Dynamic environment risk level assessment method for blind guiding robot dog

Abstract

The invention discloses a dynamic environment risk level assessment method for a blind guiding robot dog. According to the method, three core risk factors are calculated in real time by obtaining multi-modal sensor data of the blind guiding robot dog, wherein the three core risk factors include the visual scene information entropy used for quantifying visual complexity, the obstacle density used for measuring the physical congestion degree and the environmental dynamic change rate used for reflecting environmental dynamics. And then, the system performs weighted summation on the risk factors to obtain a comprehensive environmental risk score. Weighted weights can be dynamically adjusted according to context information such as the speed of the robot dog and ambient light so as to adapt to changeable scenes. And finally, the system determines a risk grade according to the risk score, and provides visual grading feedback for the visually impaired user in modes of vibration, voice and the like. According to the invention, the active perception ability, the navigation safety and the user confidence of the user can be obviously improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, service robots, assistive devices for the disabled, machine vision, and human-computer interaction, and in particular to a method for assessing the dynamic environmental risk level of a guide robot dog, which is also applicable to other guide robots. Background Technology

[0002] For visually impaired individuals, such as the blind or low-vision, walking safely and independently in complex outdoor environments is a significant challenge. Traditional assistive devices, such as guide dogs, white canes, or reliance on human companions, while providing some assistance, have significant limitations. For example, white canes can only detect low-lying obstacles at close range and cannot perceive suspended obstacles or anticipate dynamic changes in the distant environment; guide dogs are expensive to train, scarce in number, and cannot engage in complex verbal communication with their users; and human companions limit the independence of visually impaired individuals.

[0003] In recent years, the rapid development of robotics technology, especially quadruped robots represented by robotic dogs, has provided new possibilities for advanced assisted navigation. Robotic dog platforms possess excellent terrain adaptability and maneuverability, and can be equipped with various high-performance sensors, such as depth cameras, LiDAR, and inertial measurement units (IMUs). This theoretically gives them the ability to comprehensively perceive their surroundings. However, existing research mostly focuses on basic obstacle detection and path planning, i.e., a binary judgment of "whether it's passable." These systems often lack a mechanism for quantitatively assessing the "risk level" of the environment. For example, they struggle to distinguish between an empty corridor and a crowded shopping mall, and the navigation risks and psychological burdens for visually impaired users are drastically different. Existing guide robot systems cannot effectively assess and convey dynamic and comprehensive risk information such as changes in traffic density, sudden appearances of pedestrians, and visual complexity of routes. Therefore, users remain in a passive following state, lacking the ability to actively perceive and predict the environment, which to some extent affects the safety of use and the user's psychological trust. Summary of the Invention

[0004] This invention primarily addresses the technical problems of insufficient dynamic environmental risk assessment capabilities and the inability to provide users with quantitative and graded risk feedback in existing technologies. It provides a dynamic environmental risk level assessment method for a guide robot dog, which calculates the navigation risk level of the current environment in real time and provides multimodal feedback to blind users through vibration, voice, and other means, thereby improving safety and proactive perception capabilities.

[0005] The present invention addresses the aforementioned technical problems primarily through the following technical solution: a method for assessing the dynamic environmental risk level of a guide robot dog, comprising the following steps: S1: Data Acquisition: Acquire sensor data of the guide dog, wherein the sensor data includes at least visual image data and three-dimensional spatial point cloud data; S2: Risk Factor Calculation: Based on the sensor data, calculate various environmental risk factors, including: Visual scene information entropy calculated based on the aforementioned visual image data; Obstacle density calculated based on the aforementioned three-dimensional spatial point cloud data; The environmental dynamic change rate is calculated based on the three-dimensional spatial point cloud data over a continuous time period. S3: Comprehensive Score: The comprehensive environmental risk score is obtained by weighted summation of all environmental risk factors. S4: Feedback Generation: Based on the comprehensive environmental risk score, determine the risk level of the current environment and generate a user feedback signal corresponding to the risk level to be fed back to the user.

