the vertical direction, i.e., the rate of change of the score in the y-direction.

[0103] Simultaneously, the set of boundary points BDg is extracted, and the difference between it and the mean of the internal region is calculated as the magnitude of boundary abrupt change. ;in If the gradient value at a point within the region is greater than the set mutation threshold... (If set to 0.3), and its location is in the inflection point region of score increase or decrease, then the point is determined to possibly correspond to a mutation structure. Traverse all regions Dg, and extract local points that meet the following conditions as danger points: gradient change rate The location shows a significant abrupt change in score compared to its neighboring areas, or is situated within a high-score boundary zone.

[0104] All points that meet the conditions constitute the final set of dangerous points D', which serves as the core data for path risk cross-analysis, early warning decision-making, and visualization: The point set D' focuses on the locations on the road where structural abrupt changes such as local collapse, uplift, and hollowing are most likely to occur.

[0105] This invention constructs a risk zone judgment mechanism based on the spatial fitting relationship between the predicted trajectory P and the set of dangerous points D'. The technical steps are as follows:

[0106] The current position of the guide user is obtained in real time by the integrated navigation system, including GPS positioning (or visual SLAM) and attitude angle and velocity information output by the IMU sensor. Let the current position be (x0, y0) and the velocity vector be (vx, vy).

[0107] Based on the above information, a Kalman filter algorithm is used to dynamically predict the user's future short-term walking path. The state variable is defined as: Xt=[xt,yt,vxt,vyt]; the state transition model is a linear prediction equation, the control input is the user's current acceleration estimate, and the observations are updated by GPS / IMU.

[0108] The path prediction time window is set to 2 to 3 seconds, and the output continuous path points form the initial predicted trajectory P: P={p1,p2,...,pn},pi=(xi,yi); the predicted path P is then linearly interpolated and discretized with a fixed step size Δ. The step size Δ is set to 0.2 meters to 0.5 meters, and the interpolated trajectory point set P′ is generated. Where p′ represents the uniformly distributed path nodes after interpolation, used to enhance the response capability to microscale path offsets.

[0109] For each trajectory node in P′ Spatial matching is performed with all points dj in the danger point set D′, and their Euclidean distances are calculated. The expression is: If there exists any If the trajectory point p′i is considered to have entered the risk approach range, Rth is the set risk judgment radius (e.g., 1.0 meter, determined based on the user's stride and reaction distance).

[0110] Index all path points p′i that satisfy the distance condition and mark the path segments to which they belong. These are risky crossing segments; merging all such risky segments forms a set of risky segments: .

[0111] Based on the premise that the path of the guide user and the set of danger points D′ are found to intersect through spatial fitting, the early warning triggering mechanism of this invention specifically includes the following to achieve effective warning and avoidance guidance of potential dangers:

[0112] Extract the first risk path node prisk=(xr,yr) from the risk segment set Prisk, and obtain the user's current location pcurr=(xc,yc).

[0113] Calculate the direction angle from the current location to the risk point using the following formula. The direction angle The internal logic determines whether the current orientation is "forward", "left", "right" or "sharp turn", and uses this information as a directional cue for subsequent modal output scheduling.

[0114] Based on the risk score R(x,y) of each path node, the risk level is divided into the following three intervals: high risk: R≥0.7; medium risk: 0.4≤R<0.7; low risk: R<0.4.

[0115] Correspondingly, the output parameters of the built-in vibration motor of the guide cane are configured as follows:

[0116] High risk: Frequency 200Hz, amplitude 2mm (strong vibration);

[0117] Medium risk: Frequency 100Hz, amplitude 1mm (medium vibration);

[0118] Low risk: 50Hz frequency, 0.5mm amplitude (light vibration).

[0119] When a risk warning is triggered, the guide cane issues a voice reminder through its built-in speech synthesis chip (which supports offline TTS synthesis), and the content automatically switches according to the type of risk and the road environment.

