Screening method and system for pressure sensor array for human sitting posture recognition

Abstract

The application discloses a screening method and system of a pressure sensor array for human sitting posture recognition, which is used for screening P pressure sensors from a pressure sensor array to construct a sitting posture recognition hardware system; a plurality of samples of each sitting posture under a plurality of preset sitting postures are acquired, each sample is a two-dimensional matrix composed of pressure data collected by all pressure sensors at the same time; an average pressure heat map of each sitting posture is generated, pixel point position indexes meeting high pressure characteristic and symmetry constraint are extracted, and a key position index set of each sitting posture is constructed; the key position index sets of all sitting postures are subjected to category-specific clustering, global fusion clustering and discretization, and a global optimal position index set containing P position indexes is obtained; and the advantage is that the pressure sensor accurately representing pressure distribution characteristics can be screened, and the cost is reduced, the data amount is reduced, and the data processing unit complexity is reduced without reducing the sitting posture recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of human posture recognition technology, and in particular to a method and system for screening pressure sensor arrays for human posture recognition. Background Technology

[0002] With the development of health monitoring and smart home technology, smart chairs are gradually becoming important devices for improving the sitting health of people who sit for long periods. Smart chairs typically integrate a human posture recognition system, which identifies different sitting postures of the user and then adjusts the chair's position to match the current posture or provides warnings for poor posture based on the recognition results.

[0003] Existing human posture recognition systems for smart chairs generally include a pressure sensor array, a signal acquisition circuit, and a data processing unit. The pressure sensor array is typically located inside the seat cushion or backrest of the smart chair to detect the pressure signals generated when a person contacts the chair, converting these signals into electrical signals that are transmitted to the data processing unit. The data processing unit typically uses machine learning algorithms to analyze the pressure data and recognize the posture. Because different sitting postures create different pressure distribution characteristics, the data processing unit uses machine learning algorithms to analyze the electrical signals to determine these pressure distribution characteristics, thereby achieving posture recognition.

[0004] To obtain more complete pressure distribution information, existing human posture recognition systems for smart chairs typically use pressure sensor arrays composed of a large number of distributed pressure sensors. For example, Chinese invention patent application CN202110192234.0 discloses a pressure cushion-based recognition system. This system collects pressure signals from 32×32 pressure sensors through two flexible pressure pads on the seat surface and backrest; that is, its pressure sensor array consists of 32×32 pressure sensors, conforming to a 32×32 specification. By collecting pressure data from these 32×32 pressure sensors, the system uses a convolutional neural network to automatically extract features to identify the type of sitting posture. It then combines this data with a state model based on features such as pressure variation coefficient difference and center of gravity distance difference to distinguish between stable and unstable sitting postures, thereby correcting the dynamic sitting posture category and providing feedback.

[0005] However, the use of pressure sensor arrays leads to a massive amount of data and computation, significantly increasing the complexity of subsequent data processing units, raising their computational requirements, and reducing the efficiency of posture recognition. To address this issue, Xu et al., in their 2021 paper "A portable sitting posture monitoring system based on a pressure sensor array and machine learning[J]. Sensors and Actuators A:Physical, 2021, 331:112900.", proposed a selection scheme for a human sitting posture recognition system using a 44×52 pressure sensor array. This selection scheme uses an alternating row and column approach to filter through different sizes such as 22×26, 11×13, 6×7, and 3×4, ultimately selecting 11×13 as the optimal pressure sensor array size for sitting posture recognition. This screening method can predetermine the optimal pressure sensor array size. When manufacturing a smart seat, the sensors at the corresponding points of the optimal pressure sensor array size can be used, thereby reducing costs, reducing the amount of pressure data entering subsequent processing, reducing the complexity of subsequent data processing units, and improving posture recognition efficiency.

[0006] However, this screening method simply increases the density of interlaced data acquisition, and the selected pressure sensors always adopt a rectangular layout. It retains some edge pressure sensors that have little effect on posture recognition, and discards some pressure sensors in key stress areas. As a result, the pressure data collected by the pressure sensor array for posture recognition cannot accurately represent the pressure distribution characteristics, which has an adverse effect on the accuracy of posture recognition. Summary of the Invention