[0006] Preferably, the step of calculating the visual scene information entropy is as follows: S211: Convert the visual image data (color image) acquired in real time through the forward view of the depth camera into a grayscale image to eliminate color interference and focus on structural and texture information; S212: Apply edge detection operators such as Canny or Sobel to the grayscale image to perform edge detection and generate a binarized edge image; the contours, texture details and other complex structures of objects in the image will be highlighted in the form of white pixels (edges); S213: Divide the edge image into N×M non-overlapping image blocks, and calculate the proportion of edge pixels (white pixels) in each image block; S214: Visual scene information entropy H v The visual field information entropy H is defined as the Shannon entropy of the proportion of edges of all image patches, and is calculated using the following formula. v : ; In the formula, p i Let p be the proportion of edge pixels in the i-th image block. i =0, then set p i log2(p i H = 0. v The higher the value, the richer the edges and textures of the visual scene, the more complex the structure, and the higher the potential cognitive load and risk.

[0007] Preferably, the step of calculating the obstacle density is as follows: S221: Define the region of interest (AOI) in the three-dimensional point cloud data. Specifically, a fan-shaped or rectangular detection area can be set in front of the robot dog, such as the area within 5 meters of a 120-degree frontal view. S222: The point cloud within the region of interest is segmented into ground point cloud and non-ground point cloud. Specifically, algorithms such as RANSAC (Random Sample Consensus) are used to perform planar fitting on the point cloud, identify and remove ground point cloud, and retain non-ground point cloud. Then, the non-ground point cloud is clustered to obtain several obstacle clusters. Specifically, DBSCAN (Density-based Spatial Clustering) or Euclidean distance clustering algorithms are applied to the non-ground point cloud to aggregate spatially adjacent point clouds into independent clusters. Each cluster represents a potential obstacle (such as pedestrians, trees, vehicles, etc.). S223: Calculate the number of obstacle clusters per unit area in the region of interest, which is the obstacle density D. o This value directly reflects the level of congestion in the current environment.

[0008] Preferably, the step of calculating the dynamic change rate of the environment includes: S231: Between consecutive sampling moments, the number of newly appearing obstacle clusters and the number of instantly disappearing obstacle clusters are determined by tracking the obstacle clusters; in each processing cycle (i.e., between the current timestamp t and the previous timestamp t-1), the obstacle cluster list is compared using a target tracking algorithm to identify the number of newly appearing targets within this extremely short sampling interval, denoted as n. new (t), and the number of targets that disappear instantaneously, denoted as n. disappeared (t). These two values ​​reflect immediate changes in the environment; S232: Within a preset time window, the number of all instantaneously newly appearing obstacle clusters and the number of instantaneously disappearing obstacle clusters determined in step S231 will be accumulated to obtain a total change. To smoothly process and capture the overall dynamic trend over a period of time, this scheme employs a sliding window mechanism. First, the total number of frames W within the window is defined, determined by the preset time window length Δt (e.g., 1.0 seconds) and the system's sampling period Ts (e.g., 0.1 seconds): W = Δt / T s Then, the instantaneous changes within the past W frames are summed to obtain the total number N of newly appearing targets within the time window Δt. new Total number of targets that disappeared N disappeared The calculation formula is as follows: N new =Σ i=0 W-1 n new (ti); N disappeared=Σ i=0 W-1 n disappeared (ti); The above formula means: starting from the current time t, backtrack W frames (i from 0 to W-1), and sum up all the instantaneous new occurrences (or disappearances) detected in each of these W frames to obtain a total change within the recent historical window; S233: Divide the total change accumulated within the window period by the length of the time window Δt to calculate the smoothed environmental dynamic change rate C. d The calculation formula is as follows: C d =(N new +N disappeared ) / Δt; This value represents the total number of targets that change on average per second over a period of time. It can measure the dynamics of the environment, such as the entry and exit of people and the departure of vehicles.

