A scene recognition method for robot environment adaptive navigation

Abstract

The application belongs to the technical field of robots, and discloses a scene recognition method for robot environment adaptive navigation. The scene recognition method is based on a data acquisition unit, a data storage unit and a data processing unit. The data processing unit is located on a host computer, sensors are arranged on a robot, data collected by the sensors is uploaded to the host computer, the host computer extracts environment features, then uses a fuzzy decision system to judge the environment where the robot is currently located, judges which one of wide, moderate, narrow and impassable the environment where the robot is currently located belongs to, and finally automatically switches a parameter group of a dynamic window algorithm DWA into a parameter group corresponding to the scene. The scene recognition method solves the problems of incomplete environment perception, parameter adjustment lag and low calculation efficiency, realizes intelligent cooperation of scene judgment of the robot in a complex dynamic environment and the parameter group of the dynamic window algorithm DWA, and has engineering practical value.

Description

Technical Field

[0001] This invention belongs to the field of robotics technology, specifically relating to a scene recognition method for adaptive navigation of robots. Background Technology

[0002] The rapid development of robotics technology has driven the widespread application of autonomous mobile robots in industries such as manufacturing, logistics, and healthcare. Path planning, as a core component of robot navigation, directly impacts system performance in terms of efficiency and safety. Traditional path planning algorithms, such as the Dynamic Window Approach (DWA), perform stably in structured environments. However, in complex dynamic scenarios, a single environmental perception method and static parameter configuration are insufficient to meet the robot's requirements for environmental adaptability and decision-making robustness. DWA, as a typical local planning method, generates candidate trajectories by sampling the velocity space and uses an evaluation function to assess safety. However, its fixed parameter settings lead to insufficient adaptability in different scenarios.

[0003] In terms of environmental perception and parameter adjustment technologies, existing methods generally suffer from incomplete feature extraction and lagging strategies. While LiDAR perception can acquire obstacle distance data and extract the nearest obstacle distance MOD, a single indicator cannot reflect the openness of the scene and the uniformity of obstacle distribution. For example, in an open area where obstacles are concentrated on one side, relying solely on the nearest obstacle distance MOD may lead to the robot misjudging drivability. Although the fusion perception of vision and LiDAR can improve accuracy, it faces challenges such as complex data alignment and high computational resource consumption. In parameter adjustment strategies, fixed parameter configurations cause the robot to use the same speed and obstacle avoidance weights in both concentrated distances and open spaces, resulting in an imbalance between obstacle avoidance efficiency and movement speed. Rule switching based on thresholds set by human experience is susceptible to sensor noise, and frequent parameter oscillations reduce system stability. Although machine learning methods can dynamically optimize parameters, they rely on a large amount of high-quality training data, have limited model generalization ability, and are difficult to adapt to the diversity of unstructured environments.

[0004] The core shortcomings of existing technologies lie in their limited environmental perception dimensions, static parameter configurations, and low algorithmic efficiency. The single nearest obstacle distance (MOD) metric cannot quantify the uniformity of obstacle distribution in a given direction, causing robots to fail to accurately distinguish between densely packed obstacles and open spaces. Fixed dynamic window algorithm (DWA) parameter sets cannot dynamically adjust maximum linear velocity, angular velocity, and weighting coefficients based on the scene. For example, insufficient obstacle avoidance weights in narrow scenes may lead to collision risks, while speed limitations in wide scenes affect task execution efficiency. Furthermore, the DWA algorithm cannot feed local environmental information back to global planning, resulting in path failures. In addition, the high computational resource requirements of traditional methods limit their application in embedded mobile robots. For instance, deep learning-based parameter optimization models rely on GPU acceleration, making it difficult to meet real-time requirements.

[0005] Currently, there is an urgent need to develop a scene recognition method for adaptive navigation of robots. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a scene recognition method for robot environment adaptive navigation, so as to overcome the defects of the prior art.

[0007] The scene recognition method for adaptive navigation of robots according to the present invention can be roughly divided into the following steps: a. The sensors mounted on the robot collect the polar coordinates of surrounding obstacles in real time. The collected data is synchronously transmitted to the host computer data processing unit, where the polar coordinates are converted into Cartesian coordinates for subsequent geometric analysis. To suppress noise in single-frame data, a sliding window mean filtering algorithm is used to process the distance data, and the filtered distance value is calculated using a recursive formula. b. The host computer extracts features from the filtered effective points, calculates the nearest obstacle distance to assess the emergency obstacle avoidance needs, and then calculates the obstacle density within the effective radius around the robot as the obstacle density. c. The host computer extracts the nearest obstacle distance and obstacle density, and uses a fuzzy decision system for scene discrimination. First, the nearest obstacle distance and obstacle density are fuzzified using a trapezoidal membership function; then, a fuzzy inference rule base is constructed, and the activation strength of each fuzzy inference rule is calculated using the minimum operator; finally, the centroid method is used for defuzzification, converting the fuzzy output into explicit scene level values. The quantized values ​​of the four scene levels correspond to "impassable", "narrow", "moderate", and "wide", respectively. d. Based on the scene level value output by the fuzzy decision, the host computer retrieves the corresponding parameter set (including maximum linear velocity, maximum angular velocity, obstacle avoidance weight, target approach weight, etc.) of the Dynamic Window Algorithm (DWA) from the preset scene-parameter mapping table. To avoid frequent parameter switching caused by sensor noise, a hysteresis switching mechanism is introduced, triggering parameter set switching only when the same scene is determined for N consecutive frames. After parameter set switching, the robot operates according to the new parameter set while continuing to execute steps a~c, forming a closed-loop control of "data acquisition - feature extraction - fuzzy decision - parameter switching". If the scene level is continuously stable, the current parameter set is maintained; if a scene change is detected and the hysteresis condition is met, the parameter set is updated to ensure that the robot can perform path planning with optimal parameters in different environments, balancing obstacle avoidance safety and movement efficiency.

