Human body posture recognition method and system based on block hash and bloom filter

Abstract

The invention relates to the technical field of crossing of computer vision and privacy security, in particular to a human body posture recognition method and system based on block Hash and a Bloom filter, which are suitable for a private place edge device deployment scene, and the process comprises the following steps: preprocessing posture data of a private place, quantifying 40-dimensional human body posture features and processing exceptions; salt is added in blocks according to human body joint functions to generate a Hash combination, and lightweight transmission is carried out through MQTT-SN + TLS1.3; according to the human body posture recognition method and system based on the block Hash and the Bloom filter, the edge end only transmits the Hash, the retrieval time consumption of the rear end is smaller than or equal to 15 ms, the total delay is smaller than or equal to 80 ms, the misjudgment rate is smaller than or equal to 0.3%, privacy extreme protection is achieved, recognition is efficient, the deployment cost is low, and dynamic adaptation is high. And the real-time early warning requirement is met.

Description

Technical Field

[0001] This invention relates to the field of computer vision and privacy security interdisciplinary technology, specifically a human pose recognition method and system based on block hashing and Bloom filters. Background Technology

[0002] Existing human pose recognition solutions mainly include three data transmission methods: directly transmitting raw image data, transmitting only the recognition result (pose category), and transmitting quantized feature vectors. Existing feature vectors are usually high-dimensional, and no clustering deduplication and dynamic update mechanism is used when pre-storing pose templates. Traditional pose recognition requires transmitting feature vectors first and then performing weighted calculations. Related technical solutions are widely used in human pose recognition scenarios to meet the pose recognition needs of different scenarios.

[0003] Directly transmitting raw image data can easily lead to privacy leaks, and the large data transmission volume reduces system real-time performance and increases bandwidth costs. Transmitting only the recognition results prevents the backend from verifying them, and adding recognition types requires upgrading all client algorithms, increasing maintenance difficulty. While transmitting quantized feature vectors can remove some sensitive information, the vectors still contain pose-related details, posing a risk of reverse engineering to infer privacy. Furthermore, existing feature vectors are typically high-dimensional, hindering storage and rapid retrieval on edge devices. Adding pose categories or upgrading algorithm recognition accuracy requires retraining the model, and the lack of clustering deduplication and dynamic update mechanisms for pre-stored pose templates results in redundant storage. Traditional pose recognition often has a total latency exceeding 200ms and lacks multi-level threshold judgment, leading to a false positive rate exceeding 1%, impacting the reliability of practical applications. Therefore, there is an urgent need to develop a human pose recognition method and system based on block hashing and Bloom filters to overcome the shortcomings of current practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a human pose recognition method and system based on block hashing and Bloom filters to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] The human pose recognition method based on block hashing and Bloom filters includes the following steps:

[0007] (1) Preprocess the human posture data in private locations, remove invalid data and standardize the data;

[0008] (2) Quantize the human posture features of the preprocessed posture data to obtain multidimensional structured quantitative features;

[0009] (3) The multidimensional quantitative features are divided into blocks based on the correlation of human joint functions, and a hash combination is generated by combining the salt value;

[0010] (4) The hash combination is transmitted to the backend using a lightweight encrypted transmission method;

[0011] (5) The backend builds a pose-specific Bloom filter to perform fast similarity retrieval on the received hash combinations to realize human pose recognition;

[0012] (6) Dynamically manage the pose prototype library and the corresponding Bloom filter to adapt to pose updates and scene changes.

[0013] As a further aspect of the present invention: in step (1), the preprocessing includes directional acquisition optimization, redundancy removal processing, and normalization processing;

[0014] The targeted acquisition optimization achieves accurate acquisition of posture data by specifying the core area of ​​human activity and combining multi-device collaborative acquisition;

[0015] The normalization process establishes a coordinate system with a preset joint point as the origin and performs size scaling based on the maximum joint spacing.