[0009] Preferably, the weights used in the weighted summation in step S3 are dynamic weights, which are determined in the following way: S31: Through offline datasets or reinforcement learning, a set of basic weights β={β1,β2,β3} is trained, where β1 is the basic weight of visual scene information entropy, β2 is the basic weight of visual scene information entropy, and β3 is the basic weight of visual scene information entropy; they represent the relative importance of each risk factor under standard working conditions. S32: The final dynamic weights are determined based on the base weights and the context adjustment factor, using the following formula: β1'=β1·(1+k L ·max(0,L threshold -L)); β2'=β2·(1+k v ·v robot ); β3'=β3·(1+k v ·v robot ); In the formula, β1' is the dynamic weight of the visual scene information entropy, β2' is the dynamic weight of the visual scene information entropy, β3' is the dynamic weight of the visual scene information entropy, and k L k is the sensitivity coefficient for adjusting ambient light. v L is the sensitivity coefficient for adjusting the speed of the guide dog, and L is the current ambient light intensity. threshold v is the preset light intensity threshold. robotThe current speed of the guide robot dog; in low light conditions, increase the weight of visual complexity (β1'>β1); when traveling at high speed, significantly increase the weight of obstacle density and dynamic change rate (β2'>β2 and β3'>β3). Comprehensive Environmental Risk Score R t =β1'·H v +β2'·D o +β3'·C d .

[0010] Preferably, in step S4, if the comprehensive environmental risk score is less than the first threshold θ1, the risk level is green, indicating that the current environment is safe to pass through, and the user feedback is no feedback or a slight vibration warning; if the comprehensive environmental risk score is greater than or equal to the first threshold θ1 and less than the second threshold θ2, the risk level is yellow, indicating that the current environment has a medium risk and caution is required, and the user feedback is a medium vibration and a voice prompt "complex ahead"; if the comprehensive environmental risk score is greater than or equal to the second threshold θ2, the risk level is red, indicating that the current environment has a high risk and avoidance is required, and the user feedback is a strong vibration and a voice prompt "detour recommended".

[0011] L threshold A threshold is defined for what the system considers "insufficient lighting." When the actual light intensity L falls below this threshold, the system considers the performance of the visual sensor (camera) to have degraded, and the entropy H of the acquired visual information decreases. v The reliability of visual information decreases, therefore its weight in risk assessment needs to be increased to be more "vigilant" with uncertain visual information. threshold The methods for determining this mainly include: (I) Sensor Performance Calibration: The test object is the visual sensor (camera) mounted on a guide dog. In a controlled optical laboratory, using a standard light source and illuminance meter, the ambient light intensity is gradually reduced while the camera captures test patterns containing standard edges, textures, and objects. The signal-to-noise ratio (SNR), contrast, and accuracy and recall of the edge detection algorithm (step S212) output results of the images are analyzed under different light intensities. When the above evaluation indicators drop to a preset, unacceptable level (e.g., a large amount of edge features are lost, and noise is falsely detected as an edge), the corresponding light intensity value at this moment is determined as L. threshold This ensures that the threshold is directly related to the hardware and algorithm used in this scheme.

[0012] (II) Functional Field Testing: Simulating various typical low-light usage scenarios such as dusk, indoor dim light, and tunnels, the guide robot dog was tested in these scenarios, and the real-time light intensity L measured by the light sensor was recorded. Data analysis was used to identify the critical light range that could lead to navigation failure, missed obstacle detection, or significant deviations in system risk assessment. A conservative value within this range was chosen as L. threshold .

[0013] (III) Reinforcement Learning: a. Construct a simulation environment that includes variable lighting conditions and static and dynamic obstacles; b. Treat the guide dog as an intelligent agent, defining its state space, action space, and reward function for evaluating navigation safety and efficiency; c. [The text abruptly ends here, likely due to an incomplete sentence or missing information.] threshold As a trainable parameter in the agent's policy network; d. Train the agent in the simulation environment, and iteratively optimize the policy network using a reinforcement learning algorithm (such as the PPO algorithm) by maximizing the cumulative reward as the optimization objective, until the Lthreshold parameter converges to an optimal value.