[0008] The scene recognition method for adaptive navigation of robots in this invention is based on a data acquisition unit, a data storage unit, and a data processing unit. The data processing unit is located on a host computer. Sensors are installed on the robot, and the data collected by the sensors is uploaded to the host computer. The host computer extracts environmental features and then uses a fuzzy decision system to determine the robot's current environment, classifying it as an impassable, narrow, moderate, or wide scene. Finally, the parameter set of the Dynamic Window Algorithm (DWA) is automatically switched to the parameter set corresponding to the scene.

[0009] Furthermore, the sensors include lidar and ultrasonic sensors.

[0010] Furthermore, the process of extracting environmental features is as follows: S31. Collect information about surrounding obstacles using sensors mounted on the robot; S32. Perform coordinate transformation; The sensor collects the polar coordinates of surrounding obstacles. The data is simultaneously transmitted to the host computer's data processing unit; the polar coordinates are converted to Cartesian coordinates. ; The sensor scan of the first Radial distance between points (unit: meters); The sensor scan of the first Angle of a point (unit: degrees, 0° points directly in front of the robot); : Coordinates in the transformed Cartesian coordinate system (unit: meters); S33. Perform sliding window mean filtering; The Exponentially Weighted Moving Average (EWMA) algorithm is used to suppress instantaneous noise in single-frame data; the EWMA algorithm recursively calculates the filtered distance value. ; : No. The first frame The original distance values ​​of each point; : No. The filtered distance value of the frame; Forgetting factor, ranging from 0 to 1, is used to weigh the importance of current measurements against historical data; when When the value is close to 1, the Exponentially Weighted Moving Average (EWMA) algorithm focuses on the current measurement and reacts quickly to real-time changes in the data, but it introduces more noise; when... When the value is close to 0, the exponentially weighted moving average (EWMA) algorithm relies more on historical data and has a better smoothing effect on noise, but it is slower to respond to real-time changes in data. S34. Calculate the distance to the nearest obstacle. ; Calculate the Euclidean distance from the robot to the nearest obstacle to assess whether the robot is in a dangerous state requiring emergency obstacle avoidance; for all filtered valid points... In the algorithm, with the robot's position as the origin, the Euclidean distance from each point to the robot's position is calculated. Take the minimum value as the nearest obstacle distance. ; ; : effective points Euclidean distance to the center of the robot; This represents all valid points after filtering; S35. Calculate the obstacle density around the robot. ; Calculate the obstacle density within the effective radius R around the robot. Obstacle density is defined by an indicator function. : ; in, This is an indicator function, which is 1 when the condition is met and 0 otherwise; the closer the indicator function value is to 1, the denser the distribution of obstacles.

[0011] Furthermore, the fuzzy decision-making system includes the following: S41. Calculate the fuzzy membership function to obtain the precise nearest obstacle distance. Obstacle density Membership degree is transformed into a fuzzy concept; closest obstacle distance trapezoidal function : ; Nearest obstacle distance Actual input value (unit: meters); : Indicates the distance to the nearest obstacle Membership degree of a "very close" fuzzy set, range ; closest obstacle distance trapezoidal function : ; : Indicates the distance to the nearest obstacle Membership degree and range of "near" fuzzy sets ; closest obstacle distance trapezoidal function : ; : Indicates the distance to the nearest obstacle Membership degree and range of a "medium" fuzzy set ; closest obstacle distance Trapezoidal membership function : ; : Indicates the distance to the nearest obstacle Membership degree and range of the "far" fuzzy set ; obstacle density Trapezoidal membership function : ; Obstacle density Actual input value, unit: pieces / m²; : Represents obstacle density Membership degree of a "low" fuzzy set, range ; obstacle density Trapezoidal membership function : ; : representing obstacle density Membership degree belonging to the "medium" fuzzy set, range ; Obstacle density Trapezoidal membership function of : ; : representing obstacle density Membership degree belonging to the "high" fuzzy set, range ; S42. Set fuzzy inference rules and calculate the activation strength of fuzzy inference rules; Perform fuzzy inference. After the fuzzification of input variables, construct a fuzzy rule base and calculate the activation strength of each fuzzy inference rule through the minimum operator to determine the influence degree of each fuzzy inference rule on the decision under the current input conditions; The fuzzy inference rule base is a multi-input single-output system. The inputs are the fuzzy values of the nearest obstacle distance and obstacle density , and the output is the scene level value; the fuzzy inference rule is: if the nearest obstacle distance is very close and is high density, then the scene level is impassable; where, the activation strength of the th fuzzy inference rule ; : Membership degree of the nearest obstacle distance belonging to the fuzzy set ; : Membership degree of obstacle density belonging to the fuzzy set ; S43. Defuzzify and convert the fuzzy output into a definite scene level value; What is obtained through fuzzy inference is the fuzzy value of the scene level. The fuzzy value of the scene level is the fuzzy output. By defuzzifying, the fuzzy value of the scene level is converted into a definite scene level value so that the robot can make a specific decision; Use the centroid method for defuzzification, perform weighted summation of the membership degrees of different scene levels and the corresponding quantization values, and then divide by the sum of membership degrees to obtain the scene level value: ; : Scene level Membership degree; : Quantitative value of scene level These correspond to scenarios that are impassable, narrow, moderate, and wide, respectively. S44. Establish a delayed switching mechanism; Improve decision stability by delaying confirmation; prevent frequent parameter switching by switching conditions; switching conditions: continuous. The frames were determined to be from the same scene; The threshold for the number of consecutive frames; when consecutive When the deblurring results of the frames are all at the same scene level, the parameter set of the Dynamic Window Algorithm (DWA) is switched to the parameter set of the corresponding scene to suppress high-frequency noise interference.