[0120] The above directional hints Vibration parameter group Vb and voice prompt content Vc are uniformly encapsulated into a multimodal early warning signal set: The priority scheduling strategy of this invention is as follows:

[0121] If the direction deflects at an angle If the angle exceeds the set turning threshold (e.g., 30 degrees), voice and directional prompts will be triggered first; if the risk score R is greater than 0.7, a high-frequency vibration signal will be forcibly activated without user confirmation.

[0122] Example 2, please refer to Figure 2 As shown in this embodiment, a road recognition and warning system for guide canes includes:

[0123] Vibration signal acquisition module: acquires the perturbation contact vibration signal between the guide cane and the ground, and constructs a perturbation vibration feature vector group V;

[0124] Micro-anomaly identification and analysis module: Input the feature vector group V into the multi-layer fusion feature analysis model F for micro-anomaly identification, and extract the anomaly sensitive feature subset V' corresponding to the target road segment, where V' includes low-amplitude periodic disturbance features, heterogeneous feedback response features and contact delay gradient change features;

[0125] Risk map construction module: Based on the abnormal sensitive feature subset V', construct the micro-disturbance anomaly risk map R of the current road area, where R classifies the potential danger level of each unit contact point according to the risk level;

[0126] Hazard point extraction module: Extracts abnormal regions D in the risk map R whose risk level is higher than the preset threshold T, and determines the hazard point set D' based on the region connectivity and the degree of structural abrupt change;

[0127] Risk assessment module: Spatial fitting of the dangerous point set D' with the predicted walking path trajectory P of the guide user to determine whether there is a risk section that crosses D';

[0128] Multimodal warning output module: If an intersection is detected, the guide cane warning mode is triggered, and a multimodal warning signal including directional prompts, vibration intensity changes and voice prompts is output.

[0129] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A road recognition and early warning method for guide canes, characterized in that: include: The perturbation contact vibration signal between the guide cane and the ground is obtained, and the perturbation vibration feature vector group V is constructed. The feature vector set V is input into the multi-layer fusion feature analysis model F for micro-anomaly identification to extract the anomaly-sensitive feature subset V' corresponding to the target road segment; Based on the anomaly-sensitive feature subset V', a micro-disturbance anomaly risk map R for the current road area is constructed, where R classifies the potential danger level of each unit contact point according to the risk level. Extract abnormal regions D in the risk map R whose risk level is higher than the preset threshold T, and determine the set of dangerous points D' based on the connectivity of the regions and the degree of structural mutation. Spatial fitting is performed between the dangerous point set D' and the predicted walking path trajectory P of the guide user to determine whether there is a risky section that crosses D'. If an intersection is detected, the guide cane warning mode is triggered, outputting a multimodal warning signal that includes directional cues, vibration intensity changes, and voice prompts.

2. The road recognition and early warning method for guide canes according to claim 1, characterized in that: The construction of the perturbation vibration feature vector set V specifically includes: Collect micro-displacement data and vibration acceleration data after the guide cane makes initial contact with the ground; Bandpass filtering was applied to the vibration data; Short-time Fourier transform is used to extract time-frequency domain features from the filtered signal, forming an initial feature set including vibration energy density, dominant frequency offset, and periodic rate of change. The initial feature set is normalized and dimensionality is reduced by principal component analysis to construct a perturbation vibration feature vector set V.

3. The road recognition and early warning method for guide canes according to claim 1, characterized in that: The construction of the multi-layer fusion feature analysis model F includes: Based on the perturbation vibration feature vector group V, a multi-layer fusion feature analysis network structure is constructed, which includes an input layer, a dual-branch feature extraction layer and a fusion discriminant layer. The dual-branch feature extraction layer is used to extract time-domain dynamic features and frequency-domain steady-state features, respectively. In the temporal feature branch, a one-dimensional convolutional neural network is used to perform convolution operations on the time series structure of V to obtain the temporal feature matrix Vt. In the frequency domain feature branch, V is decomposed into a multi-resolution frequency domain based on multi-scale wavelet packet transform to obtain a frequency domain feature matrix Vf containing different frequency band response modes; The Vt and Vf are input into the fusion discriminant layer, and a comprehensive feature vector Vc is formed by weighted feature concatenation. A multi-layer fusion feature analysis model F is then generated based on the Softmax classifier for subsequent extraction of anomaly-sensitive features.