[0014] k L and k V These are "adjustment knobs," which determine the system's sensitivity to insufficient lighting and high-speed motion. L This determines how quickly the weight of visual information entropy increases in poorer lighting conditions; k V This determines how quickly the weights of obstacle density and environmental dynamics increase with increasing speed. L and k V The determination method can be any of the following: (a) Empirical / Heuristic Tuning Method: Based on engineering experience, set a small initial value (e.g., 0.1 or 0.5). Design a series of test cases covering different combinations of lighting and speed. For example: Case 1: Well-lit, walking at low speed.

[0015] Case 2: Well-lit, running fast.

[0016] Case 3: Extremely poor lighting, slow exploration.

[0017] Case 4: Extremely poor lighting, running at high speed (high-risk scenario).

[0018] Run the system under each case and observe the output comprehensive environmental risk score R. t Does it align with the intuitive judgment of human experts? If the system is not "vigilant" enough at high speeds (Rt value is too low), then increase k. V If the system is too "sensitive" in low light (R...t If the value is too high (leading to frequent false alarms), then decrease k. L Through repeated iterative testing and adjustments, a set of k-values ​​that exhibit balanced and robust performance across various scenarios was finally obtained. L and k V value.

[0019] (II) Data-Driven Approach Based on Optimization Algorithms: Construct a large-scale, diverse test dataset. Each data point in the dataset includes sensor readings (images, point clouds, velocity, illumination) and a "true" risk level (e.g., from 1 to 10) labeled by a human safety expert. Define a loss function to measure the risk score R calculated by the system. t The gap between the actual risk level and the expert-assigned risk level. (k) L and k V (And the basic weights β1, β2, β3) are used as variables to be optimized. Algorithms such as grid search and Bayesian optimization are used to find a set of parameter values ​​that minimize the total loss across the entire dataset. This method is more objective and systematic.

[0020] (III) Based on reinforcement learning: k L and k V By incorporating a framework that trains basic weights through reinforcement learning and incorporating the comprehensive risk score as part of the reward function, the agent (guide dog) learns the optimal parameter strategy autonomously through extensive training in a simulated environment, maximizing its long-term task success rate and safety.

[0021] The substantial effects of this invention are: 1. Real-time quantification and comprehensive evaluation: This invention constructs a three-factor risk model that includes visual complexity, physical congestion, and environmental dynamics, achieving for the first time a real-time, quantitative evaluation of navigation environment risks. This far surpasses the binary judgment of traditional obstacle avoidance systems, and the evaluation results are more comprehensive and closer to human perception.

[0022] 2. Enhanced proactive perception capability: Through a graded, multimodal feedback mechanism (such as vibrations of different intensities and key voice prompts), this invention enables visually impaired users to clearly "perceive" the risk level of the environment ahead, thereby enabling them to proactively make predictive decisions such as slowing down and pausing, transforming passive following into active navigation, greatly improving safety and user confidence.

[0023] 3. Adaptability and High Robustness: The introduced dynamic weight adjustment mechanism enables the risk assessment model to adaptively adjust based on real-time changing contextual information such as lighting and speed. For example, it relies more on lidar in low light conditions and pays more attention to dynamic obstacles when traveling at high speeds, significantly improving the robustness and assessment accuracy of this solution in changing environments.

[0024] 4. High integration and ease of use: The method of this invention can be directly deployed on a guide dog platform, eliminating the need for users to carry additional complex equipment. The feedback device is designed as a wearable device, offering intuitive interaction that aligns with the usage habits of visually impaired users.