[0012] Furthermore, the parameter group switching method for the Dynamic Window Algorithm (DWA) includes the following: S51. The host computer receives the quantized value of the scene level output by the fuzzy decision system. It uses a pre-defined scene-parameter mapping table for fast indexing; the scene-parameter mapping table is stored in key-value pairs, where the key is the scene level and the value is the parameter set of the corresponding Dynamic Window Algorithm (DWA); the mapping logic of the scene-parameter mapping table follows the matching principle between scene characteristics and control strategies: For low-level scenarios with a quantization value of 1 or 2, parameter sets with low speed and high obstacle avoidance weights are matched first to restrict robot movement and ensure safety; for high-level scenarios with a quantization value of 3 or 4, parameter sets with high speed and low steering sensitivity are matched to unleash robot movement capabilities and improve robot movement efficiency. S52. To suppress high-frequency parameter group switching caused by sensor noise or transient environmental changes, a continuous frame count confirmation mechanism is introduced: when the quantization value of the scene level is the same for N consecutive frames, the scene is considered stable and parameter group switching is triggered; if the quantization value of the scene level changes in any frame before N frames, the counter is reset and counting starts again until the condition of N consecutive frames is met; the continuous frame count confirmation mechanism ensures that parameter group switching only occurs when the scene is continuously stable through redundant judgment in the time dimension, avoiding control quantity oscillation caused by accidental noise; S53. The host computer sends the target parameter set to the Dynamic Window Algorithm (DWA) execution module through standardized communication interfaces, including CAN bus, serial port, or ROS parameter server, triggering the initialization process: S531. Stop the current path planning, interrupt the running local path planning thread, clear historical trajectory data and motion command queue to avoid control conflicts caused by the mixing of old and new parameter groups; S532. Load new parameter set: Write the parameter set, including maximum linear velocity, maximum angular velocity, obstacle avoidance weight, and target approach weight, into the core configuration module of the Dynamic Window Algorithm (DWA), and update the feasible speed window boundary of the speed planner and the weight coefficient of the steering controller. S533. Regenerate Local Path: Starting from the robot's current pose, and combining the motion constraints defined by the new parameter set, based on the real-time environment map, recalculate the set of feasible trajectories containing velocity and angular velocity, regenerate the local path, and select the optimal path to output to the actuator.

[0013] Furthermore, the parameters in the parameter group of the Dynamic Window Algorithm (DWA) are modified according to the robot's dynamic characteristics and task requirements. The maximum linear velocity in narrow scenes is set to 0.3 m / s and the obstacle avoidance weight is set to 0.8. For heavy-duty robots, the maximum linear velocity threshold is reduced, and for high-speed moving scenes, the obstacle avoidance weight is adjusted to balance obstacle avoidance and efficiency.

[0014] The scene recognition method for adaptive navigation of robots of the present invention has the following characteristics: 1. Obstacle data is collected in real time by sensors to construct a feature system that includes the nearest obstacle distance and obstacle density. The nearest obstacle distance reflects the emergency obstacle avoidance needs, and the obstacle density reflects the usability of the environment. The two work together to achieve a comprehensive characterization of the scene complexity. 2. Fuzzy reasoning is adopted to transform the precise values ​​of the nearest obstacle distance and obstacle density into four fuzzy concepts of "impassable", "narrow", "moderate" and "wide" through trapezoidal membership functions, and a fuzzy rule base containing 9 fuzzy reasoning rules is constructed. The centroid method is used to defuzzify and generate parameter switching instructions to realize the dynamic adjustment of parameters such as maximum linear velocity, maximum angular velocity, obstacle avoidance weight, and target approach weight in the parameter group of the Dynamic Window Algorithm (DWA). 3. A hysteresis switching mechanism for continuous N-frame stability determination is introduced to avoid frequent parameter oscillations caused by sensor noise; at the same time, a closed-loop control process of "data acquisition - feature extraction - fuzzy decision - parameter switching" is established to ensure that the robot can respond quickly and maintain parameter stability when the scene changes.

