A tactile sensing device and a robot terrain sensing method
Through a robot terrain perception method based on array haptic sensors, the tactile perception components are used to collect and process data, and combined with machine learning algorithms to identify the soft and hard properties of the terrain, the vibration noise problem in terrain perception of rotating leg robots is solved, achieving high accuracy and low cost perceived effect.
Patent Information
- Application Number
- CN202110974557.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-24
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-08-24
AI Technical Summary
Due to vibration noise problems, existing terrain perception methods such as vibration and audio-based methods are not applicable, and the prior art has failed to achieve more than 95% recognition accuracy, and the cost and real-time indicators are not effective.
Using a robot terrain perception method based on array haptic sensors, pressure data is collected through haptic perception components, time sequence data preprocessing, peak window segmentation, feature extraction and machine learning classification are used to identify the soft and hard attribute categories of the terrain.
It realizes low-cost terrain perception, reduces the sensor cost of robot gait switching and secure navigation tasks, improves the real-time terrain perception, and achieves recognition accuracy of more than 94.0%.
Smart Images

Figure CN113673447B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of robot terrain perception, and in particular to a tactile perception device and a robot terrain perception method. Background Art
[0002] In recent years, rotating legged robots have shown better performance than legged robots and wheeled robots in terrain traversal tasks, especially on unstructured terrain. However, due to the rhythmic movement pattern of the rotating legs, the robot body usually vibrates violently and produces huge noise, which makes terrain perception methods based on vibration and audio unfeasible. For rotating legged robots, there is currently no reliable and low-cost terrain perception method designed. Therefore, how to combine the movement characteristics of rotating legged robots for terrain perception has become an important problem to be solved.
[0003] There are several terrain perception methods in the prior art:
[0004] 1. Chinese invention patent "Terrain recognition method and device based on the interactive sound between a mobile robot and the ground" (publication number: CN112288870A), which discloses a terrain recognition method based on the interactive sound between a mobile robot and the ground. It collects audio data of the interaction between the mobile robot and various terrains, and after pre-processing, it recognizes the terrain by constructing a terrain classification network module based on the attention mechanism, eliminating the influence of illumination changes on system detection.
[0005] 2. Chinese invention patent "A method for terrain classification and identification of vibration information based on multi-layer perceptron" (publication number: CN111027627A), which discloses a method for terrain classification and identification of vibration information based on multi-layer perceptron, collects the original vibration data of the vibration sensor in different terrain environments, converts it to frequency and uses multi-layer perceptron as a model for training, and finally can perform online terrain type classification and identification, and the overall classification accuracy is maintained at about 84%.
[0006] 3. Chinese invention patent "A terrain perception method for a legged robot based on virtual sensors" (publication number: CN111260057A), which discloses a terrain perception method for a legged robot based on virtual sensors. The angles of the leg joints, the angular velocities of the leg joints, and the motor current of the legged robot are used as data samples of the virtual sensors. The ground detection neural network model and the soil classification machine learning model are trained as the terrain perception system of the legged robot. The SVM model with a linear kernel function is used to achieve a classification accuracy of 92%.
[0007] The above-mentioned existing technologies are all terrain perception methods designed based on legged robots. In actual applications, since rotary legged robots rely on periodic rotation to move forward, the robot body will vibrate violently and produce huge noise, which not only makes the terrain perception methods based on vibration and audio no longer applicable to rotary legged robots; but also the above-mentioned methods have not achieved a high recognition accuracy of more than 95%, and a certain degree of error will still occur in actual applications; the above-mentioned methods have not considered the real-time and cost indicators of the methods, and the effect is poor in actual terrain perception. Summary of the invention
[0008] In order to solve the above technical problems, one of the objects of the present invention is to provide a tactile perception device for a robot, comprising a tactile perception component for generating and collecting data, a time series data preprocessing module for preprocessing data, a peak window segmentation module for segmenting data, a feature extraction module for extracting features, and a machine learning classification algorithm module for identifying terrain soft and hard attribute categories.
