A fall detection classification method based on HMM and SVM cascade voter

Abstract

The application discloses a kind of based on HMM and SVM cascade voter's fall detection classification method, it is related to coal mine safety technical field, including: obtaining IMU sensor acquisition IMU data in human activity process;Using fixed length sliding window to read IMU data and parse, conversion is float number two-dimensional array, quaternion fusion is carried out to float number two-dimensional array, to update float number two-dimensional array, and data filtering is carried out to updated float number two-dimensional array;Based on SMV and HMM cascade voter output fall probability, and with the set proportion weight of SVM and HMM determines final output, if final output is greater than or equal to set value 0.5, then voter output is fall, if final output is less than set value 0.5, then cycle enters step one, until voter output is fall;The fall detection classification method improves fall detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety technology, and in particular to a fall detection and classification method based on a cascaded HMM and SVM voting system. Background Technology

[0002] I. Safety Issues in the Coal Mining Industry

[0003] Coal mining is a high-risk industry. Workers in the underground environment face a variety of potential safety hazards, including gas explosions, mine collapses, fires, mechanical injuries, and falls. Falls are not only a major cause of injury and death among coal miners, but they can also trigger secondary accidents, leading to even more serious consequences. Therefore, timely and accurate detection and identification of falls are crucial for ensuring the safety of coal miners.

[0004] II. Limitations of Existing Fall Detection Technology

[0005] Currently, fall detection technology is mainly divided into wearable device-based detection and video surveillance-based detection. Wearable device detection monitors workers' movement in real time using devices such as accelerometers and gyroscopes, but it faces problems such as sensor malfunction, improper device wearing, and environmental interference, leading to a decrease in detection accuracy. While video surveillance-based detection methods do not rely on workers wearing devices, in complex environments like coal mines, factors such as changes in lighting, dust, and obstructions significantly affect the quality of video images, thus impacting the effectiveness of fall detection. Summary of the Invention

[0006] Based on the technical problems existing in the background technology, this invention proposes a fall detection classification method based on HMM and SVM cascaded voting machines, which improves the accuracy of fall detection.

[0007] This invention proposes a fall detection classification method based on a cascaded HMM and SVM voting system, comprising the following steps:

[0008] Step 1: Acquire IMU data during human activity collected by the IMU sensor. The IMU data includes triaxial acceleration and triaxial angular velocity.

[0009] Step 2, Feature Processing Module: Read and parse IMU data using a fixed-length sliding window, convert it into a floating-point two-dimensional array, perform quaternion fusion on the floating-point two-dimensional array to generate three-dimensional attitude angles based on roll, yaw, and pitch, perform first-order difference on the three-dimensional attitude angles, vectorize the obtained first-order difference data with the three-dimensional attitude angle features to update the floating-point two-dimensional array, and perform data filtering on the updated floating-point two-dimensional array;

[0010] Step 3, the first stage of the attitude detection module: The SMV calculates the synthetic motion vector value based on the filtered floating-point two-dimensional array. If the synthetic motion vector value is less than the set SMV violent motion threshold, the collected IMU data is buffered and collection continues. If the synthetic motion vector value is greater than or equal to the set SMV violent motion threshold, the SMV outputs the binary classification of violent motion and the timestamp. When the binary classification of violent motion is determined to be true, the second stage of the attitude detection module is entered. The SMV is a support vector machine.

[0011] Step 4, the second stage of the attitude detection module: The filtered floating-point two-dimensional array is used to determine the explicit and implicit states through the emission matrix and transition matrix of the HMM. When the explicit and implicit state is the fall state, the HMM outputs the second fall probability. At the same time, the filtered floating-point two-dimensional array is calculated by linear SVM. The SVM outputs the first fall probability. The HMM is a Hidden Markov Model.

[0012] Step 5: Determine the final output based on the set weight ratio of SVM and HMM. If the final output is greater than or equal to the set value of 0.5, the voting device outputs "fall down". If the final output is less than the set value of 0.5, the process loops back to Step 1 until the voting device outputs "fall down".

[0013] Furthermore, in step three, the SMV calculates the synthetic motion vector value based on the filtered floating-point two-dimensional array. Specifically, it uses the most recent data in the filtered floating-point two-dimensional array to perform support vector machine calculations on the three-axis acceleration to obtain the synthetic motion vector value.

[0014] Furthermore, in step four, the Hidden Markov Model is constructed as follows:

[0015] The hidden state of the HMM is set to fall and non-fall, which are two features; the explicit state of the HMM is set to six computable features, namely triaxial acceleration and triaxial angular velocity.

[0016] The parameters of the HMM consist of three parts: the initialization matrix B, the emission matrix S (2*6), and the transition matrix Q (6*6). The three parts B, S, and Q are obtained by training and converging the Hidden Markov Model on the host computer.

[0017] The Viterbi path is calculated based on the B, S, and Q matrices, specifically as follows:

[0018] Initialization: Initialize the probability of a given first observation being in each state, where the observation is a value in a filtered two-dimensional array of floating-point numbers from the HMM input;

[0019] Recursion: For each subsequent observation, calculate the probability that each state becomes the next state in the hidden state sequence;

[0020] Termination: After processing all observations, backtrack to find the hidden state sequence with the highest probability. The highest hidden state sequence is taken as the probability of the second fall, and the second fall probability is judged as a fall state if it exceeds 0.5.