[0007] This invention aims to provide a method and system for screening pressure sensor arrays for human posture recognition. This screening method and system can select pressure sensors that accurately characterize pressure distribution features, thereby reducing costs, decreasing the amount of pressure data entering subsequent processing, and lowering the complexity of subsequent data processing units without compromising posture recognition accuracy. This solves the problems of large pressure data volume, cumbersome data processing, and high computing power requirements of deployment equipment in existing technologies.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for screening a pressure sensor array for human sitting posture recognition, the pressure sensor array comprising, according to... M OK N column distribution M×NA pressure sensor, and the screening method is used to screen out P pressure sensors from the pressure sensor array to construct a human sitting posture recognition hardware system, where P is a preset positive integer and P < M×N, representing the optimal number of pressure sensors finally screened, and its specific value can be determined according to hardware costs, data processing volume, and sitting posture recognition accuracy requirements; obtaining multiple samples of each sitting posture under multiple preset sitting postures, and each sample is a two-dimensional matrix composed of pressure data collected by all pressure sensors at the same moment; generating an average pressure heat map for each sitting posture based on all samples of each sitting posture to eliminate differences between sample individuals, so as to accurately characterize the pressure characteristics of the pressure sensor array in spatial distribution; extracting the position indexes of pixel points that meet the preset high-pressure characteristics and symmetry constraints from the average pressure heat map of each sitting posture to form a key position index set for each sitting posture; based on the key position index sets of all sitting postures, performing class-specific clustering, global fusion clustering, and discretization to obtain a globally optimal position index set; the globally optimal position index set contains P position indexes, and the P position indexes correspond to P pressure sensors among the M×N pressure sensors, and are used to arrange the P pressure sensors according to the P position indexes to construct the human sitting posture recognition hardware system.

[0009] Further, the preset sitting postures are seven kinds, namely: leaning backward, leaning forward, leaning left, leaning right, left leg crossed over right leg, right leg crossed over left leg, and sitting upright; among them, leaning backward means that the upper body leans backward, leaning forward means that the upper body leans forward, leaning left means that the center of gravity of the upper body shifts to the left, leaning right means that the center of gravity of the upper body shifts to the right, left leg crossed over right leg means that the left leg is placed on top of the right leg, right leg crossed over left leg means that the right leg is placed on top of the left leg, and sitting upright means that the upper body remains basically symmetric and upright.

[0010] Further, the specific method for obtaining multiple samples of each sitting posture under multiple preset sitting postures is: selecting testers with a number not less than 100 and having diversity in body type, gender, and age distribution to ensure the representativeness and generalization ability of the samples; each tester sits on the intelligent seat equipped with the pressure sensor array in each sitting posture, the pressure sensor array is a flexible pressure sensor array, arranged on the seat cushion of the intelligent seat, and maintains this sitting posture within the preset sampling time, controlling the pressure sensor array to sample according to the preset sampling period, and obtaining the pressure data collected by all pressure sensors in each sampling period according to M rows N columns to form a two-dimensional matrix.

[0011] Further, set the position index of the pressure sensor in the m-th row and n-th column as (m, n), where m = 1, 2,..., M and n = 1, 2,..., N, and the pixel size of the average pressure heat map is M×NThe pixel in the m-th row and n-th column represents the pressure sensor in the m-th row and n-th column, with a position index of (m, n). The specific method for generating the average pressure heatmap for each sitting posture based on all samples is as follows: the minimum pressure data among all samples of each sitting posture is taken as the global minimum pressure value for that posture, and the maximum pressure data is taken as the global maximum pressure value for that posture. Based on the global minimum pressure value and global maximum pressure value for each sitting posture, each sample of each sitting posture is linearly scaled and mapped to [0, ..., ... The values ​​in the [1] interval are then mapped to the [0,255] interval and converted to 8-bit grayscale values ​​to obtain the normalized sample corresponding to each sample of each sitting posture. The 8-bit grayscale value in the m-th row and n-th column of the normalized sample corresponding to each sample of each sitting posture is used as the grayscale value of the pixel in the m-th row and n-th column of the pressure heat map of that sample to obtain the pressure heat map of each sample of each sitting posture. The average value of the grayscale value of the pixel in the m-th row and n-th column of the pressure heat map of all samples of each sitting posture is used as the grayscale value of the pixel in the m-th row and n-th column of the average pressure heat map of that sitting posture to obtain the average pressure heat map of each sitting posture.

[0012] Furthermore, the specific method for extracting the position indices of pixels that satisfy the preset high-pressure characteristics and symmetry constraints from the average pressure heatmap of each sitting posture to form the key position index set for each sitting posture is as follows: For the average pressure heatmap of each sitting posture, all pixels are sorted in descending order of their grayscale values. If multiple pixels have the same grayscale value, these multiple pixels are randomly arranged, and the pixels with the highest ranking (preferably 60%, or rounded up if the highest 60% is a decimal) are taken as high-pressure pixels, thus forming the high-pressure pixels for each sitting posture. Pixel set; For each sitting posture's high-pressure pixel set, determine whether the pixels that are mirror-symmetrical about the longitudinal central axis of the average pressure heatmap of that sitting posture are located in the high-pressure pixel set for that sitting posture. If not, the pixel is a pixel in the high-pressure pixel set for that sitting posture that meets the left-right symmetry constraint. The position indices of the pixels in the high-pressure pixel set for each sitting posture and all pixels that meet the left-right symmetry constraint together constitute the key position index set for each sitting posture to ensure the physical symmetry of the sensor array layout.