[0015] This invention presents a scene recognition method for robot environmental adaptive navigation, proposing a parameter adaptation technique driven by multi-dimensional environmental feature fusion and fuzzy decision-making. By introducing obstacle density and combining it with the nearest obstacle distance to construct a two-dimensional feature system, the complexity of the scene is comprehensively characterized. Fuzzy membership functions are used to transform precise environmental features into fuzzy concepts, which are then mapped to four scene categories—"impassable," "narrow," "moderate," and "wide"—through a fuzzy rule base. A centroid method is used to defuzzify and generate parameter switching instructions. Simultaneously, a hysteresis switching mechanism is designed to avoid noise interference through stable multi-frame judgment, improving decision stability. This method solves the problems of incomplete environmental perception, lag in parameter adjustment, and low computational efficiency, achieving intelligent collaboration between scene judgment and parameter groups of the Dynamic Window Algorithm (DWA) in complex dynamic environments, and has practical engineering value. Attached Figure Description

[0016] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0017] Figure 1 This is a flowchart of the scene recognition method for adaptive navigation of robots according to the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, the scene recognition method for robot environment adaptive navigation of the present invention is based on a data acquisition unit, a data storage unit, and a data processing unit. The data processing unit is located on the host computer. Sensors are installed on the robot. The data collected by the sensors is uploaded to the host computer. The host computer extracts environmental features and then uses a fuzzy decision system to judge the current environment of the robot. It determines whether the current environment of the robot belongs to the category of impassable, narrow, moderate, or wide scene. Finally, the parameter group of the Dynamic Window Algorithm (DWA) is automatically switched to the parameter group of the corresponding scene.

[0020] Furthermore, the sensors include lidar and ultrasonic sensors; single-line lidar is used in low-cost scenarios, while visual sensors are added for fusion perception in high-precision scenarios.

[0021] Furthermore, the process of extracting environmental features is as follows: S31. Collect information about surrounding obstacles using sensors mounted on the robot; S32. Perform coordinate transformation; The sensor collects the polar coordinates of surrounding obstacles. The data is simultaneously transmitted to the host computer's data processing unit. Polar coordinates are important for intuitively describing the distance and direction relationship between the robot and surrounding obstacles, but Cartesian coordinates are often more convenient for subsequent complex geometric analysis and calculations. Therefore, it is necessary to convert polar coordinates to Cartesian coordinates using the Cartesian coordinate transformation formula: ; The sensor scan of the first Radial distance between points (unit: meters); The sensor scan of the first Angle of a point (unit: degrees, 0° points directly in front of the robot); : Coordinates in the transformed Cartesian coordinate system (unit: meters); S33. Perform sliding window mean filtering; The Exponentially Weighted Moving Average (EWMA) algorithm is employed to suppress instantaneous noise in single-frame data. During sensor data acquisition, various noises are unavoidable, potentially stemming from sensor measurement errors or electromagnetic interference from the surrounding environment. This noise causes fluctuations and uncertainties in the acquired data, severely impacting the accurate extraction and analysis of environmental features. To address this issue, EWMA filtering is used. The EWMA algorithm recursively calculates the filtered distance value: ; : No. The first frame The original distance values ​​of each point; : No. The filtered distance value of the frame; Forgetting factor, ranging from 0 to 1, is used to weigh the importance of current measurements against historical data; when When the value is close to 1, the Exponentially Weighted Moving Average (EWMA) algorithm focuses on the current measurement and reacts quickly to real-time changes in the data, but it introduces more noise; when... When the value is close to 0, the exponentially weighted moving average (EWMA) algorithm relies more on historical data and has a better smoothing effect on noise, but it is slower to respond to real-time changes in data. S34. Calculate the distance to the nearest obstacle. ; Calculate the Euclidean distance from the robot to the nearest obstacle to assess whether the robot is in a dangerous state requiring emergency obstacle avoidance; for all filtered valid points... In the algorithm, with the robot's position as the origin, the Euclidean distance from each point to the robot's position is calculated. Take the minimum value as the nearest obstacle distance. ; ; : effective points Euclidean distance to the center of the robot; This represents all valid points after filtering; S35. Calculate the obstacle density around the robot. ; Calculate the obstacle density within the effective radius R around the robot. Obstacle density is defined by an indicator function. : ; in, This is an indicator function, which is 1 when the condition is met and 0 otherwise; the closer the indicator function value is to 1, the denser the distribution of obstacles.