[0025] 5. Good scalability: The risk assessment framework of this invention is an open model, which can be easily integrated with more risk factors in the future, such as sound source localization (to determine the direction of oncoming vehicles), ground material recognition (to determine slippery road surfaces), road width recognition, etc., thereby continuously improving the risk model and making it more intelligent. Attached Figure Description

[0026] Figure 1 This is a flowchart of a dynamic environmental risk level assessment method for a guide robot dog according to the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0028] Example: This example provides a method for assessing the dynamic environmental risk level of a guide robot dog. The method is deployed on a quadruped guide robot dog (e.g., Unitree Robotics' Unitree Go1), which is equipped with at least: (1) a perception unit: including a depth camera (e.g., Intel RealSense D435, which can simultaneously provide RGB images and depth information), a 360° lidar (e.g., Slamtec RPLIDAR S2, which provides 2D / 3D point cloud data of the surrounding environment) and an inertial measurement unit (IMU, which provides the robot's posture and velocity information). (2) a computing unit: a high-performance onboard computer (e.g., NVIDIA Jetson series) for running the algorithm of this invention. (3) a feedback unit: a wearable vibration module (which can be worn on the wrist or waist belt) connected to the computing unit via Bluetooth and a text-to-speech (TTS) module for outputting feedback signals to the user.

[0029] like Figure 1 As shown, the detailed process of the method provided in this embodiment is as follows: S1: Data Acquisition: Executed cyclically at a preset frequency (e.g., 10Hz). At the beginning of each cycle, the computing unit acquires the latest sensor data from the sensing unit, primarily including RGB color images and depth images acquired by the forward-facing depth camera and 3D spatial point cloud data of the surrounding environment generated by LiDAR scanning. The robot's current linear velocity v is acquired by the IMU and wheeled odometry (if any). robot .

[0030] S2: Risk Factor Calculation: Calculate three core environmental risk factors in parallel or serially.

[0031] S2.1: Visual scene information entropy H v The calculation of this factor aims to quantify the visual complexity of a scene within the user's field of vision. A scene with rich texture and complex edges (such as a commercial street full of billboards) requires more cognitive resources from the user than a scene with a simple structure (such as an empty corridor), and therefore carries a higher risk. The calculation steps are as follows: S211: Image preprocessing: Convert the real-time acquired RGB color image into a grayscale image to eliminate interference from color information and focus on the structural and textural features of the scene.

[0032] S212: Edge Detection: Apply an edge detection operator (such as the Canny or Sobel operator) to a grayscale image to generate a binarized edge image. In this image, complex structures such as the outline and texture details of objects are highlighted as white pixels (edges).

[0033] S213: Blocking and Proportion Calculation: Divide the edge image into N×M non-overlapping image blocks (e.g., N=8, M=8, a total of 64 blocks). Then, for the i-th image block, calculate the proportion of white pixels (edge ​​pixels) to the total number of pixels in the block, denoted as p. i .

[0034] S214: Information Entropy Calculation: Calculate the information entropy H of the entire visual scene according to the Shannon entropy formula. v : ; According to the definition of information entropy, if a certain p i Then the item p i log2(p i The value is 0. v The higher the value, the richer the edges and textures of the scene, the more complex the structure, and the higher the potential cognitive load and risk.

[0035] S2.2: Obstacle density D o The calculation of this factor: This factor is used to quantify the crowding level of physical space. The calculation steps are as follows: S221: Define Region of Interest (AOI): Define a key detection area in the 3D point cloud data in front of the robot dog. For example, a fan-shaped or rectangular area with a radius of 5 meters and a 120-degree field of view in front. This area is where the user will enter, and its congestion level directly affects the ease of passage.

[0036] S222: Ground Segmentation and Obstacle Clustering: First, algorithms such as RANSAC (Random Sample Consensus) are used to perform planar fitting on the point cloud within the AOI, identifying and removing ground point clouds. Then, clustering algorithms, such as DBSCAN (Density-Based Spatial Clustering) or Euclidean distance clustering, are applied to the remaining non-ground point clouds. Spatially adjacent point clouds are aggregated into independent clusters, each cluster representing a potential obstacle (such as pedestrians, trees, pillars, etc.).