[0009] Preferably, the tactile sensing component includes a bottom support plate, an inner support frame, an outer support frame, a sensor signal acquisition module and a plurality of tactile sensors, the inner support frame and the outer support frame are arranged in parallel and at intervals, the two ends of the bottom support plate are respectively connected to the bottom ends of the inner support frame and the outer support frame, the plurality of tactile sensors are arranged at intervals on the bottom surface of the bottom support plate, the sensor signal acquisition module is arranged on the inner support frame or the outer support frame, and the sensor signal acquisition module is electrically connected to the plurality of tactile sensors.
[0010] The second object of the present invention is to provide a robot terrain perception method based on an array tactile sensor, which uses the above-mentioned tactile perception device for perception, and includes step 1: the tactile perception component collects pressure data and generates time series data; step 2: the time series data preprocessing module preprocesses the time series data; step 3: the peak window segmentation module segments the data points of the peak window; step 4: the feature extraction module outputs the tactile pressure features; step 5: the machine learning classification algorithm module identifies the soft and hard attribute categories of the terrain. The robot terrain perception method based on the array tactile sensor has the advantages of reducing the sensor cost of robot gait switching, safe navigation and other tasks, and improving the real-time performance of terrain perception.
[0011] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is as follows:
[0012] A robot terrain perception method based on an array tactile sensor comprises the following steps:
[0013] Step 1: collecting pressure data of multiple different positions of the robot when it contacts the ground through the tactile sensing component, and generating time series data of multiple channels;
[0014] Step 2: The time series data preprocessing module preprocesses the time series data;
[0015] Step 3: The peak window segmentation module segments the preprocessed time series data into data points of peak windows of multiple channels;
[0016] Step 4: the feature extraction module calculates the multi-dimensional tactile pressure features related to the hardness and softness of the terrain according to the data points in the peak window;
[0017] Step 5: The machine learning classification algorithm module identifies the softness and hardness attributes of different terrains according to the tactile pressure characteristics.
[0018] Preferably, step 1 comprises:
[0019] Step 1.1: The tactile sensing component generates pressure data at multiple different positions when the robot contacts the ground;
[0020] Step 1.2: The pressure data includes a pressure analog signal, and the tactile sensing component collects the pressure analog signal and converts the pressure analog signal into a quantifiable pressure digital signal;
[0021] Step 1.3: The tactile sensing component generates time series data based on the correspondence between the timestamp and the pressure digital signal.
[0022] Preferably, step 2 comprises:
[0023] Step 2.1: The time series data preprocessing module performs linear interpolation on the time series data;
[0024] Step 2.2: the time series data preprocessing module performs neighbor interpolation on the time series data at the head end and the tail end;
[0025] Step 2.3: The time series data preprocessing module uses median filtering to remove noise in the time series data that is seriously deviated from the steady-state value and has a negative amplitude;
[0026] Step 2.4: The time series data preprocessing module uses Savitzky-Golay filtering to smooth the time series data while maintaining the periodicity and trend of the time series data itself.
[0027] Preferably, step 3 comprises:
[0028] Step 3.1: The peak window segmentation module uses the peak detection function in SciPy to detect the peaks in the preprocessed time series data;
[0029] Step 3.2: The peak window segmentation module continuously outputs the position index of the peak;
[0030] Step 3.3: the peak window segmentation module traverses the position indexes of the peaks of multiple channels;
[0031] Step 3.4: The peak window segmentation module removes the position index of the first and last peaks of each channel;
[0032] Step 3.5: The peak window segmentation module checks whether the time difference between the position index of the current position in each channel and the position index of the same position in the adjacent channel is within 20%; if so, the position index of the current position in each channel is retained; otherwise, the position index out of range is removed and the position index at the next moment is used as a replacement value;
[0033] Step 3.6: The peak window segmentation module checks whether the number of position indexes in each channel is the same; if so, the position index is used as the center of the peak window, the length is set as the size of the peak window, and the data points of the segmented peak window are output; otherwise, the first position index of the first channel is used as the starting point, the minimum number of position indexes in multiple channels is used as the length, the position indexes of multiple channels are adjusted to the same size, and then the data points of the segmented peak window are output.
[0034] Preferably, the parameters of the peak detection function are set as follows: the lowest height of the peak is the median of the time series data, the distance between adjacent peaks is 300, the prominence coefficient of the peak is 0.1, and the width coefficient of the peak is 1.