[0022] Furthermore, the fuzzy decision-making system includes the following: S41. Calculate the fuzzy membership function to obtain the precise nearest obstacle distance. Obstacle density Membership degree is transformed into a fuzzy concept; closest obstacle distance trapezoidal function : ; Nearest obstacle distance Actual input value (unit: meters); : Indicates the distance to the nearest obstacle Membership degree of a "very close" fuzzy set, range ; closest obstacle distance trapezoidal function : ; : Indicates the distance to the nearest obstacle Membership degree and range of "near" fuzzy sets ; closest obstacle distance trapezoidal function : ; : Represents the nearest obstacle distance Degree of membership belonging to the "medium" fuzzy set, range ; Nearest obstacle distance Trapezoidal membership function of : ; : Represents the nearest obstacle distance Degree of membership belonging to the "far" fuzzy set, range ; Obstacle density Trapezoidal membership function of : ; : Obstacle density Actual input value of, unit: number / m²; : Represents obstacle density Degree of membership belonging to the "low" fuzzy set, range ; Obstacle density Trapezoidal membership function of : ; : Represents obstacle density Degree of membership belonging to the "medium" fuzzy set, range ; Obstacle density Trapezoidal membership function of : ; : Represents obstacle density Degree of membership belonging to the "high" fuzzy set, range ; S42. Set fuzzy inference rules and calculate the activation strength of fuzzy inference rules; Perform fuzzy inference. After the fuzzification of input variables, construct the fuzzy rule base shown in Table 1, and through the minimum operator Logically calculate the activation strength of each fuzzy inference rule to determine the influence degree of each fuzzy inference rule on the decision under the current input conditions; The fuzzy rule base is a multi-input single-output system, and the inputs are the nearest obstacle distance 、obstacle density The fuzzy value is output as the scene level value; the fuzzy inference rule is: if the nearest obstacle distance is... It is very close, and If it is high density, then the scene level cannot be approved; Among them, the Activation strength of fuzzy inference rules : ; Nearest obstacle distance Belongs to fuzzy set Membership degree; Obstacle density Belongs to fuzzy set Membership degree; Table 1 Fuzzy Rule Base ; S43. Deblurring converts the blurred output into a clear scene level value; The fuzzy values ​​obtained through fuzzy reasoning are fuzzy values ​​at the scene level. These fuzzy values ​​at the scene level are fuzzy outputs. Defuzzification transforms these fuzzy values ​​at the scene level into explicit scene level values, enabling the robot to make specific decisions. The centroid method is used for deblurring. The membership degrees of different scene levels are weighted and summed with their corresponding quantized values, and then divided by the sum of membership degrees to obtain the scene level value. ; Scene Level Membership degree; : Quantitative value of scene level These correspond to scenarios that are impassable, narrow, moderate, and wide, respectively. S44. Establish a delayed switching mechanism; Improve decision stability by delaying confirmation; prevent frequent parameter switching by switching conditions; switching conditions: continuous. The frames were determined to be from the same scene; For the consecutive frame number threshold, take When continuous When the deblurring results of the frames are all at the same scene level, the parameter set of the Dynamic Window Algorithm (DWA) is switched to the parameter set of the corresponding scene to suppress high-frequency noise interference.

[0023] Furthermore, the parameter group switching method of the Dynamic Window Algorithm (DWA) includes the following: S51. The host computer receives the quantized value of the scene level output by the fuzzy decision system. The scene-parameter mapping table shown in Table 2 is used for fast indexing. The scene-parameter mapping table is stored in key-value pair format, where the key is the scene level and the value is the parameter group of the corresponding Dynamic Window Algorithm (DWA). The mapping logic of the scene-parameter mapping table follows the matching principle between scene characteristics and control strategies. For low-level scenarios with a quantization value of 1 or 2, parameter sets with low speed and high obstacle avoidance weights are matched first to restrict robot movement and ensure safety; for high-level scenarios with a quantization value of 3 or 4, parameter sets with high speed and low steering sensitivity are matched to unleash robot movement capabilities and improve robot movement efficiency. Table 2 Scene-Parameter Mapping Table ; S52. To suppress high-frequency parameter group switching caused by sensor noise or transient environmental changes, a continuous frame count confirmation mechanism is introduced: when the quantization value of the scene level is the same for N consecutive frames, the scene is considered stable and parameter group switching is triggered; if the quantization value of the scene level changes in any frame before N frames, the counter is reset and counting starts again until the condition of N consecutive frames is met; the continuous frame count confirmation mechanism ensures that parameter group switching only occurs when the scene is continuously stable through redundant judgment in the time dimension, avoiding control quantity oscillation caused by accidental noise; S53. The host computer sends the target parameter set to the Dynamic Window Algorithm (DWA) execution module through standardized communication interfaces, including CAN bus, serial port, or ROS parameter server, triggering the initialization process: S531. Stop the current path planning, interrupt the running local path planning thread, clear historical trajectory data and motion command queue to avoid control conflicts caused by the mixing of old and new parameter groups; S532. Load new parameter set: Write the parameter set, including maximum linear velocity, maximum angular velocity, obstacle avoidance weight, and target approach weight, into the core configuration module of the Dynamic Window Algorithm (DWA), and update the feasible speed window boundary of the speed planner and the weight coefficient of the steering controller. S533. Regenerate Local Path: Starting from the robot's current pose, and combining the motion constraints defined by the new parameter set, based on the real-time environment map, recalculate the set of feasible trajectories containing velocity and angular velocity, regenerate the local path, and select the optimal path to output to the actuator.