[0037] S223: Density Calculation: Count the total number N of obstacle clusters formed within the AOI. obstacles Obstacle density D o Defined as the number of obstacles per unit area: D o =N obstacles / Area AOI Area AOI It is the area of ​​the region of interest. D o The higher the value, the more crowded the environment ahead, requiring more frequent obstacle avoidance and resulting in higher risk.

[0038] S2.3: Environmental dynamic change rate C d The calculation of this factor measures the dynamics of the environment, specifically the inflow and outflow rates of objects. A rapidly changing environment (such as a busy intersection) poses a higher risk than a static environment (such as a room filled with furniture). The calculation steps are as follows: S231: Between consecutive sampling moments, the number of newly appearing obstacle clusters and the number of instantly disappearing obstacle clusters are determined by tracking the obstacle clusters; a Kalman filter can be used for state prediction and updating, or a simple IOU (Intersection over Union) matching can be used to associate the same obstacle in two consecutive frames; in each processing cycle of the system (i.e., between the current timestamp t and the previous timestamp t-1), the system compares the obstacle cluster list using a target tracking algorithm to identify the number of newly appearing targets within this extremely short sampling interval, denoted as n. new (t), and the number of targets that disappear instantaneously, denoted as n. disappeared (t). These two values ​​reflect immediate changes in the environment; S232: Within a preset time window, the number of all instantaneously newly appearing obstacle clusters and the number of instantaneously disappearing obstacle clusters determined in step S231 will be accumulated to obtain a total change. To smoothly process and capture the overall dynamic trend over a period of time, the system employs a sliding window mechanism. First, the total number of frames W within the window is defined, determined by the preset time window length Δt (e.g., 1.0 seconds) and the system's sampling period Ts (e.g., 0.1 seconds): W = Δt / T sThen, the system sums up the instantaneous changes over the past W frames to obtain the total number N of newly appearing targets within the time window Δt. new Total number of targets that disappeared N disappeared The calculation formula is as follows: N new =Σ i=0 W-1 n new (ti); N disappeared =Σ i=0 W-1 n disappeared (ti); The above formula means: starting from the current time t, backtrack W frames (i from 0 to W-1), and sum up all the instantaneous new occurrences (or disappearances) detected in each of these W frames to obtain a total change within the recent historical window; S233: Divide the total change accumulated within the window period by the length of the time window Δt to calculate the smoothed environmental dynamic change rate C. d The calculation formula is as follows: C d =(N new +N disappeared ) / Δt; This value represents the total number of targets whose environment changes on average per second over a period of time. It can measure the dynamics of the environment, such as the entry and exit of people and the departure of vehicles. C d The higher the value, the more active the dynamic factors such as pedestrian and vehicle traffic, the more unpredictable the environment, and the higher the risk.

[0039] S3: Overall Score: This step integrates the three independent risk factors mentioned above into a single, comprehensive risk score R. t To make the model more adaptable, this embodiment employs a dynamic weighting mechanism.

[0040] S31: Obtaining Basic Weights: First, obtain a set of basic weights β={β1,β2,β3} offline. This can be achieved by training a model (e.g., linear regression) on a dataset containing a large amount of scene data and expert risk annotations, or by using reinforcement learning to allow the robot to learn the optimal weight allocation in a simulated environment. This set of basic weights represents the relative importance of each risk factor under "standard" conditions (e.g., good lighting, medium speed). For example, through learning, we might obtain β1=0.6, β2=0.3, β3=0.1.

[0041] S32: Dynamic weight adjustment: The system obtains context information in real time and uses this information to adjust the basic weights to obtain the final dynamic weights used for calculation.

[0042] This embodiment considers two context adjustment factors: (1) Ambient light intensity L: can be obtained by calculating the average pixel value of the camera's grayscale image. (2) Robot's own speed v robot : Obtained from IMU or odometer.