[0013] Furthermore, based on the key position index set for all sitting postures, category-specific clustering, global fusion clustering, and discretization are performed to obtain the globally optimal position index set. The specific method is as follows: the union of the key position index sets for all sitting postures is used as the global key position index set; the global key position index set is expanded to obtain an extended key position index set; K-means clustering is performed on the key position index set for each sitting posture to achieve category-specific clustering, resulting in a global candidate position index set; K-means clustering is then performed on the global candidate position index set to achieve global fusion clustering, resulting in a globally preliminary screening position index set; a conflict resolution mechanism and a point compensation mechanism are introduced to discretize the globally preliminary screening position index set to ensure that the final position indexes are within the effective position index range of the pressure sensor, thus obtaining the globally optimal position index set. Based on obtaining the key position index sets for all sitting postures, the global key position index set, and the extended key position index set, a two-stage clustering process is performed. Euclidean distance is used as the similarity metric during the clustering process to achieve optimized selection and spatially uniform distribution of the pressure sensor position indexes.

[0014] Furthermore, the global key location index set is expanded by position indexing. The specific method for obtaining the expanded key location index set is as follows: A positive integer k is set as the neighborhood expansion number; ... M×N The location index of each pressure sensor corresponds to... M OK N The columns are distributed to form a position index matrix. For each position index in the global key position index set, in the position index matrix, with the position index as the center, k adjacent position indices are obtained in the eight directions of up, down, left, right and four diagonal directions as extended position indices. If the number of adjacent position indices in a certain direction is less than k, the actual number of adjacent position indices is obtained. All extended position indices are added to the global key position index set and deduplication is performed to obtain the extended key position index set.

[0015] Furthermore, K-means clustering is performed on the key position index set for each sitting posture to obtain the global candidate position index set. The specific method is as follows: the key position index set for each sitting posture is used as the sample space, and each position index within the sample space is used as a clustering sample point; the formula is used... Determine the number of clusters K for each sitting posture. The symbol represents the rounding up sign; where C is the number of sitting posture categories, and F is a preset redundancy coefficient used to control the redundancy of the initial screening position index in K-means clustering, F>1; K-means clustering is performed on each sample space to obtain K cluster centers for each sample space. The cluster centers calculated at this time are the mean positions of each sample point, so they are usually represented as floating-point coordinates. Each cluster center has the same form as the position index and serves as an initial screening position index for each sitting posture; the initial screening position indices of all sitting postures are summarized, and duplicate initial screening position indices are deduplicated to obtain a global candidate position index set.

[0016] Furthermore, the specific method for performing K-means clustering on the global candidate location index set to obtain the global preliminary screening location index set is as follows: using the global candidate location index set as the sample space, each candidate location index in the sample space is used as a cluster sample point; using the total number of target points P as the number of clusters, K-means clustering is performed on the sample space to obtain P cluster centers, each cluster center having the same form as the location index, serving as a preliminary screening location index; using the P cluster centers, i.e., the P preliminary screening location indices, constitutes the global preliminary screening location index set.

[0017] Furthermore, a conflict resolution mechanism and a point compensation mechanism are introduced to discretize the global initial screening location index set to obtain the global optimal location index set. The specific method is as follows: Each initial screening location index in the global initial screening location index set is rounded to the nearest integer, resulting in P candidate location indices. The P candidate location indices are then deduplicated. If duplicates exist, conflict resolution and point compensation are performed according to the following priority order: Priority 1: For each duplicate candidate location index in the global key location index set, a new location index with the closest Euclidean distance and no duplicates is selected; Priority 2: If Priority 1 cannot be achieved, the same operation is performed in the expanded key location index set; Priority 3: If Priority 2 still cannot be achieved, the same operation is performed in the location index matrix; The P location indices obtained after deduplication, conflict resolution, and point compensation are taken as the global optimal location index set.

[0018] Secondly, the present invention also provides a system for implementing the above-mentioned screening method, including a data acquisition circuit and a data processing module. The data acquisition circuit is used to read the pressure data collected by the pressure sensor array and output it to the data processing module. The data processing module is used to construct samples and perform the screening work of the screening method.