[0024] Furthermore, the parameters in the parameter group of the Dynamic Window Algorithm (DWA) are modified according to the robot's dynamic characteristics and task requirements. The maximum linear velocity in narrow scenes is set to 0.3 m / s and the obstacle avoidance weight is set to 0.8. For heavy-duty robots, the maximum linear velocity threshold is reduced, and for high-speed moving scenes, the obstacle avoidance weight is adjusted to balance obstacle avoidance and efficiency.

[0025] Example: This example uses the autonomous navigation of a power distribution room inspection robot. It is assumed that the robot's preset movement state is to move along a predetermined route in an indoor environment. The robot is equipped with a LiDAR sensor, and the host computer is a Raspberry Pi 4B. Sensor data is transmitted to the Raspberry Pi 4B via a USB interface. Suppose the robot encounters a scene with relatively concentrated data points and needs to switch parameters to ensure safe passage. The specific process is as follows: 1. Set the initial navigation parameters on the host computer; The initial navigation parameters for the robot are set on the Raspberry Pi 4B. These initial navigation parameters are a general set of parameters suitable for open environments. The Raspberry Pi 4B sends the initial navigation parameters to the Dynamic Window Algorithm (DWA) execution module. The DWA execution module loads the initial navigation parameters, generates an initial local path starting from the current robot pose, and controls the robot to begin moving. Simultaneously, the Raspberry Pi 4B stores the initial navigation parameters for subsequent comparison and switching decisions. 2. Sensor data acquisition; The robot is equipped with a lidar sensor to collect the polar coordinates of surrounding obstacles. For ease of calculation, 10 valid data points were selected, as shown in Table 3 below: Table 3 Polar coordinates of obstacles ; Data is transferred to the Raspberry Pi 4B via the USB interface. The Raspberry Pi 4B converts the polar coordinates to Cartesian coordinates; 3. Fuzzy decision output scenario level: The Raspberry Pi 4B extracts environmental features based on the collected LiDAR data. The specific steps are as follows: Polar coordinate to Cartesian coordinate transformation: Raspberry Pi 4B converts polar coordinates Convert to Cartesian coordinates : ; The sensor scan of the first Radial distance between points (unit: meters); The sensor scan of the first Angle of a point (unit: degrees, 0° points directly in front of the robot); : Coordinates in the transformed Cartesian coordinate system (unit: meters); The specific data after conversion is shown in Table 4 below: Table 4 Cartesian coordinates of obstacles ; The transformed Cartesian coordinates are filtered using the Exponentially Weighted Moving Average (EWMA) algorithm to reduce noise interference; the filtered distance value is calculated recursively. ; : No. The first frame The original distance values ​​of each point; : No. The filtered distance value of the frame; Forgetting factor, take ; The specific data after filtering are shown in Tables 5.1 and 5.2 below: Table 5.1 Filtered Cartesian coordinates ( ) ; Table 5.2 Filtered Cartesian coordinates ( ) ; Calculate the distance to the nearest obstacle : ; : effective points Euclidean distance to the center of the robot; This represents all the effective points after filtering; Table 6 shows the Euclidean distance from each effective point to the robot center. Table 6. Euclidean distances from each valid point to the robot center ; The nearest obstacle distance Substitute the nearest obstacle distance trapezoidal function : ; : Indicates the distance to the nearest obstacle Membership degree and range of "near" fuzzy sets ; Calculate obstacle density : Calculate the obstacle density within an effective radius R = 4 meters centered on the robot. Since the distances of all 10 data points are... meters, therefore, obstacle density Incorporating obstacle density Trapezoidal membership function : ; : Represents obstacle density The membership degree and range of a "high" fuzzy set. ; can be obtained ; Based on the nearest obstacle distance obstacle density Get the nearest obstacle distance Obstacle density Membership function , Only fuzzy inference rule 3 is activated. Assuming the membership degree in other scenarios is 0, the activation strength of fuzzy inference rule 3 is as follows: ; Nearest obstacle distance Belongs to fuzzy set Membership degree; Obstacle density Belongs to fuzzy set Membership degree; get The centroid method is used for deblurring to obtain the scene level value: ; The scene level value is 2, and the final output is the scene level of the current environment. Narrow; 4. Scene level mapping and delayed confirmation; Scene level mapping is performed. The Raspberry Pi 4B receives scene level 2 as narrow. Through the pre-established scene-parameter mapping table, the corresponding parameter group of the Dynamic Window Algorithm (DWA) is matched. To avoid frequent parameter switching caused by sensor noise, a hysteresis switching mechanism is established by employing continuous switching. Frames are determined to be in the same scene, so N=5; the Raspberry Pi 4B starts counting scene levels, and when 5 consecutive frames are all in the same scene level... When the time is right, a parameter switching signal is triggered; if the scene level changes during the counting process, the counting restarts until the condition of 5 consecutive frames is met. 5. Parameter distribution and initialization; Once the hysteresis switching mechanism is satisfied, the Raspberry Pi 4B sends the parameter set of the Dynamic Window Algorithm (DWA) to the DWA algorithm execution module.

[0026] The algorithm execution module initializes according to the following process: Stop the current path planning, clear historical trajectory data to avoid interference from old paths to the new path. Load the parameter set of the new Dynamic Window Algorithm (DWA), update the feasible velocity window boundaries of the velocity planner and the weight coefficients of the steering controller; based on the parameter set of the new DWA algorithm, generate a new local path starting from the current robot pose. The new local path will more carefully avoid obstacles, ensuring the robot can safely and smoothly pass through scenes with relatively concentrated obstacles.