[0035] Preferably, step 4 comprises:
[0036] Step 4.1: The feature extraction module calculates multiple tactile pressure indicators according to the data points of the peak window to obtain multi-dimensional tactile pressure features;
[0037] Step 4.2: The feature extraction module uses a chi-square test to measure the correlation between each tactile pressure feature and the hardness category of the terrain;
[0038] Step 4.3: The feature extraction module removes the tactile pressure features that are irrelevant to the hardness category of the terrain, and outputs multi-dimensional tactile pressure features that are relevant to the hardness category of the terrain.
[0039] Preferably, in step 5, the machine learning classification module uses an unsupervised Gaussian mixture clustering algorithm to calculate and analyze the tactile pressure features output from step 4, and ultimately identifies categories of different softness and hardness attributes of the terrain.
[0040] Preferably, in step 5, the machine learning classification module uses a supervised random forest algorithm to calculate and analyze the tactile pressure features output from step 4, and ultimately identifies categories of different softness and hardness attributes of the terrain.
[0041] Compared with the prior art, the present invention has achieved beneficial technical effects:
[0042] The present invention is aimed at the robot's outdoor natural environment and proposes a novel robot terrain perception method based on an array tactile sensor. A tactile perception device is used for perception. The pressure data of multiple different positions of the robot when it contacts the ground during movement is collected through the tactile perception component, and time series data is generated to be output to the software algorithm module of the host computer. During the algorithm execution process, the time series data is sequentially input into the time series data preprocessing module, the peak window segmentation module, and the feature extraction module. Finally, the machine learning classification module identifies the categories of different hardness and softness attributes of the terrain. The present invention can use low-cost tactile sensors to complete the robot's terrain perception task in the tactile modality, which is beneficial to reducing the sensor cost of robot gait switching, safe navigation and other tasks, and improving the real-time performance of terrain perception. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 is a flow chart of a robot terrain perception method according to an embodiment of the present invention;
[0044] Figure 2 It is a topographic map of different softness and hardness involved in the embodiment of the present invention;
[0045] Figure 3 is a waveform diagram of time series data obtained by sensing different terrains in an embodiment of the present invention;
[0046] Figure 4 is a schematic diagram of a peak window segmentation module according to an embodiment of the present invention segmenting multi-channel peak window data points;
[0047] Figure 5 is a list diagram of tactile pressure indicators according to an embodiment of the present invention;
[0048] Figure 6 is a confusion matrix diagram of classification accuracy compared with other methods in the embodiment of the present invention;
[0049] Figure 7 is a performance comparison diagram of the embodiment of the present invention and other methods;
[0050] Figure 8 Schematic diagram of a tactile sensing component according to an embodiment of the present invention.
[0051] The technical features indicated by the reference numerals are as follows:
[0052] 1. Tactile sensing component; 11. Bottom support plate; 12. Inner support frame; 13. Outer support frame; 14. Sensor signal acquisition module; 15. Tactile sensor. DETAILED DESCRIPTION
[0053] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments, but the scope of protection claimed in the present invention is not limited to the following specific embodiments.
[0054] refer to Figure 1-8 The present embodiment discloses a tactile perception device of a robot, including a tactile perception component 1 for generating and collecting data, a time series data preprocessing module for preprocessing data, a peak window segmentation module for segmenting data, a feature extraction module for extracting features, and a machine learning classification algorithm module for identifying soft and hard attribute categories of terrain. The tactile perception component 1, the time series data preprocessing module, the peak window segmentation module, the feature extraction module, and the machine learning classification algorithm module are electrically connected or wirelessly connected, and the wireless connection methods include WiFi connection, Bluetooth connection, and the like.
[0055] Furthermore, the tactile sensing component 1 includes a bottom support plate 11, an inner support frame 12, an outer support frame 13, a sensor signal acquisition module 14 and a plurality of tactile sensors 15. The inner support frame 12 and the outer support frame 13 are arranged in parallel and at intervals. The two ends of the bottom support plate 11 are respectively connected to the bottom ends of the inner support frame 12 and the outer support frame 13. The plurality of tactile sensors 15 are arranged at intervals on the bottom surface of the bottom support plate 11. The sensor signal acquisition module 14 is arranged on the inner support frame 12 or the outer support frame 13, and the sensor signal acquisition module 14 is electrically connected to the plurality of tactile sensors 15.