[0016] As a further aspect of the present invention: in step (2), the multidimensional quantification feature is 40-dimensional, including four categories: binary attributes, proportional attributes, angular attributes, and trend attributes;

[0017] The 40-dimensional features are constructed based on the positional relationships, movement trends, and morphological features of 16 key human body points.

[0018] If feature extraction fails in a certain dimension, the average feature value of historical frames is used to fill the gap.

[0019] As a further aspect of the present invention: in step (3), the block division adopts a sliding window method, which groups features that are functionally coordinated and spatially adjacent into the same window, and sets overlapping features in adjacent windows;

[0020] The salt value is generated by the edge device through a hardware random number generator, stored in a secure storage area and updated periodically. During the update, only the salt value hash is synchronized, not the original salt value.

[0021] As a further aspect of the present invention: the hash combination is generated in the following way:

[0022] After the quantization features of each window are superimposed with salt values, a hash code is generated by a hash algorithm, and the hash codes of multiple windows are combined to form the hash combination.

[0023] The hash result for each window also has a checksum added for anomaly detection.

[0024] As a further aspect of the present invention: In step (4), the lightweight encrypted transmission adopts the MQTT-SN protocol combined with the TLS 1.3 encryption layer, and the transmitted data includes hash combination, device identifier, timestamp and check code, and is encapsulated in TLV format.

[0025] As a further aspect of the present invention: in step (5), the construction of the Bloom filter includes:

[0026] The filter bit length and hash function number are calculated based on the attitude core hash combination. The hash combination is mapped to the corresponding index bit of the filter and set to 1 through multiple independent hash functions.

[0027] The similarity retrieval adopts a multi-level threshold determination mechanism, calculates a similarity score based on the proportion of the matching window, and determines the posture category based on the score result.

[0028] As a further aspect of the present invention: in step (6), the dynamic management includes incremental learning of the prototype library, elimination and updating of the Bloom filter, and synchronous restoration of the salt value;

[0029] The prototype library incremental learning obtains core pose features through clustering deduplication.

[0030] The filter replacement and update are based on the call frequency and recognition accuracy.

[0031] The salt value is synchronously restored by transmitting a new salt value encrypted after device authentication.

[0032] A human pose recognition system based on block hashing and Bloom filters includes:

[0033] The data acquisition module is used to collect human posture data in a targeted manner in private locations;

[0034] The preprocessing module is used to perform redundancy removal and normalization preprocessing on the acquired attitude data;

[0035] The feature quantization module is used to quantize the preprocessed pose data and generate multi-dimensional structured quantized features.

[0036] The block-salted hash module is used to divide quantified features into blocks based on human joint function and generate hash combinations by combining salt values.

[0037] A lightweight transmission module is used to transmit hash combinations to the backend via an encrypted transmission protocol;

[0038] The backend retrieval module is used to build a Bloom filter to perform fast similarity retrieval on the received hash combinations and realize pose recognition.

[0039] The dynamic management module is used for incremental updates, elimination, and salt value restoration management of the attitude prototype library and Bloom filters.

[0040] As a further aspect of the present invention: the data acquisition module includes an infrared camera and a millimeter-wave radar, the infrared camera supports environmental adaptive adjustment, and the millimeter-wave radar is used for blind spot acquisition in occluded scenarios;

[0041] The backend retrieval module has a retrieval time of ≤15ms, a total system latency of ≤80ms, and a false positive rate of ≤0.3%.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. Extreme privacy protection: Only 80-byte hash combination is transmitted at the edge, without transmitting any feature vectors or sensitive information. Sensitive information is destroyed locally in real time. The salt value is stored in the secure storage area of ​​the edge device and is not synchronized with the original salt value. Even if the transmission link is hijacked, attackers cannot restore the posture or privacy, effectively solving the data leakage risk of traditional solutions.