[0027] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. For those skilled in the art, all features disclosed in the present invention, or all steps in all methods or processes disclosed, except for mutually exclusive features and / or steps, can be combined in any way without departing from the principles of the present invention. The present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A scene recognition method for adaptive navigation of robots, characterized in that, The scene recognition method described above is based on a data acquisition unit, a data storage unit, and a data processing unit; The data processing unit is located on the host computer, and sensors are installed on the robot. The data collected by the sensors is uploaded to the host computer. The host computer performs environmental feature extraction, which includes: Based on the obstacle distance data collected by the sensors, calculate the nearest obstacle distance MOD and the obstacle density OD within a predetermined effective radius around the robot; The host computer uses the nearest obstacle distance (MOD) and obstacle density (OD) as inputs to the fuzzy decision system. After fuzzifying these two inputs, it performs inference through the fuzzy inference rule base, defuzzifies the inference result, and outputs a quantitative value of the scene level. The scene level is one of four types: impassable, narrow, moderate, and wide. Furthermore, a delayed switching mechanism is introduced, which triggers parameter group switching only when the quantization values ​​of the scene level are the same for N consecutive frames, switching the parameter group of the Dynamic Window Algorithm (DWA) to the parameter group of the corresponding scene. The parameter set should include at least: maximum linear velocity, maximum angular velocity, maximum linear acceleration, and target orientation weight coefficient and obstacle avoidance distance weight coefficient in the Dynamic Window Algorithm (DWA) evaluation function; Among them, the maximum linear velocity, maximum angular velocity, and maximum linear acceleration constitute the kinematic constraint layer, while the target orientation weight coefficient and obstacle avoidance distance weight coefficient constitute the decision preference layer. The kinematic constraint layer and the decision preference layer are adjusted in coordination according to the scene level.

2. The scene recognition method for robot environment adaptive navigation according to claim 1, characterized in that, The sensors mentioned include lidar and ultrasonic sensors.

3. The scene recognition method for adaptive navigation of robots according to claim 1, characterized in that, The process of environmental feature extraction is as follows: S31. Collect information about surrounding obstacles using sensors mounted on the robot; S32. Perform coordinate transformation; The sensor collects the polar coordinates of surrounding obstacles. The data is simultaneously transmitted to the host computer's data processing unit; the polar coordinates are converted to Cartesian coordinates. ； The sensor scan of the first Radial distance between points (unit: meters); The sensor scan of the first Angle of a point (unit: degrees, 0° points directly in front of the robot); : Coordinates in the transformed Cartesian coordinate system (unit: meters); S33. Perform sliding window mean filtering; The Exponentially Weighted Moving Average (EWMA) algorithm is used to suppress instantaneous noise in single-frame data; the EWMA algorithm recursively calculates the filtered distance value. ； : No. The first frame The original distance values ​​of each point; : No. The filtered distance value of the frame; Forgetting factor, ranging from 0 to 1, is used to weigh the importance of current measurements against historical data; when When the value is close to 1, the Exponentially Weighted Moving Average (EWMA) algorithm focuses on the current measurement and reacts quickly to real-time changes in the data, but it introduces more noise; when... When the value is close to 0, the exponentially weighted moving average (EWMA) algorithm relies more on historical data and has a better smoothing effect on noise, but it is slower to respond to real-time changes in data. S34. Calculate the distance to the nearest obstacle. ; Calculate the Euclidean distance from the robot to the nearest obstacle to assess whether the robot is in a dangerous state requiring emergency obstacle avoidance; for all filtered valid points... In the algorithm, with the robot's position as the origin, the Euclidean distance from each point to the robot's position is calculated. Take the minimum value as the nearest obstacle distance. ; ； : effective points Euclidean distance to the center of the robot; This represents all valid points after filtering; S35. Calculate the obstacle density around the robot. ; Calculate the obstacle density within the effective radius R around the robot. Obstacle density is defined by an indicator function. : ； in, This is an indicator function, which is 1 when the condition is met and 0 otherwise; the closer the indicator function value is to 1, the denser the distribution of obstacles.