[0056] Furthermore, in this embodiment, the bottom support plate 11 is an arc-shaped plate, and the two ends of the bottom support plate 11 along the axial direction are fixedly connected to the bottom ends of the inner support frame 12 and the outer support frame 13 respectively, and a plurality of tactile sensors 15 are arranged at equal angles along the circumference of the bottom support plate 11. The bottom support plate 11, the inner support frame 12 and the outer support frame 13 constitute the variable stiffness leg of the robot's arc-foot-like leg.
[0057] The sensor signal acquisition module 14 includes a microcontroller, a capacitance-to-digital converter and a Bluetooth communication module. The tactile sensor 15 , the microcontroller, the capacitance-to-digital converter and the Bluetooth communication module are electrically connected to each other.
[0058] In this embodiment, the touch sensor 15 includes a flexible capacitive pressure sensor.
[0059] refer to Figure 1-8 This embodiment also discloses a robot terrain perception method based on an array tactile sensor, which uses the above-mentioned tactile perception device for perception, and includes the following steps:
[0060] Step 1: The tactile sensing component 1 installed on the robot collects pressure data at multiple different positions when the robot contacts the ground, and generates time series data of multiple channels;
[0061] Step 2: The time series data preprocessing module preprocesses the time series data;
[0062] Step 3: The peak window segmentation module segments the preprocessed time series data into data points of peak windows of multiple channels;
[0063] Step 4: The feature extraction module calculates the multi-dimensional tactile pressure features related to the hardness and softness of the terrain based on the data points in the peak window;
[0064] Step 5: The machine learning classification algorithm module identifies the softness and hardness attributes of different terrains based on the tactile pressure characteristics.
[0065] The present invention is aimed at the robot's outdoor natural environment and proposes a novel robot terrain perception method based on an array tactile sensor. The method performs perception through a tactile perception device, and uses a tactile perception component 1 to obtain tactile pressure information at multiple different positions when the robot contacts the ground during movement, and generates time series data to output to a software algorithm module of a host computer. During the execution of the algorithm, the time series data is sequentially input into a time series data preprocessing module, a peak window segmentation module, and a feature extraction module. Finally, a machine learning classification module identifies the categories of different hardness and softness attributes of the terrain. The present invention can utilize a low-cost tactile sensor 15 to complete the robot's terrain perception task in tactile mode, which is beneficial to reducing the sensor cost of robot gait switching, safe navigation and other tasks, and improving the real-time performance of terrain perception.
[0066] Step 1 includes:
[0067] Step 1.1: The tactile sensing component 1 generates pressure data of multiple different positions when the robot contacts the ground. Specifically, multiple tactile sensors 15 generate pressure data of multiple different positions when the robot contacts the ground.
[0068] Step 1.2: The pressure data includes a pressure analog signal. The tactile sensing component 1 collects the pressure analog signal and converts the pressure analog signal into a quantifiable pressure digital signal. Specifically, the sensor signal acquisition module 14 collects the pressure analog signal and converts it into a pressure digital signal. The pressure analog signal is collected by a microcontroller and transmitted to a capacitance digital converter. The microcontroller controls the capacitance digital converter to convert the pressure analog signal into a quantifiable pressure digital signal.
[0069] Step 1.2: The tactile perception component 1 generates time series data based on the correspondence between the timestamp and the pressure digital signal, and then uploads the time series data to the software algorithm module of the host computer. Specifically, the sensor signal acquisition module 14 generates time series data and uploads it to the software algorithm module of the host computer. The microcontroller controls the capacitance digital converter to generate time series data based on the correspondence between the timestamp and the pressure digital signal and transmits it to the Bluetooth communication module. The microcontroller controls the Bluetooth communication module to upload the time series data to the software algorithm module of the host computer wirelessly.
[0070] Furthermore, in this embodiment, taking the example of five tactile sensors 15 provided on the bottom surface of the bottom support plate 11, the five tactile sensors 15 can measure the pressure data of five different positions when the robot is in contact with the ground during movement, and the sensor signal acquisition module 14 collects the pressure data of five different positions and generates five channels of time series data.
[0071] Since there are abnormal values and missing values in the original time series data collected by the sensor signal acquisition module 14, the present invention constructs a time series data preprocessing module to process the time series data and generate time series data that is easier to process. Figure 1 The second part of the process of robotic terrain perception methods.