[0043] The formula for calculating dynamic weights is as follows: β1'=β1·(1+k L ·max(0,L threshold -L)); β2'=β2·(1+k v ·v robot ); β3'=β3·(1+k v ·v robot ); Formula Explanation and Parameter Acquisition: The first formula means: when the ambient light intensity L is lower than the preset threshold L... threshold When there is insufficient light, the visual information entropy H v While the reliability of LiDAR decreases, its importance needs to be increased (because ambiguous and complex environments are more dangerous), and the physical obstacles detected by LiDAR also need to be addressed. o and C d We should also be more vigilant. Therefore, we should increase the weight of β1'. The max function ensures that the weight only increases when there is insufficient light. The meaning of the second and third formulas is: when the robot dog's speed v robot The faster the obstacle, the shorter the reaction time to surrounding obstacles (especially dynamic obstacles), and therefore the greater the potential risk. Therefore, it is necessary to increase the obstacle density D accordingly. o and the rate of dynamic change of the environment C d The weights β2' and β3'.

[0044] Obtaining hyperparameters: L threshold (Light intensity threshold): This can be calibrated experimentally. Run the edge detection algorithm under different lighting conditions, observe the average image brightness when its performance begins to decline significantly, and set this value as the threshold. For example, it can be set to 60 (within the range of 0-255).

[0045] k L ,k V (Adjusting Sensitivity Coefficients): These two coefficients control the drasticness of weight adjustments. Optimization can be achieved through grid search on labeled datasets, or through expert manual tuning and repeated testing in real-world scenarios, until the system's response is neither too sensitive nor too sluggish. For example, k can be set... L =0.5,k V =0.8.

[0046] S33: Final Score Calculation: Using adjusted dynamic weights, calculate the final comprehensive environmental risk score R. t : R t =β1'·H v +β2'·D o +β3'·C d ; S4: Feedback Generation: This step generates continuous numerical scores R. t The feedback signal is converted into a discrete, user-friendly representation. Risk level classification: The system presets two scoring thresholds, a first threshold θ1 and a second threshold θ2 (e.g., θ1=1.2, θ2=2.0). Based on R... t The value of R divides the environment into three risk levels: Green level (safe): if R t <θ1 indicates the current environment is safe and passage is permitted. Yellow level (medium risk): If θ1 ≤ R t ≤θ2 indicates that the environment ahead is becoming more complex or crowded, requiring increased vigilance. Red level (high risk): If R t ≥θ2 indicates that the environment ahead is highly complex, crowded, or dynamically changing, posing a high risk, and it is not recommended to proceed directly.

[0047] Multimodal Feedback: Based on the determined risk level, the feedback unit provides users with different modes of feedback: Green Level: No feedback, or a slight, short vibration from the wearable device, indicating "all is normal". Yellow Level: Moderate, intermittent vibration from the wearable device, along with a brief voice prompt, such as "Complex ahead, please be careful". Red Level: Strong, continuous vibration from the wearable device, along with clear action suggestions from the voice module, such as "High risk, detour recommended".

[0048] Real-world scenario example: When a user is crossing a square: at the edge of the square where there are fewer people, the system calculates R... t The value is low, at the green level, and users hardly perceive any feedback. As you approach the center of the square, the flow of people increases, and the visual scene becomes more complex due to the buildings and crowds. t As the value rises to the yellow level, the user's wristband begins to vibrate intermittently, and a "complex ahead" warning is heard, causing them to subconsciously slow down. When a group of children suddenly appears directly in front, causing D... o and C d R rises sharply t If the value instantly exceeds the red threshold, the system immediately issues a strong vibration and a voice alert suggesting a detour, guiding the user to avoid the high-risk area and thus preventing a potential collision.

[0049] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0050] Although this document uses a variety of terms, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of the invention; interpreting them as any additional limitation would contradict the spirit of the invention.

Claims

1. A method for assessing the dynamic environmental risk level of a guide dog, characterized in that, Includes the following steps: S1: Data Acquisition: Acquire sensor data of the guide dog, wherein the sensor data includes at least visual image data and three-dimensional spatial point cloud data; S2: Risk Factor Calculation: Based on the sensor data, calculate various environmental risk factors, including: Visual scene information entropy calculated based on the aforementioned visual image data; Obstacle density calculated based on the aforementioned three-dimensional spatial point cloud data; The environmental dynamic change rate is calculated based on the three-dimensional spatial point cloud data over a continuous time period. S3: Comprehensive Score: The comprehensive environmental risk score is obtained by weighted summation of all environmental risk factors. S4: Feedback Generation: Based on the comprehensive environmental risk score, determine the risk level of the current environment and generate a user feedback signal corresponding to the risk level to be fed back to the user.