[0044] 2. Significantly improved recognition efficiency: The backend uses a posture-specific Bloom filter to achieve fast hash code matching, with a retrieval time of ≤15ms, which is 60% faster than traditional feature vector comparison (≥50ms). The total system latency is ≤80ms, which can meet the real-time early warning needs of elderly people falling.

[0045] 3. Extremely low deployment cost: The amount of hash calculation for edge devices is reduced by 30% (SHA-128 takes less time than SHA-256), the storage size of a single-pose Bloom filter in the backend is ≤360 bytes, and the storage size of historical data for one year is ≤500MB, which is beneficial for the storage and lightweight deployment of edge devices;

[0046] 4. Strong dynamic adaptability: Supports salt value loss recovery and dynamic updates of Bloom filters. The time for adding a pose is ≤30 seconds, the time for updating the filter is ≤15 seconds, the time for salt value loss recovery is ≤10 seconds, and the false judgment rate is ≤0.3%. It is suitable for adding and removing poses in multiple scenarios such as home and medical. By clustering and deduplication, the storage of the prototype library is reduced, avoiding storage redundancy. At the same time, it can cope with the decline in pose recognition accuracy and ensure the stability of recognition results. Attached Figure Description

[0047] Figure 1 This is an overall flowchart of the human pose recognition method based on block hashing and Bloom filter in an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0050] Please see Figure 1 The human posture recognition method and system provided by the embodiments of the present invention for privacy settings (home care, hospital wards and private fitness spaces, etc.) are particularly suitable for small and medium-sized scenarios with high requirements for edge device deployment, real-time early warning, and strong privacy sensitivity. It aims to solve the problems of contradiction between privacy protection and data transmission, low efficiency of feature storage and retrieval, and imbalance between real-time performance and accuracy in the prior art, and achieve the goals of extreme privacy protection, efficient similarity retrieval, lightweight deployment, and dynamic adaptation and fault tolerance.

[0051] I. Preprocessing of posture data in private locations

[0052] (I) Optimization of Targeted Data Acquisition

[0053] Equipment selection: An infrared camera with AI region cropping is used (such as Hikvision DS-2CD3T46WD-L). This camera supports 10-30℃ environmental adaptation, which can effectively avoid light interference and ensure stable acquisition of attitude data in different temperature environments. At the same time, it is paired with millimeter-wave radar (such as TI IWR6843) to assist in blind spot compensation, which can effectively solve the problem of attitude data acquisition in occluded scenarios and ensure the integrity of the acquired data.

[0054] Region delineation: The "core area of ​​human activity" can be manually selected through an authorized terminal (such as the caregiver's mobile phone). The system will automatically save the coordinates of the area. For example, in a home setting, it can be set to "x:0-500px, y:200-800px". During the collection process, only the data of the core area is retained, and the background removal rate is ≥95%, which greatly reduces the amount of invalid data and reduces the pressure of subsequent data processing.

[0055] (ii) Redundancy Removal and Normalization

[0056] Edge detection: The Canny operator threshold is adaptively adjusted by ambient light. When the light intensity is >500 lux, the upper limit of the Canny operator is set to 190 and the lower limit is set to 90; when the light intensity is <200 lux, the upper limit is set to 160 and the lower limit is set to 70. This setting ensures that the contour extraction accuracy is ≥98% under different lighting conditions, providing a clear contour basis for subsequent pose feature extraction.

[0057] Normalization: With the hip joint as the origin (automatically located by the MobilePose model with a positioning error ≤2px), the Z-axis of the three-dimensional coordinate system is set to be perpendicular to the ground. The size is scaled using the "maximum skeleton length normalization" method. The normalization value is calculated as: Normalized value = Original value / Maximum joint spacing. This normalization process can effectively eliminate the influence of different human body shapes on posture recognition and improve the universality of recognition.