[0072] Step 2 includes:
[0073] Step 2.1: The time series data preprocessing module performs linear interpolation on the time series data;
[0074] Step 2.2: The time series data preprocessing module performs nearest neighbor interpolation on the time series data at the beginning and end;
[0075] The above steps 2.1 and 2.2 are used to fill the missing values in the time series data;
[0076] Step 2.3: The time series data preprocessing module uses median filtering to remove the noise in the time series data that deviates seriously from the steady-state value and has a negative amplitude;
[0077] Step 2.4: The time series data preprocessing module uses Savitzky-Golay filtering to smooth the time series data while maintaining the periodicity and trend of the time series data itself.
[0078] The peak window segmentation module is able to extract and align the five channels of pressure data generated each time the robot contacts the ground.
[0079] Step 3 includes:
[0080] Step 3.1: The peak window segmentation module uses the peak detection function in SciPy to detect the peaks in the preprocessed time series data;
[0081] Step 3.2: The peak window segmentation module continuously outputs the position index of the peak;
[0082] Step 3.3: The peak window segmentation module traverses the position indexes of the peaks of multiple channels, that is, traverses the position indexes of the peaks of five channels;
[0083] Step 3.4: The peak window segmentation module removes the position index of the first and last peak of each channel;
[0084] Step 3.5: The peak window segmentation module checks whether the time difference between the position index of the current position in each channel and the position index of the same position in the adjacent channel is within 20%; if so, the position index of the current position in each channel is retained; otherwise, the position index out of range is removed and the position index at the next moment is used as a replacement value;
[0085] Step 3.6: The peak window segmentation module checks whether the number of position indexes in each channel is the same; if so, the position index is used as the center of the peak window, and the set length is used as the size of the peak window, and the data points of the segmented peak window are output. In this embodiment, the set length of the peak window is 300; otherwise, the first position index of the first channel is used as the starting point, and the minimum number of position indexes in multiple channels is used as the length. The position indexes of multiple channels are adjusted to the same size and then the data points of the segmented peak window are output. The visualization result of the data output by the peak window segmentation module is as follows: Figure 4 shown.
[0086] The peak window segmentation module not only ensures that the robot can quickly detect the latest terrain it passes through, but the peak extraction step also avoids the truncation of a complete rotation cycle of the robot leg when segmenting the data with a rectangular window of fixed length. The peak extraction step also adapts to the impact of the robot generating data of different frequencies in different gaits.
[0087] Furthermore, the parameters of the peak detection function used in step 3.1 are set as follows: the minimum height of the peak is the median of the time series data, the distance between adjacent peaks is 300, the prominence coefficient of the peak is 0.1, and the width coefficient of the peak is 1.
[0088] The feature extraction module can calculate high-level tactile pressure features from the pressure data.
[0089] Step 4 includes:
[0090] Step 4.1: The feature extraction module calculates multiple tactile pressure indicators based on the data points of the peak window to obtain multi-dimensional tactile pressure features;
[0091] Step 4.2: The feature extraction module uses the chi-square test to measure the correlation between each tactile pressure feature and the hardness category of the terrain;
[0092] Step 4.3: The feature extraction module removes the tactile pressure features that are irrelevant to the hardness category of the terrain, and outputs multi-dimensional tactile pressure features that are relevant to the hardness category of the terrain.
[0093] Specifically, the list of tactile pressure indicators calculated by the present invention is as follows: Figure 5 As shown in the figure, the tactile pressure indicators include absolute energy, autocorrelation coefficient, complexity invariant distance, kurtosis, skewness, sample entropy, maximum value, minimum value, mean value, median value, variance, and standard deviation. The absolute energy uses the sum of squares of the peak data to characterize the energy exchanged when the robot leg collides with the terrain; the autocorrelation coefficient characterizes the autocorrelation of the peak data in the time dimension; the complexity invariant distance estimates the complexity within the peak window; the kurtosis characterizes the Fisher-Pearson standard moment coefficient G1 of the peak data; the skewness characterizes the Fisher-Pearson standard moment coefficient G2 of the peak data; the son standard moment coefficient G2; sample entropy characterizes the complexity of the time series by calculating the probability of generating a new pattern in a peak window; maximum value, minimum value, mean value, median value, variance, and standard deviation are all conventional statistical indicators of the peak window; that is, after calculating the twelve tactile pressure indicators for the data points of the data windows of the five channels in step 4.1, a sixty-dimensional tactile pressure feature is obtained, and then after the chi-square test correlation in step 4.2, step 4.3 removes irrelevant tactile pressure features, which plays a role in dimensionality reduction, and finally outputs a twenty-six-dimensional tactile pressure feature vector, such as Figure 1 As shown in the fourth part of .