4. The road recognition and early warning method for guide canes according to claim 1, characterized in that: The abnormally sensitive feature subset V' includes low-amplitude periodic perturbation features, heterogeneous feedback response features, and contact delay gradient change features.

5. A road recognition and early warning method for guide canes according to claim 1, characterized in that: The method for constructing a perturbation anomaly risk map of the current road area based on an anomaly-sensitive feature subset V' includes: The abnormally sensitive feature subset V' is associated with the spatial coordinates of each cane touch point to form a set of spatiotemporal data units D containing feature and location information; Based on dataset D, a density-based spatial clustering algorithm is used to identify areas with abnormally dense local features and to preliminarily divide potential high-risk area clusters C. Weighted statistical analysis was performed on the data units within each region cluster C, and a multi-factor risk scoring function R(x, y) was constructed based on the low-amplitude periodic disturbance value, energy non-equilibrium degree, and gradient change amplitude. The risk score is mapped to the two-dimensional coordinate system of the road to generate a perturbation anomaly risk map R, where each unit contact point is marked as high, medium and low according to the risk level range.

6. A road recognition and early warning method for guide canes according to claim 5, characterized in that: The determination of the hazard point set D' includes: In the risk map R, each unit contact point is screened using a preset risk threshold T, high-risk points with risk scores not lower than T are extracted, and a preliminary abnormal area set D is formed. Based on set D, connectivity analysis is performed on adjacent risk points according to the condition that the spatial distance does not exceed the set connectivity radius, and several continuous abnormal region blocks Dg are formed by merging eight neighboring regions. For each anomalous region block Dg, calculate its internal risk score gradient change rate and boundary mutation amplitude, and determine the local point with the most significant structural mutation in the region based on the gradient change peak and mutation threshold. Local points in each region block Dg whose gradient change rate exceeds a set mutation threshold are extracted as a dangerous point set D'.

7. A road recognition and early warning method for guide canes according to claim 6, characterized in that: The step of spatially fitting the set of dangerous points D' with the predicted walking path trajectory P of the guide user to determine whether there is a risky section that crosses D' includes: Based on the user's current pose, inertial navigation data, and historical path features, a short-term path prediction trajectory P for the user is constructed using the Kalman filter algorithm. Spatial interpolation and path segment discretization are performed on the predicted trajectory P to generate a trajectory point set P′ consisting of multiple path nodes. Spatial matching is performed between the path point set P′ and the danger point set D′. If the Euclidean distance between any path node and any danger point is less than the set risk proximity radius Rth, it is determined that the potential path crosses the risk section. Record the set of all path segments that meet the proximity condition as the risk segment set Prisk.

8. A road recognition and warning system for guide canes, used to implement the road recognition and warning method for guide canes as described in any one of claims 1-7, characterized in that: include: Vibration signal acquisition module: acquires the perturbation contact vibration signal between the guide cane and the ground, and constructs a perturbation vibration feature vector group V; The micro-anomaly identification and analysis module: inputs the feature vector group V into the multi-layer fusion feature analysis model F for micro-anomaly identification, and extracts the anomaly sensitive feature subset V' corresponding to the target road segment; Risk map construction module: Based on the abnormal sensitive feature subset V', construct the micro-disturbance anomaly risk map R of the current road area, where R classifies the potential danger level of each unit contact point according to the risk level; Hazard point extraction module: Extracts abnormal regions D in the risk map R whose risk level is higher than the preset threshold T, and determines the hazard point set D' based on the region connectivity and the degree of structural abrupt change; Risk assessment module: Spatial fitting of the dangerous point set D' with the predicted walking path trajectory P of the guide user to determine whether there is a risk section that crosses D'; Multimodal warning output module: If an intersection is detected, the guide cane warning mode is triggered, and a multimodal warning signal including directional prompts, vibration intensity changes and voice prompts is output.

Images