[0019] In a further technical solution, the data processing module is equipped with a program to implement the filtering method, and the filtering method is implemented by executing the program.

[0020] Compared with the prior art, the advantages of the present invention are as follows: while significantly reducing the hardware cost, reducing the amount of data, and reducing the complexity of the data processing unit, the sitting posture recognition accuracy is maintained, effectively solving the technical problems in the prior art that due to the excessive number of pressure sensors, the cost is high, the amount of data is large, and the recognition efficiency is low, and the pressure sensors in the key areas are discarded due to the interlaced and inter-column screening, and the rectangular layout cannot accurately represent the pressure distribution characteristics. The specific advantages are described as follows: (1) The present invention breaks the rectangular layout caused by the interlaced and inter-column screening of pressure sensors in the existing screening method, avoids the edge pressure sensors with less effect on sitting posture discrimination, and the problem that the pressure sensors in the key stress areas are discarded, making the P pressure sensors selected show an irregular distribution pattern, accurately fitting the actual pressure distribution characteristics of the human sitting posture; (2) By performing class-specific clustering, global fusion clustering, and discretization on the key position index sets of all preset sitting postures, the comprehensively optimal nature of the screening results for all sitting postures is ensured, making the sitting posture recognition accuracy not lower than the accuracy when using M×N pressure sensors; (3) Since P<M×N, the number of pressure sensors used in the human sitting posture recognition hardware system of the intelligent seat is significantly reduced. Not only the pressure data entering the subsequent processing is reduced, the computing power requirement for the data processing unit is correspondingly reduced, the complexity of the subsequent data processing unit is reduced, but also the hardware cost and power consumption are greatly reduced, and it can be widely applied to the design of the human sitting posture recognition hardware system of various intelligent seats. Description of the Drawings

[0021] Figure 1 It is the structural diagram of the system of the present invention; Figure 2 It is the flowchart of the screening method of the present invention; Figure 3 It is the average pressure heat map of seven types of sitting postures in the screening method of Embodiment 1 of the present invention; Figure 4 It is the schematic diagram of the key position index set of seven types of sitting postures in the screening method of Embodiment 1 of the present invention; Figure 5 It is the schematic diagram of the global key position index set in the screening method of Embodiment 1 of the present invention; Figure 6 It is the schematic diagram of the extended key position index set in the screening method of Embodiment 1 of the present invention; Figure 7 It is the schematic diagram of the screening result of class-specific clustering in the first stage in the screening method of Embodiment 1 of the present invention; Figure 8 It is the schematic diagram of the global candidate position index set in the screening method of Embodiment 1 of the present invention; Figure 9This is a schematic diagram of the preliminary results of the second stage of global fusion clustering in the screening method of Embodiment 1 of the present invention; Figure 10 This is a schematic diagram illustrating the distribution of the globally optimal position index set in the filtering method of Embodiment 1 of the present invention; Figure 11 This is an average pressure thermogram of four sitting postures in the screening method of Embodiment 2 of the present invention; Figure 12 This is a schematic diagram of the key position index set for four types of sitting postures in the screening method of Embodiment 2 of the present invention; Figure 13 This is a schematic diagram of the global key location index set in the filtering method of Embodiment 2 of the present invention; Figure 14 This is a schematic diagram of the expanded key position index set in the filtering method of Embodiment 2 of the present invention; Figure 15 This is a schematic diagram of the first-stage category-specific clustering screening results in the screening method of Embodiment 2 of the present invention; Figure 16 This is a schematic diagram of the global candidate position index set in the filtering method of Embodiment 2 of the present invention; Figure 17 This is a schematic diagram of the preliminary results of the second stage of global fusion clustering in the screening method of Embodiment 2 of the present invention; Figure 18 This is a schematic diagram of the distribution of the global optimal position index set in the filtering method of Embodiment 2 of the present invention. Detailed Implementation

[0022] The core concept of this invention lies in creatively constructing, layer by layer, a high-pressure pixel set, a key position index set, a global key position index set, and an extended key position index set. This is combined with discretization using category-specific clustering, global fusion clustering, conflict resolution mechanisms, and point compensation mechanisms. Within the effective range of the pressure sensor array, this allows for the effective differentiation between edge pressure sensors and pressure sensors in key stress areas, enabling the selection of pressure sensors that accurately reflect the actual pressure distribution characteristics of a human sitting posture. This invention provides smart chair manufacturers with a cost-reducing and efficiency-enhancing solution for new human posture recognition hardware systems.

[0023] The technical solution of the present invention will be described in detail and completely below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can understand and implement it. The embodiments described in this section are only used to illustrate the technical solution of the present invention, and are not intended to limit the scope of protection of the present invention.