[0094] refer to Figure 8The applicant uses the AmphiHex-Ⅱ robot platform to construct data samples for machine learning classification modules. The robot platform is mainly composed of five parts: a waterproof shell system, a control circuit system, a power module, a drive module and a tactile sensing device. Among them, the waterproof shell system is used for waterproof sealing to protect the operation of internal components; the control circuit system is responsible for the operation control of the entire robot; the power module is used to power the drive module and the control circuit system; the drive module is responsible for outputting the power of the entire robot; the bottom support plate 11, the inner support frame 12 and the outer support frame 13 of the tactile sensing device constitute the variable stiffness legs of the robot platform. The inner support frame 12 and the outer support frame 13 are sleeved on the outside of the drive shaft of the drive module and fixed by locating pins. When the robot walks on the terrain, it mainly relies on the arc-shaped foot composed of the inner support frame 12, the outer support frame 13 and the bottom support plate 11 to achieve propulsion movement, which improves the overall system Reliability and stability enable the robot to adapt to various complex working conditions for effective support and propulsion movements. The tactile sensor 15 and the sensor signal acquisition module 14 are used to generate and collect pressure data. Six groups of tactile sensing components 1 are provided on the robot platform, and the six groups of tactile sensing components 1 are respectively arranged on the opposite sides of the robot platform, that is, the robot platform is provided with six variable stiffness legs, and the six variable stiffness legs of the robot platform are divided into two groups. The robot platform moves forward on the terrain with a three-legged gait. Specifically, the three-legged gait is that the robot always supports its main body with three non-adjacent legs, so that the center of gravity of the body falls within the triangle formed by the contact points of the three legs and the ground; when one group is in the supporting phase, the other group is in the swinging phase, and the movements of the three legs in the same group are exactly the same. The three-legged gait has the advantages of fast movement speed, high efficiency and wide application range, so the present invention selects the movement gait of the robot with a three-legged gait to construct the data sample.
[0095] In order to cope with different scenarios, the present invention uses the tactile array sensing device of the robot platform to collect data on a variety of terrains with different degrees of hardness and softness. Figure 2 As shown in the figure, they are named "very hard", "hard", "soft", and "very soft" respectively. When constructing the data set of the present invention, the influence of environmental factors such as illumination changes and weather on system detection is excluded, and finally 169 training samples and 109 test samples are included. Figure 3 As shown, the time series data collected by the tactile array sensing device for terrains of different hardness and softness present distinctive features visible to the naked eye.
[0096] In one embodiment, in step 5, the machine learning classification module uses an unsupervised Gaussian mixture clustering algorithm and a supervised random forest algorithm to calculate and analyze the tactile pressure features output in step 4, and finally identifies the categories of different softness and hardness attributes of the terrain. The Gaussian mixture clustering algorithm is a prior art and will not be described in detail here.
[0097] In one embodiment, in step 5, the machine learning classification module uses a supervised random forest algorithm to calculate and analyze the tactile pressure features output in step 4, and finally identifies the categories of different softness and hardness attributes of the terrain. The random forest algorithm is a prior art and will not be described in detail here.
[0098] Specifically, the Gaussian mixture model is used for unsupervised classification in the machine learning classification module, which shows that even without prior information about the map, the designed feature extraction module can map the data samples into separable clusters, proving that the extracted tactile pressure features have a good effect on distinguishing terrains with different attributes. The random forest algorithm has a faster training and testing speed, and can usually achieve higher accuracy while not being prone to overfitting. Therefore, the random forest algorithm is further adopted for supervised recognition of the hardness and softness categories of the terrain to achieve higher recognition accuracy.