4. The scene recognition method for robot environment adaptive navigation according to claim 1, characterized in that, The fuzzy decision system includes the following: S41. Calculate the fuzzy membership function to obtain the precise nearest obstacle distance. Obstacle density Membership degree is transformed into a fuzzy concept; closest obstacle distance trapezoidal function : ； Nearest obstacle distance Actual input value (unit: meters); : Indicates the distance to the nearest obstacle Membership degree of a "very close" fuzzy set, range ; closest obstacle distance trapezoidal function : ； : Indicates the distance to the nearest obstacle Membership degree and range of "near" fuzzy sets ; closest obstacle distance trapezoidal function : ； : Indicates the distance to the nearest obstacle Membership degree and range of "medium" fuzzy sets ; closest obstacle distance Trapezoidal membership function : ； : Indicates the distance to the nearest obstacle Membership degree and range of the "far" fuzzy set ; obstacle density Trapezoidal membership function : ； Obstacle density Actual input value, unit: pieces / m²; : Represents obstacle density Membership degree of a "low" fuzzy set, range ; obstacle density Trapezoidal membership function : ； : represents the obstacle density Membership degree belonging to the "medium" fuzzy set, range ; obstacle density Trapezoidal membership function : ； : Represents obstacle density Membership degree of a "high" fuzzy set, range ; S42. Set fuzzy inference rules and calculate the activation strength of the fuzzy inference rules; To perform fuzzy inference, after fuzzifying the input variables, a fuzzy rule base is constructed, and then the minimum operator is used. The activation intensity of each fuzzy inference rule is calculated logically to determine the degree of influence of each fuzzy inference rule on the decision under the current input conditions. The fuzzy inference rule base is a multi-input, single-output system, where the input is the nearest obstacle distance. Obstacle density The fuzzy value is output as the scene level value; the fuzzy inference rule is: if the nearest obstacle distance is... It is very close, and If it is high density, then the scene level cannot be approved; Among them, the Activation strength of fuzzy inference rules : ； Nearest obstacle distance Belongs to fuzzy set Membership degree; Obstacle density Belongs to fuzzy set Membership degree; S43. Deblurring converts the blurred output into a clear scene level value; The fuzzy values ​​obtained through fuzzy reasoning are fuzzy values ​​at the scene level. These fuzzy values ​​at the scene level are fuzzy outputs. Defuzzification transforms these fuzzy values ​​at the scene level into explicit scene level values, enabling the robot to make specific decisions. The centroid method is used for deblurring. The membership degrees of different scene levels are weighted and summed with their corresponding quantized values, and then divided by the sum of membership degrees to obtain the scene level value. ； Scene Level Membership degree; : Quantitative value of scene level These correspond to scenarios that are impassable, narrow, moderate, and wide, respectively. S44. Establish a delayed switching mechanism; Improve decision stability by delaying confirmation; prevent frequent parameter switching by switching conditions; switching conditions: continuous. The frames were determined to be from the same scene; The threshold is the number of consecutive frames; when consecutive When the deblurring results of the frames are all at the same scene level, the parameter set of the Dynamic Window Algorithm (DWA) is switched to the parameter set of the corresponding scene to suppress high-frequency noise interference.

5. The scene recognition method for adaptive navigation of robots according to claim 1, characterized in that, The parameter group switching method of the Dynamic Window Algorithm (DWA) includes the following: S51. The host computer receives the quantized value of the scene level output by the fuzzy decision system. It uses a pre-defined scene-parameter mapping table for fast indexing; the scene-parameter mapping table is stored in key-value pairs, where the key is the scene level and the value is the parameter set of the corresponding Dynamic Window Algorithm (DWA); the mapping logic of the scene-parameter mapping table follows the matching principle between scene characteristics and control strategies: For low-level scenarios with a quantization value of 1 or 2, parameter sets with low speed and high obstacle avoidance weights are matched first to restrict robot movement and ensure safety; for high-level scenarios with a quantization value of 3 or 4, parameter sets with high speed and low steering sensitivity are matched to unleash robot movement capabilities and improve robot movement efficiency. S52. To suppress high-frequency parameter group switching caused by sensor noise or transient environmental changes, a continuous frame count confirmation mechanism is introduced: when the quantization value of the scene level is the same for N consecutive frames, the scene is considered stable and parameter group switching is triggered; if the quantization value of the scene level changes in any frame before N frames, the counter is reset and counting starts again until the condition of N consecutive frames is met; the continuous frame count confirmation mechanism ensures that parameter group switching only occurs when the scene is continuously stable through redundant judgment in the time dimension, avoiding control quantity oscillation caused by accidental noise; S53. The host computer sends the target parameter set to the Dynamic Window Algorithm (DWA) execution module through standardized communication interfaces, including CAN bus, serial port, or ROS parameter server, triggering the initialization process: S531. Stop the current path planning, interrupt the running local path planning thread, clear historical trajectory data and motion command queue to avoid control conflicts caused by the mixing of old and new parameter groups; S532. Load new parameter set: Write the parameter set, including maximum linear velocity, maximum angular velocity, obstacle avoidance weight, and target approach weight, into the core configuration module of the Dynamic Window Algorithm (DWA), and update the feasible speed window boundary of the speed planner and the weight coefficient of the steering controller. S533. Regenerate Local Path: Starting from the robot's current pose, and combining the motion constraints defined by the new parameter set, based on the real-time environment map, recalculate the set of feasible trajectories containing velocity and angular velocity, regenerate the local path, and select the optimal path to output to the actuator.

6. The scene recognition method for robot environment adaptive navigation according to claim 5, characterized in that, The parameters in the parameter group of the Dynamic Window Algorithm (DWA) are modified according to the robot's dynamic characteristics and task requirements. The maximum linear velocity in narrow scenes is set to 0.3 m / s and the obstacle avoidance weight is set to 0.

8. For heavy-duty robots, the maximum linear velocity threshold is reduced, and for high-speed moving scenes, the obstacle avoidance weight is adjusted to balance obstacle avoidance and efficiency.

Images