[0024] Building such Figure 1The screening system shown uses a computer as the data processing module. A pressure sensor array is mounted on the seat cushion of the smart chair. A data acquisition circuit is connected to both the pressure sensor array and the computer. This data acquisition circuit utilizes mature products and consists of a main component such as an MCU, analog switches, and operational amplifiers. The analog switches form the conduction circuit for the pressure sensor array, while the operational amplifiers and peripheral circuits such as capacitors and resistors form the voltage acquisition circuit. The MCU is responsible for both circuit conduction and voltage acquisition, transmitting the collected pressure data from the pressure sensor array to the computer. The computer is equipped with a Python 3.8+ environment, and Visual Studio Code is used for algorithm development and debugging. The main open-source libraries used in the experiment include: NumPy (numerical computation and matrix operations), Pandas (data structuring), Scikit-learn (K-means clustering algorithm implementation), Matplotlib (pressure heatmap plotting and visualization), and OpenCVPython (image grayscale conversion and normalization). One hundred participants were selected as experimenters, exhibiting diversity in body type, gender, and age to ensure representativeness and generalization ability of the sample.

[0025] Example 1: As Figure 2 As shown, a method for screening pressure sensor arrays for human posture recognition is used to screen a 32×32 pressure sensor array to obtain a target number of pressure sensor points to construct a human posture recognition hardware system. The method includes the following steps: Step 1: Pre-store the position index matrix of the pressure sensor array on the computer and set seven typical sitting postures: leaning back, leaning forward, leaning left, leaning right, crossing the left leg, crossing the right leg, and sitting upright. Each tester sits on the pressure sensor array in each posture for a preset sampling time. The pressure sensor array samples 20 times according to a preset sampling period. The data acquisition circuit acquires the pressure data from all 20 pressure sensor samples and transmits it to the computer. Each time the computer receives pressure data, it distributes the pressure data according to the 32 rows and 32 columns corresponding to the position index matrix to form a pressure matrix (sample). Each tester obtains 20 samples for each sitting posture, for a total of 140 samples, and all testers obtain a total of 14,000 samples.

[0026] Step 2: Perform linear normalization on each pressure data point in each sample, mapping it to the [0, 1] interval, then further mapping it to the [0, 255] interval and converting it to 8-bit grayscale values. Then, convert each normalized 32×32 pressure matrix into a corresponding pressure heatmap, mapping each pixel position in the heatmap to the pressure sensor index position, with the grayscale value equal to the normalized pressure value. For the same sitting posture, accumulate the grayscale values ​​of the same pixel position in 2000 pressure heatmaps point by point, and divide by the total number of samples corresponding to that posture (2000) to obtain the average pressure value at that pixel position. By performing the above operation on all pixel positions in 2000 pressure heatmaps, a 32×32 average pressure heatmap can be constructed, as shown below. Figure 3 As shown. Figure 3 In the diagram, 1-7 represent the average pressure heatmaps for leaning back, leaning forward, leaning left, leaning right, crossing the left leg, crossing the right leg, and sitting upright, respectively. These heatmaps characterize the spatial distribution of pressure sensor arrays under seven sitting postures. The grayscale values ​​of all pixels in the average pressure heatmap for each posture are sorted from largest to smallest, and the top 60%, or 614 pixels, are selected to form a high-pressure pixel set.

[0027] Step 3: For the average pressure heatmap of each sitting posture, take the vertical central axis of the average pressure heatmap as the axis of symmetry. For any high-pressure pixel, determine whether its mirror pixel about the central axis is in the set of high-pressure pixels. If not, its mirror pixel is a left-right symmetrical pixel.

[0028] Step 4: The set of high-pressure pixels for each sitting posture is merged with the position indices of all its symmetrical pixels to form a key position index set, such as... Figure 4 The location of the white area in the middle is shown. Figure 4 In the table, 1-7 represent the key position index sets for leaning back, leaning forward, leaning left, leaning right, crossing the left leg, crossing the right leg, and sitting upright, respectively.

[0029] Step 5: Perform a union operation on the key position index sets obtained for each of the seven sitting postures to obtain the global key position index set, such as... Figure 5 The location of the white area in the middle is shown.

[0030] Step 6: Set the neighborhood expansion number k to 1. In the position index matrix, using each position index in the global key position index set as the center, expand by 1 position index in each of the eight directions (up, down, left, right, and four diagonal directions). If there is no expandable position index in the position index matrix, do not expand. Merge the expanded position index with the global key position index set to obtain the expanded key position index set, such as... Figure 6 As shown, Figure 6The white area represents the set of global key location indices, while the gray area represents the extended location index area.