[0099] The present invention conducts experiments in an actual outdoor environment. The robot platform generates about 300 data points in each step during the movement. In order to contrast with the segmentation method of the peak window segmentation module of the present invention, a traditional rectangular window segmentation method with lengths of 300, 1000 and 2000 is introduced. Figure 6 The confusion matrix diagram of the classification accuracy of the present invention compared with other methods is shown in FIG. Figure 6 As shown in (a), the recognition accuracy of the rectangular window segmentation method with a length of 300 is 76.5%; Figure 6 As shown in (c), the recognition accuracy of the rectangular window segmentation method with a length of 2000 is 94.0%. Although the recognition accuracy is improved with the expansion of the rectangular window, the important real-time performance in the robot terrain perception task is sacrificed at the same time. For example, for the rectangular window segmentation method with a length of 2000, eight to ten steps are required to correctly predict the attributes of the latest passed terrain, such as Figure 6 As shown in (d) and (e), the segmentation method of the peak window segmentation module of the present invention combined with the unsupervised Gaussian mixture algorithm and the supervised random forest algorithm achieved recognition accuracies of 94.0% and 95.5% respectively. Figure 7 The performance comparison shows that the method of the present invention can perform faster reasoning and more accurate recognition of terrain compared to other methods, although it uses fewer data points.
[0100] According to the disclosure and teaching of the above description, those skilled in the art to which the present invention belongs may also make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and some modifications and changes to the invention should also fall within the scope of protection of the claims of the present invention. In addition, although some specific terms are used in this specification, these terms are only for the convenience of description and do not constitute any limitation to the invention.
Claims
1. A tactile sensing device for a robot, characterized in that: It includes a tactile sensing component (1) for generating and collecting data, a time series data preprocessing module for preprocessing data, a peak window segmentation module for segmenting data, a feature extraction module for extracting features, and a machine learning classification algorithm module for identifying soft and hard attribute categories of terrain; The tactile sensing component (1) collects pressure data at multiple different positions when the robot is in contact with the ground, and generates time series data of multiple channels; The time series data preprocessing module preprocesses the time series data; the peak window segmentation module segments the preprocessed time series data into data points of peak windows of multiple channels; the feature extraction module outputs multi-dimensional tactile pressure features related to the hardness and softness of the terrain according to the data points of the peak windows; The machine learning classification algorithm module identifies the softness and hardness attributes of different terrains according to the tactile pressure characteristics; The method for the peak window segmentation module to segment the preprocessed time series data into data points of the peak windows of multiple channels comprises the following steps: the peak window segmentation module uses the peak detection function in SciPy to detect the peaks in the preprocessed time series data; the peak window segmentation module continuously outputs the position index of the peaks; The peak window segmentation module traverses the position indexes of the peaks of multiple channels; The peak window segmentation module removes the position index of the first and last peaks of each channel; The peak window segmentation module checks whether the time difference between the position index of the current position in each channel and the position index of the same position in the adjacent channel is within 20%; if so, the position index of the current position in each channel is retained; Otherwise, remove the position index that is out of range and use the position index at the next moment as the replacement value; The peak window segmentation module checks whether the number of position indexes in each channel is the same; if so, the position index is used as the center of the peak window, the length is set as the size of the peak window, and the data points of the segmented peak window are output; Otherwise, the first position index of the first channel is used as the starting point, the minimum number of position indexes in multiple channels is used as the length, the position indexes of multiple channels are adjusted to the same size, and then the data points of the segmented peak window are output.
2. The tactile sensing device according to claim 1, characterized in that: The tactile sensing component (1) comprises a bottom support plate (11), an inner support frame (12), an outer support frame (13), a sensor signal acquisition module (14) and a plurality of tactile sensors (15); the inner support frame (12) and the outer support frame (13) are arranged in parallel and at intervals; the two ends of the bottom support plate (11) are respectively connected to the bottom ends of the inner support frame (12) and the outer support frame (13); the plurality of tactile sensors (15) are arranged at intervals on the bottom surface of the bottom support plate (11); the sensor signal acquisition module (14) is arranged on the inner support frame (12) or the outer support frame (13); and the sensor signal acquisition module (14) is electrically connected to the plurality of tactile sensors (15).