[0058] II. Quantification of Human Posture Comparison Features

[0059] (I) Selection of key points of the human body

[0060] Sixteen key points of the human body located by the MobilePose model were selected and numbered. The correspondence between the key point numbers and names is shown in the table below:

[0061]

[0062] (II) 40-dimensional feature quantization

[0063] Feature quantization methods: The 40-dimensional features are divided into four categories: binary attributes, proportional attributes, angular attributes, and trend attributes. The quantization methods and examples for each type of feature are shown in the table below:

[0064]

[0065] Feature definition basic parameters: Define a coordinate system with the hip center (number 10) as the origin, with the Z-axis perpendicular to the ground and upward (height direction), the X-axis horizontal left and right (left negative and right positive, consistent with the line connecting the shoulder center and hip center), and the Y-axis horizontal front and back (front positive and back negative, perpendicular to the XZ plane); define the maximum joint distance Lmax as the maximum value of the pairwise Euclidean distance between all 16 key points; set the key point positioning confidence Conf≥0.8 as "valid recognition", and Conf<0.8 as "occlusion / recognition failure".

[0066] The specific definition of 40-dimensional features is as follows: The sequence number, name, and calculation basis (key point number) of the 40-dimensional features are shown in the table below:

[0067]

[0068]

[0069]

[0070] (III) Quantitative Anomaly Handling

[0071] If feature extraction in a certain dimension fails (e.g., due to keypoint occlusion), the average value of the feature in that dimension from the previous 3 frames is used to fill the gap. The filling error is ≤5%, which can effectively avoid subsequent hash generation failure due to missing features and ensure the stability and integrity of the feature quantization process.

[0072] III. Block-based salted hash generation

[0073] (I) Salinity generation and management

[0074] Salt value generation: When the edge device starts up for the first time, a 16-byte salt value is generated by a hardware random number generator. The salt value format is "8-bit device SN + 8-bit random number" (such as SN12345678_9A7B5C3D), which ensures the uniqueness and randomness of the salt value.

[0075] Salt value storage and update: The salt value is stored in a secure storage area on the edge device (such as eFuse). This storage area is immutable, which can prevent the salt value from being maliciously modified. The salt value is automatically updated once every 6 months. During the update, it is synchronized to the backend via TLS 1.3 encryption. Only the salt hash is synchronized, not the original salt value, to avoid the salt value being leaked during the synchronization process and to ensure the security of the salt value.

[0076] (II) Implementation of Block Hash

[0077] Blocking Criteria: Based on the correlation of human joint functions and the robustness of feature recognition, combined with the requirements of posture recognition and hash calculation efficiency, the 40-dimensional quantized features are grouped according to human joint functions. Features with functional coordination and spatial proximity are grouped into the same window. For example, the head-neck (features 1-8) corresponds to the upper trunk core control unit, and the knee-ankle (features 13-20) corresponds to the lower limb motor unit. This ensures that features in the same window can accurately reflect the local posture state and avoids the decrease in hash recognition caused by the mixing of cross-functional features. At the same time, a sliding window overlap compensation mechanism is adopted, with adjacent windows overlapping 4-dimensional features (such as the last 4 dimensions of the head-neck window overlapping with the first 4 dimensions of the shoulder-elbow window). This compensates for the feature fragmentation problem that may be caused by fixed blocking. The overlapping part can strengthen the posture correlation at the joint connection (such as the angular linkage between the neck and the shoulder) and improve the adaptability of the hash code to continuous posture changes.

[0078] Hash calculation: The 8-dimensional quantization value of each window (after being converted to ASCII code) is superimposed with a 16-byte salt value and a 128-bit hash code is generated by the SHA-128 hash algorithm; the 40-dimensional quantization features are divided into 5 sliding windows according to "head-neck, shoulder-elbow, wrist-hip, knee-ankle, and whole body trend", forming a "5×128-bit edge hash combination" (80 bytes in total). This hash calculation method can effectively compress the amount of feature data while retaining key posture information.