2. The method for assessing the dynamic environmental risk level of a guide dog according to claim 1, characterized in that, The steps for calculating the entropy of visual scene information are as follows: S211: Convert the real-time acquired visual image data into a grayscale image; S212: Perform edge detection on the grayscale image to generate a binarized edge image; S213: Divide the edge image into N×M non-overlapping image blocks and calculate the proportion of edge pixels in each image block; S214: Calculate the visual field information entropy H using the following formula. v : ； In the formula, p i denoted as the proportion of edge pixels in the i-th image block.

3. The method for assessing the dynamic environmental risk level of a guide dog according to claim 2, characterized in that, The steps for calculating the obstacle density are as follows: S221: Define the region of interest in three-dimensional point cloud data; S222: Divide the point cloud within the region of interest into ground point cloud and non-ground point cloud, and cluster the non-ground point cloud to obtain several obstacle clusters; S223: Calculate the number of obstacle clusters per unit area in the region of interest, which is the obstacle density D. o .

4. The method for assessing the dynamic environmental risk level of a guide dog according to claim 3, characterized in that, The steps for calculating the dynamic rate of change of the environment include: S231: Between consecutive sampling moments, by tracking the obstacle clusters, determine the number of newly appearing obstacle clusters and the number of instantly disappearing obstacle clusters; S232: Within a preset time window, the number of all instantaneously newly appearing obstacle clusters and the number of instantaneously disappearing obstacle clusters determined in step S231 will be accumulated to obtain a total change. S233: Divide the total change by the length of the preset time window to obtain the environmental dynamic change rate C. d .

5. The method for assessing the dynamic environmental risk level of a guide dog according to claim 4, characterized in that, The weights used in the weighted summation in step S3 are dynamic weights, which are determined in the following way: S31: Through offline datasets or reinforcement learning, a set of basic weights β={β1,β2,β3} is obtained by training, where β1 is the basic weight of visual scene information entropy, β2 is the basic weight of visual scene information entropy, and β3 is the basic weight of visual scene information entropy. S32: The final dynamic weights are determined based on the base weights and the context adjustment factor, using the following formula: β1’=β1·(1+k L ·max(0,L threshold -L)); β2’=β2·(1+k v ·v robot ); β3’=β3·(1+k v ·v robot ); In the formula, β1' is the dynamic weight of the visual scene information entropy, β2' is the dynamic weight of the visual scene information entropy, β3' is the dynamic weight of the visual scene information entropy, and k L k is the sensitivity coefficient for adjusting ambient light. v L is the sensitivity coefficient for adjusting the speed of the guide dog, and L is the current ambient light intensity. threshold v is the preset light intensity threshold. robot This refers to the current speed of the guide dog. Comprehensive Environmental Risk Score R t =β1'·H v +β2'·D o +β3'·C d .

6. A method for assessing the dynamic environmental risk level of a guide dog according to claim 1 or 5, characterized in that, In step S4, if the comprehensive environmental risk score is less than the first threshold θ1, the risk level is green, indicating that the current environment is safe to pass through, and the user feedback is no feedback or a slight vibration prompt. If the comprehensive environmental risk score is greater than or equal to the first threshold θ1 and less than the second threshold θ2, the risk level is yellow, indicating that the current environment has a medium risk and caution is required. User feedback includes a medium vibration and a voice prompt "complex ahead". If the comprehensive environmental risk score is greater than or equal to the second threshold θ2, the risk level is red, indicating that the current environment is high-risk and needs to be avoided. User feedback will include strong vibrations and a voice prompt "It is recommended to detour".

Images