[0031] Step 7: Set P=400, redundancy factor F=1.5, and the number of clusters in each cluster class is... The first stage of clustering (category-specific clustering) was performed on the key position index set for each sitting posture. At this stage, 86 candidate position indices were obtained for each sitting posture, resulting in a total of 86 × 7 = 602 candidate position indices. Figure 7 The black dot is shown in the white area.

[0032] It should be noted that in practical applications, P can be selected through multiple experiments based on the sensor array size to ensure the detection accuracy of the subsequent pressure sensor array.

[0033] Step 8: Summarize and deduplicate the 602 candidate location indices to obtain a global candidate location index set, such as... Figure 8 The black dot is shown in the white area.

[0034] Step 9: Perform a second-stage clustering and filtering (global fusion clustering) on ​​the global candidate location index set to obtain the global initial screening location index set, such as... Figure 9 The black dot is shown in the white area.

[0035] Step 10: Since the cluster centers obtained from the second stage clustering are continuous coordinates, while the actual pressure sensors need to be deployed discretely, the global initial screening location index set needs to be discretized. At this time, the initial screening location index is rounded to the nearest integer to map it to a discrete location, resulting in 400 candidate location indices.

[0036] Step 11: Deduplicate the 400 candidate location indices. If duplicates exist, resolve conflicts and compensate for the locations according to priority. The specific method is as follows: Step 11-1: Randomly arrange the 400 candidate position indices in chronological order, and take the first candidate position index as the optimal position index; and traverse from the second one in chronological order. For each index, check if there is an optimal position index with the same value. If not, take the candidate position index as the optimal position index. If it exists, proceed to step 11-2. Step 11-2: Find a location index that is the same as the candidate location index in the global key location index set. Perform a neighborhood search in the global key location index set with this location index as the center. If there is a location index that is different from all the current best location indexes, select the location index with the closest Euclidean distance to this location index as the best location index. If there is no location index that is different from all the current best location indexes, proceed to step 11-3. Step 11-3: Find a location index that is the same as the candidate location index in the extended key location index set. Perform a neighborhood search in the global key location index set with this location index as the center. If there is a location index that is different from all the current best location indexes, select the location index with the closest Euclidean distance to this location index as the best location index.

[0037] After this step is completed, all duplicate candidate location indices will have a corresponding optimal location index found in the global key location index set or the extended key location index set.

[0038] It should be noted that in practical applications, if after this step, if there are duplicate candidate location indices and no optimal location index is found in either the global key location index set or the extended key location index set, the search range is expanded. A location index that is the same as the candidate location index is found in the location index matrix. A neighborhood search is performed in the global key location index set with this location index as the center. Among the location indices that are different from all the current optimal location indices, the location index with the closest Euclidean distance to this location index is taken as the optimal location index.

[0039] At this point, 400 optimal location indices are obtained. These 400 optimal location indices correspond to the location indices of 400 pressure sensors, forming a globally optimal location index set, the distribution of which is shown in the figure below. Figure 10 As shown. From Figure 10 As can be seen, the position indices do not overlap in the discrete space and are all located within the global key position index set or the extended key position index set. This indicates that the screening method in this embodiment accurately and effectively screens the 32×32 pressure sensor array, not only reducing the number of pressure sensors used from 1024 to 400, significantly reducing the number of pressure sensors, but also ensuring that the screened pressure sensors are all located in the key stress area, thus guaranteeing the effective representation of human sitting posture characteristics and not reducing the accuracy of subsequent sitting posture recognition.

[0040] Example 2: This example is basically the same as Example 1, with the main differences being: the pressure sensor array size is 20×20; the sitting posture types are set to four: leaning forward, sitting upright, leaning left, and leaning right; P=144, redundancy coefficient F=1.5, and the number of clusters per class is... .

[0041] Based on the above differences, in this embodiment, 8000 samples are first obtained. Based on these 8000 samples, normalization, heatmap generation, averaging calculation, high-pressure extraction, and symmetry constraint processes are performed sequentially to obtain average pressure heatmaps for four types of sitting postures, as follows: Figure 11 and Figure 12 As shown. Figure 11 In the diagram, 1-4 represent the average pressure thermograms for leaning forward, sitting upright, leaning to the left, and leaning to the right, respectively. Figure 12 In the table, 1-4 represent the key position index sets for leaning forward, sitting upright, leaning left, and leaning right, respectively.

[0042] Perform a union operation on the key position index sets obtained for the four sitting postures to obtain a global key position index set, as shown below. Figure 13 The location of the white area is shown in the image. Based on this, the global key location index set is expanded to obtain an expanded key location index set, as shown below. Figure 14 As shown, Figure 14 The white area represents the set of global key location indices, while the gray area represents the extended location index area.