3. A robot terrain perception method based on an array tactile sensor, characterized in that: The robot terrain perception method adopts the tactile perception device according to claim 1 or 2 for perception, and comprises the following steps: Step 1: collecting pressure data of multiple different positions of the robot when it contacts the ground through the tactile sensing component (1), and generating time series data of multiple channels; Step 2: The time series data preprocessing module preprocesses the time series data; Step 3: The peak window segmentation module segments the preprocessed time series data into data points of peak windows of multiple channels; Step 4: the feature extraction module outputs a multi-dimensional tactile pressure feature related to the hardness and softness of the terrain according to the data points in the peak window; Step 5: The machine learning classification algorithm module identifies the softness and hardness attributes of different terrains according to the tactile pressure characteristics.
4. The robot terrain perception method according to claim 3, characterized in that: Step 1 includes: Step 1.1: the tactile sensing component (1) generates pressure data at multiple different positions when the robot contacts the ground; Step 1.2: The pressure data includes a pressure analog signal, and the tactile sensing component (1) collects the pressure analog signal and converts the pressure analog signal into a quantifiable pressure digital signal; Step 1.3: The tactile sensing component (1) generates time series data based on the correspondence between the timestamp and the pressure digital signal.
5. The robot terrain perception method according to claim 3, characterized in that: Step 2 includes: Step 2.1: The time series data preprocessing module performs linear interpolation on the time series data; Step 2.2: the time series data preprocessing module performs neighbor interpolation on the time series data at the head end and the tail end; Step 2.3: The time series data preprocessing module uses median filtering to remove noise in the time series data that is seriously deviated from the steady-state value and has a negative amplitude; Step 2.4: The time series data preprocessing module uses Savitzky-Golay filtering to smooth the time series data while maintaining the periodicity and trend of the time series data itself.
6. The robot terrain perception method according to claim 3, characterized in that: Step 3 includes: Step 3.1: The peak window segmentation module uses the peak detection function in SciPy to detect the peaks in the preprocessed time series data; Step 3.2: The peak window segmentation module continuously outputs the position index of the peak; Step 3.3: the peak window segmentation module traverses the position indexes of the peaks of multiple channels; Step 3.4: The peak window segmentation module removes the position index of the first and last peaks of each channel; Step 3.5: The peak window segmentation module checks whether the time difference between the position index of the current position in each channel and the position index of the same position in the adjacent channel is within 20%; if so, the position index of the current position in each channel is retained; otherwise, the position index out of range is removed and the position index at the next moment is used as a replacement value; Step 3.6: The peak window segmentation module checks whether the number of position indexes in each channel is the same; if so, the position index is used as the center of the peak window, the length is set as the size of the peak window, and the data points of the segmented peak window are output; otherwise, the first position index of the first channel is used as the starting point, the minimum number of position indexes in multiple channels is used as the length, the position indexes of multiple channels are adjusted to the same size, and then the data points of the segmented peak window are output.
7. The robot terrain perception method according to claim 6, characterized in that: The parameters of the peak detection function are set as follows: the minimum height of the peak is the median of the time series data, the distance between adjacent peaks is 300, the peak prominence coefficient is 0.1, and the peak width coefficient is 1.
8. The robot terrain perception method according to claim 3, characterized in that: Step 4 includes: Step 4.1: The feature extraction module calculates multiple tactile pressure indicators according to the data points of the peak window to obtain multi-dimensional tactile pressure features; Step 4.2: The feature extraction module uses a chi-square test to measure the correlation between each tactile pressure feature and the hardness category of the terrain; Step 4.3: The feature extraction module removes the tactile pressure features that are irrelevant to the hardness category of the terrain, and outputs multi-dimensional tactile pressure features that are relevant to the hardness category of the terrain.
9. The robot terrain perception method according to claim 3, characterized in that: In step 5, the machine learning classification algorithm module uses an unsupervised Gaussian mixture clustering algorithm to calculate and analyze the tactile pressure features output from step 4, and finally identifies the categories of different softness and hardness attributes of the terrain.
10. The robot terrain perception method according to claim 3, characterized in that: In step 5, the machine learning classification algorithm module uses a supervised random forest algorithm to calculate and analyze the tactile pressure features output from step 4, and finally identifies the categories of different softness and hardness attributes of the terrain.
Citation Information
Patent Citations
Vibration information terrain classification and identification method based on multilayer perceptron
CN111027627A
Foot-type robot terrain perception method based on virtual sensor
CN111260057A
Terrain recognition method and device based on sound interaction between mobile robot and ground
CN112288870A