[0079] Anti-avalanche optimization: Add 2 "check bits" (the sum of the quantized values ​​within the window modulo 4) to the end of the hash result of each window. If a window's hash is abnormal, the abnormal window can be quickly located through the check bits without recalculating all windows, thus improving the efficiency of hash anomaly handling.

[0080] IV. Lightweight Encrypted Transmission

[0081] (a) Transmission Data Structure

[0082] The transmitted data only contains "edge hash combination (80 bytes) + device identifier (last 6 bits of MAC, such as A1B2C3) + millisecond-level timestamp (13 bytes, such as 202510251430567) + CRC32 checksum (4 bytes)", with a total data volume of 80+6+13+4=103 bytes / strip. Only lightweight data such as hash combinations are transmitted, and feature vectors or original images are not transmitted, which greatly reduces the amount of data transmitted and reduces bandwidth usage.

[0083] (ii) Data Format

[0084] Data is encapsulated and transmitted using the TLV (Type-Length-Value) format, for example, Type=0x01 (hash combination), Length=80, Value=... This format facilitates rapid data parsing by the backend and improves data processing efficiency.

[0085] (iii) Transmission Protocol

[0086] The MQTT-SN lightweight protocol is adopted, which is suitable for the low power consumption requirements of edge devices; the Keep-Alive interval is set to 30 seconds to ensure the stability of the connection between the device and the backend; the QoS level is set to 1 (at least one delivery) to avoid data loss in weak network environments; at the same time, a TLS1.3 encryption layer is superimposed (only protecting the link to avoid double encryption loss) to ensure the security of data transmission and prevent data from being hijacked or tampered with.

[0087] V. Backend Bloom Filter Construction and Retrieval

[0088] (a) Constructing filters

[0089] Parameter calculation: For a single posture (e.g., "bedridden"), 20 core hash combinations are pre-stored (n=20), with a false positive rate of p=0.001, calculated using the formula:

[0090] The number of digits m = -n × lnp / (ln2) 2 ;

[0091] The calculation yields:

[0092] m≈-20×ln0.001 / (0.693) 2≈2880 bits (360 bytes);

[0093] Through the formula:

[0094] The number of hash functions k = ceil(m / n × ln2);

[0095] The calculation yields:

[0096] k≈ceil(2880 / 20×0.693)≈10;

[0097] Ten different hash functions, such as MurMurHash3-32 and CityHash, are selected to ensure the uniformity of hash mapping and reduce the risk of misjudgment.

[0098] Construction process: First, obtain the 128-bit hash code of each of the 20 core hash combinations, convert it to big-endian byte order (consistent with the byte order requirements of the MQTT-SN protocol), and then convert it into a continuous binary bit stream without delimiters;

[0099] The second step is to select 10 independent hash functions (MurMurHash3-32, CityHash, etc.), input the 128-bit hash code into the 10 functions respectively, and each function outputs a 32-bit hash value;

[0100] The third step is to use the formula:

[0101] Index bit = (hash function output value) mod 2880;

[0102] The hash value is mapped to the valid index range of the 2880-bit filter, and the corresponding index is set to 1 to complete the construction of a single pose Bloom filter. This construction method ensures that the filter can accurately match the target pose hash code.

[0103] (ii) Similarity Search

[0104] Data processing and matching determination: After receiving the edge hash combination, the backend performs the same processing as the filter construction process on each hash code, and records the number of "set to 1" bits in the 10 mapped index bits as t;

[0105] When t≥8, the window is considered to match;

[0106] The formula for calculating the global similarity score is:

[0107] Score = (Number of matching windows / 5) × 100%;

[0108] Two threshold levels are set: when the score is ≥80%, it is directly determined to be the pose; when the score is 60%≤ and <80%, the hash combination of the previous two frames of the edge device is retrieved and the pose is re-searched. If the scores are both ≥60%, it is determined to be the pose; when the score is <60%, it is determined to be "unrecognized pose" and a backend warning is triggered. This multi-level determination mechanism can improve the accuracy and reliability of pose recognition.