[0043] The first stage of clustering (category-specific clustering) was performed on the key position index set for each sitting posture. At this stage, 54 candidate position indices were obtained for each sitting posture, resulting in a total of 54 × 4 = 216 candidate position indices. Figure 15 The black dot is shown in the white area.

[0044] The 216 candidate location indices are aggregated and deduplicated to obtain a global candidate location index set, such as... Figure 16 The black dots in the white area are shown. Using the target number of 144 as the cluster number, a second-stage clustering and filtering (global fusion clustering) is performed on the global candidate location index set to obtain the initial global location index set, as shown below. Figure 17 The white area is shown as a black dot. Discretizing this area yields a global optimal position index set consisting of 144 optimal position indices.

[0045] The distribution diagram of the above globally optimal location index set is as follows: Figure 18 As shown. From Figure 18 It can be seen that the position indices do not overlap in the discrete space and are all located within the global key position index set or the extended key position index set. This indicates that the screening method in this embodiment accurately and effectively screens the 20×20 pressure sensor array, significantly reducing the number of pressure sensors from 400 to 144. Furthermore, the screened pressure sensors are all located in key stress areas, ensuring effective representation of human sitting posture characteristics and not reducing the accuracy of subsequent sitting posture recognition.

Claims

1. A screening method for a pressure sensor array for human sitting posture recognition, which is used to screen out P pressure sensors from the pressure sensor array to construct a human sitting posture recognition hardware system. The pressure sensor array includes M rows N distributed in columns M×N pressure sensors, where P < M × N; It is characterized in that Multiple samples are obtained for each of the various preset sitting postures. Each sample is a two-dimensional matrix composed of pressure data collected by all pressure sensors at the same time. An average pressure heatmap is generated for each sitting posture based on all samples. The location indices of pixels that satisfy the preset high pressure characteristics and symmetry constraints are extracted from the average pressure heatmap of each sitting posture to form a set of key location indices. Based on the set of key location indices for all sitting postures, category-specific clustering, global fusion clustering, and discretization are performed to obtain a globally optimal set of location indices containing P location indices.

2. The screening method for pressure sensor arrays for human sitting posture recognition according to claim 1, characterized in that, The specific method for obtaining multiple samples of each of the various preset sitting postures is as follows: At least 100 test subjects with diversity in body type, gender, and age are selected; each test subject sits in a smart chair equipped with the pressure sensor array in each posture, maintaining that posture for a preset sampling time. The pressure sensor array is controlled to sample according to a preset sampling period, and the pressure data collected by all pressure sensors in each sampling period is obtained. M OK N The columns form a two-dimensional matrix.

3. The screening method for pressure sensor arrays for human sitting posture recognition according to claim 1, characterized in that, The position index of the pressure sensor in the m-th row and n-th column is set to (m, n), where m = 1, 2, ..., M, and n = 1, 2, ..., N; the pixel size of the average pressure heatmap is... M×N The pixel in the m-th row and n-th column represents the pressure sensor in the m-th row and n-th column, with a position index of (m, n). The specific method for generating the average pressure heatmap for each sitting posture based on all samples is as follows: the minimum pressure data among all samples for each sitting posture is taken as its global minimum pressure value, and the maximum pressure data is taken as its global maximum pressure value. Based on the global minimum pressure value and global maximum pressure value for each sitting posture, each sample for each sitting posture is linearly scaled and mapped to [0, ..., ... The values ​​in the [1] interval are then mapped to the [0,255] interval and converted to 8-bit grayscale values ​​to obtain the normalized sample corresponding to each sample of each sitting posture. The 8-bit grayscale value in the m-th row and n-th column of the normalized sample corresponding to each sample of each sitting posture is used as the grayscale value of the pixel in the m-th row and n-th column of the pressure heat map of that sample to obtain the pressure heat map of each sample of each sitting posture. The average value of the grayscale value of the pixel in the m-th row and n-th column of the pressure heat map of all samples of each sitting posture is used as the grayscale value of the pixel in the m-th row and n-th column of the average pressure heat map of that sitting posture to obtain the average pressure heat map of each sitting posture.

4. The screening method for pressure sensor arrays for human sitting posture recognition according to claim 1, characterized in that, The specific method for extracting the position indices of pixels that satisfy the preset high-pressure characteristics and symmetry constraints from the average pressure heatmap of each sitting posture to form the key position index set for each sitting posture is as follows: For all pixels in the average pressure heatmap of each sitting posture, sort them in descending order of their grayscale values. If multiple pixels have the same grayscale value, then these multiple pixels are randomly arranged, and the pixels with the highest ranking according to the preset proportion are taken as high-pressure pixels, forming the high-pressure pixel set for each sitting posture. For the high-pressure pixel set for each sitting posture, determine whether the pixels that are mirror-symmetrical about the vertical central axis of the average pressure heatmap of that sitting posture are located in the high-pressure pixel set for that sitting posture. If not, then the pixel is the left-right symmetrical pixel in the high-pressure pixel set for that sitting posture. The position indices of the pixels in the high-pressure pixel set for each sitting posture and all the left-right symmetrical pixels together form the key position index set for each sitting posture.

5. The screening method for pressure sensor arrays for human sitting posture recognition according to claim 1, characterized in that, Based on the key position index set of all sitting postures, the specific method for obtaining the globally optimal position index set through category-specific clustering, global fusion clustering, and discretization is as follows: The union of the key position index sets of all sitting postures is taken as the global key position index set; the global key position index set is expanded to obtain an expanded key position index set; K-means clustering is performed on the key position index set of each sitting posture to obtain a global candidate position index set; K-means clustering is performed on the global candidate position index set to obtain a global preliminary screening position index set; a conflict resolution mechanism and a point compensation mechanism are introduced to discretize the global preliminary screening position index set to obtain the globally optimal position index set.

6. The screening method for pressure sensor arrays for human sitting posture recognition according to claim 5, characterized in that, The specific method for expanding the global key location index set to obtain the expanded key location index set is as follows: Set a positive integer k as the neighborhood expansion number; then... M×N The location index of each pressure sensor corresponds to... M OK N The columns are distributed to form a position index matrix. For each position index in the global key position index set, in the position index matrix, with the position index as the center, k adjacent position indices are obtained in the eight directions of up, down, left, right and four diagonal directions as extended position indices. If the number of adjacent position indices in a certain direction is less than k, the actual number of adjacent position indices is obtained. All extended position indices are added to the global key position index set and deduplication is performed to obtain the extended key position index set.

7. The method for screening pressure sensor arrays for human sitting posture recognition according to claim 5, characterized in that, The specific method for performing K-means clustering on the key position index set for each sitting posture to obtain the global candidate position index set is as follows: The key position index set for each sitting posture is used as the sample space, and each position index within the sample space is treated as a clustering sample point; the formula is used... Determine the number of clusters K for each sitting posture. is the rounding up sign; where C is the number of sitting posture categories, and F is a preset redundancy coefficient used to control the redundancy of the initial screening position index in K-means clustering, F>1; K-means clustering is performed on each sample space to obtain K cluster centers for each sample space. Each cluster center has the same form as the position index and serves as an initial screening position index for each sitting posture; the initial screening position indices of all sitting postures are summarized, and duplicate initial screening position indices are deduplicated to obtain a global candidate position index set.

8. The method for screening pressure sensor arrays for human sitting posture recognition according to claim 5, characterized in that, The specific method for performing K-means clustering on the global candidate location index set to obtain the global preliminary screening location index set is as follows: using the global candidate location index set as the sample space, each candidate location index in the sample space is used as a cluster sample point; using the total number of target points P as the number of clusters, K-means clustering is performed on the sample space to obtain P cluster centers, each cluster center having the same form as the location index, serving as a preliminary screening location index; using P cluster centers, i.e., P preliminary screening location indices, to form the global preliminary screening location index set.

9. The method for screening pressure sensor arrays for human sitting posture recognition according to claim 5, characterized in that, The global initial screening location index set is discretized using conflict resolution and location compensation mechanisms to obtain the globally optimal location index set. Specifically, each initial screening location index in the global initial screening location index set is rounded to the nearest integer, resulting in P candidate location indices. These P candidate location indices are then deduplicated. If duplicates exist, conflict resolution and location compensation are performed according to the following priority order: Priority 1: For each duplicate candidate location index in the global key location index set, a new, non-duplicate location index with the closest Euclidean distance is selected; Priority 2: If Priority 1 cannot be implemented, the same operation is performed in the expanded key location index set; Priority 3: If Priority 2 still cannot be implemented, the same operation is performed in the location index matrix; The P location indices obtained after deduplication, conflict resolution, and location compensation are taken as the globally optimal location index set.

10. A system for implementing the screening method according to any one of claims 1-9, characterized in that, It includes a data acquisition circuit and a data processing module. The data acquisition circuit is used to read the pressure data collected by the pressure sensor array and output it to the data processing module. The data processing module is used to construct samples and perform the screening work of the screening method.

Images