[0109] Retrieval efficiency: The backend uses a Bloom filter to achieve fast hash code matching, with a retrieval time of ≤15ms, which is 60% faster than traditional feature vector comparison (≥50ms), greatly improving the real-time performance of posture recognition and meeting the real-time early warning needs of elderly people falling.

[0110] (III) Handling of Misjudgments

[0111] If the posture is determined to be the target posture, but the backend finds through historical data that the posture does not match the time pattern (e.g., "standing" is determined at 3 am, but usually one is lying in bed at this time), a secondary search is automatically triggered (adding 2 hash functions). By increasing the number of hash functions, the probability of false judgment is further reduced, and the false judgment rate is reduced to ≤0.3%, thus improving the accuracy of posture recognition.

[0112] VI. Dynamic Management of Prototype Library and Filters

[0113] (a) Incremental learning of the prototype library

[0114] Data Acquisition and Clustering: Authorized users label new postures (such as "walking with support against a wall") through their terminals, and edge devices automatically collect 100 sets of quantitative features for this posture. The backend performs K-means clustering (K=20, distance metric is Euclidean distance) on the 100 sets of quantitative features, removes outliers (such as samples whose cluster center distance is greater than 2 standard deviations), and obtains 20 sets of core quantitative features. Before clustering, the quantitative features are normalized (mean subtraction and standard deviation division) to ensure that the clustering results are not affected by the feature magnitude and to ensure the representativeness of the core features.

[0115] Filter construction: Generate a 5×128-bit hash combination based on the core features, construct a new Bloom filter according to the aforementioned filter construction method, and add it to the backend filter set. The whole process takes ≤30 seconds, enabling the rapid addition of new poses and improving the system's adaptability to new poses.

[0116] (II) Filter replacement and renewal

[0117] Elimination mechanism: The backend counts the call frequency of each filter weekly. If a filter is called less than once per day for 30 consecutive days, it is marked as a "redundant filter" and deleted after confirmation by an authorized user, thus releasing storage resources, avoiding storage redundancy, and reducing storage costs.

[0118] Update mechanism: If the recognition accuracy of a certain posture is less than 90% (such as the "falling" posture whose accuracy decreases due to changes in clothing), the system automatically collects 50 new samples, re-clusters them to generate core hash combinations, and updates the corresponding filters. The update time is ≤15 seconds, ensuring that the posture recognition accuracy is always maintained at a high level and guaranteeing the system's recognition performance.

[0119] (III) Synchronous recovery of salinity

[0120] If the edge device loses its salt value, the backend verifies the device's identity using "device SN + backend pre-stored salt value hash" and then sends a salt value update command (the salt value is transmitted encrypted via TLS 1.3). After receiving the command, the edge device writes the salt value to a secure storage area. The recovery time is ≤10 seconds, which quickly restores the device to normal operation and improves the system's fault tolerance.

[0121] It should be noted that, in this invention, although the specification describes the embodiments, not every embodiment contains only one independent technical solution. This way of describing the specification is only for clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A human pose recognition method based on block hashing and Bloom filters, characterized in that, Includes the following steps: (1) Preprocess the human posture data in private locations, remove invalid data and standardize the data; (2) Quantize the human posture features of the preprocessed posture data to obtain multidimensional structured quantitative features; (3) The multidimensional quantitative features are divided into blocks based on the correlation of human joint functions, and a hash combination is generated by combining the salt value; (4) The hash combination is transmitted to the backend using a lightweight encrypted transmission method; (5) The backend builds a pose-specific Bloom filter to perform fast similarity retrieval on the received hash combinations to realize human pose recognition; (6) Dynamically manage the pose prototype library and the corresponding Bloom filter to adapt to pose updates and scene changes.

2. The human pose recognition method based on block hashing and Bloom filter according to claim 1, characterized in that, In step (1), the preprocessing includes directional acquisition optimization, redundancy removal, and normalization. The targeted acquisition optimization achieves accurate acquisition of posture data by specifying the core area of ​​human activity and combining multi-device collaborative acquisition; The normalization process establishes a coordinate system with a preset joint point as the origin and performs size scaling based on the maximum joint spacing.

3. The human pose recognition method based on block hashing and Bloom filter according to claim 1, characterized in that, In step (2), the multidimensional quantization feature is 40-dimensional, including four categories: binary attribute, proportional attribute, angular attribute and trend attribute; The 40-dimensional features are constructed based on the positional relationships, movement trends, and morphological features of 16 key human body points. If feature extraction fails in a certain dimension, the average feature value of historical frames is used to fill the gap.

4. The human pose recognition method based on block hashing and Bloom filter according to claim 1, characterized in that, In step (3), the block division adopts a sliding window method, which groups features that are functionally coordinated and spatially adjacent into the same window, and sets overlapping features in adjacent windows; The salt value is generated by the edge device through a hardware random number generator, stored in a secure storage area and updated periodically. During the update, only the salt value hash is synchronized, not the original salt value.

5. The human pose recognition method based on block hashing and Bloom filter according to claim 4, characterized in that, The hash combination is generated as follows: After the quantization features of each window are superimposed with salt values, a hash code is generated by a hash algorithm, and the hash codes of multiple windows are combined to form the hash combination. The hash result for each window also has a checksum added for anomaly detection.

6. The human pose recognition method based on block hashing and Bloom filter according to claim 1, characterized in that, In step (4), the lightweight encrypted transmission uses the MQTT-SN protocol combined with the TLS 1.3 encryption layer. The transmitted data includes hash combination, device identifier, timestamp and check code, and is encapsulated in TLV format.

7. The human pose recognition method based on block hashing and Bloom filter according to claim 1, characterized in that, In step (5), the construction of the Bloom filter includes: The filter bit length and hash function number are calculated based on the attitude core hash combination. The hash combination is mapped to the corresponding index bit of the filter and set to 1 through multiple independent hash functions. The similarity retrieval adopts a multi-level threshold determination mechanism, calculates a similarity score based on the proportion of the matching window, and determines the posture category based on the score result.

8. The human pose recognition method based on block hashing and Bloom filter according to claim 1, characterized in that, In step (6), the dynamic management includes incremental learning of the prototype library, Bloom filter elimination and update, and salt value synchronization and recovery; The prototype library incremental learning obtains core pose features through clustering deduplication. The filter replacement and update are based on the call frequency and recognition accuracy. The salt value is synchronously restored by transmitting a new salt value encrypted after device authentication.

9. A human pose recognition system based on block hashing and Bloom filters, characterized in that, include: The data acquisition module is used to collect human posture data in a targeted manner in private locations; The preprocessing module is used to perform redundancy removal and normalization preprocessing on the acquired attitude data; The feature quantization module is used to quantize the preprocessed pose data and generate multidimensional structured quantized features. The block-salted hash module is used to divide quantized features into blocks based on human joint function and generate hash combinations by combining salt values. A lightweight transmission module is used to transmit hash combinations to the backend via an encrypted transmission protocol; The backend retrieval module is used to build a Bloom filter to perform fast similarity retrieval on the received hash combinations and realize pose recognition. The dynamic management module is used for incremental updates, elimination, and salt value restoration management of the attitude prototype library and Bloom filters.

10. The human pose recognition system based on block hashing and Bloom filters according to claim 9, characterized in that, The data acquisition module includes an infrared camera and a millimeter-wave radar. The infrared camera supports environmental adaptive adjustment, and the millimeter-wave radar is used for blind spot acquisition in occluded scenes. The backend retrieval module has a retrieval time of ≤15ms, a total system latency of ≤80ms, and a false positive rate of ≤0.3%.

Images