[0021] Furthermore, in step five, where the final output is determined based on the set weight ratio of SVM and HMM, the final output is specifically as follows:

[0022] The first fall probability is obtained by performing linear SVM on the filtered two-dimensional array of floating-point numbers.

[0023] The HMM is assigned a weight of 0.8, and the SVM is assigned a weight of 0.2. The final output is the sum of the weights of the second fall probability output by the HMM and the first fall probability output by the SVM.

[0024] Furthermore, in step two, the two-dimensional array of floating-point numbers is specifically generated as follows:

[0025] IMU data is read from the IMU's FIFO using a fixed-length FIFO interrupt method, where the FIFO interrupt threshold is set to 20 data packets, with a cycle of 200ms at 100Hz.

[0026] Using a sliding window approach, when the set 1-second window period is met, all data in the window is parsed and converted into a floating-point two-dimensional array Sliding_data_fifo

[100] [6], which is a 100*6 matrix data.

[0027] Furthermore, in step two, the updated floating-point two-dimensional array is subjected to a first-order Butterworth filter to obtain a filtered floating-point two-dimensional array.

[0028] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the fall detection and classification method as described above.

[0029] A computer-readable storage medium storing a plurality of classification programs, the plurality of classification programs being invoked by a processor and executed as described above for the fall detection classification method.

[0030] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0031] The advantages of the fall detection and classification method based on a cascaded HMM and SVM voting mechanism provided by this invention are as follows: By capturing the time-series features of worker actions through a Hidden Markov Model, it can more accurately distinguish falls from other similar actions (such as squatting or bending over). Combining the classification capabilities of Support Vector Machines with the comprehensive judgment of the voting mechanism can further improve the accuracy of fall detection and reduce false positives and false negatives. Moreover, this embodiment comprehensively utilizes the advantages of both HMM and SVM models, maintaining high detection performance in the complex and variable coal mine environment. The voting mechanism effectively reduces misjudgments by a single model in special circumstances, improving the overall robustness and stability of the system. Furthermore, timely and accurate fall detection can trigger early warning and rescue measures at the first moment, reducing injuries to workers caused by falls. This helps coal mining enterprises establish more intelligent safety monitoring systems, improve overall safety management levels, and protect the lives of workers. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the structure of the present invention;

[0033] Figure 2 A schematic diagram illustrating data transmission between the terminal equipment and the host computer system of the well monitoring platform;

[0034] Figure 3 Define the representation intent for the HMM's visible and hidden states; 3D attitude angles for roll, yaw, and pitch.

[0035] Figures 4a to 7h This serves as a reference image in feature visualization processing.

[0036] Figures 4a to 4c The following is a trend graph showing the changes in attitude angles (Roll, Pitch, Yaw) over time in the first fall paradigm: Figure 4a This is a graph showing the trend of the attitude angle Roll (pitch angle) over time in the first fall paradigm. Figure 4b This is a graph showing the trend of pitch over time in the first fall paradigm. Figure 4c This is a trend curve of the yaw (deflection angle) changing over time in the first fall paradigm;

[0037] Figures 5a to 5c The following is a graph showing the trend of the attitude angles Roll, Pitch, and Yaw over time in the second fall paradigm: Figure 5a This is a graph showing the trend of the attitude angle Roll (pitch angle) over time in the second fall paradigm. Figure 5bThis is a graph showing the trend of pitch over time in the second fall paradigm. Figure 5c This is a trend curve of the yaw (deflection angle) changing over time in the second fall paradigm;

[0038] Figures 6a to 6c The following is a trend graph showing the changes in attitude angles (Roll, Pitch, Yaw) over time in the third fall paradigm: Figure 6a This is a graph showing the trend of the attitude angle Roll (pitch angle) over time in the third fall paradigm. Figure 6b This is a trend graph of pitch over time in the third fall paradigm. Figure 6c This is a trend curve of the yaw (deflection angle) changing over time in the third fall paradigm.

[0039] Figures 7a to 7h The following is a trend curve showing how different data characteristics change over time for a single set of data: Figure 7a This is a trend graph of label numbers changing over time, but it has no real data significance. Figure 7b This is a trend curve of SMV changing over time, i.e., the SMV threshold for intense movement. Figure 7c This is a trend curve of the first-order Roll difference value changing over time in the first-order difference data; Figure 7d This is a trend curve of the second-order Roll difference value changing over time. Figure 7e This is a trend curve of the first-order pitch difference value changing over time in the first-order difference data. Figure 7f This is a trend curve of the second-order pitch difference value changing over time. Figure 7g This is a trend curve of the first-order Yaw difference value changing over time in the first-order difference data. Figure 7h This is a trend curve of the second-order Yaw difference value changing over time. Detailed Implementation

[0040] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0041] In mining operations, to detect falls, miners wear terminal devices on their waists. These devices include data acquisition, feature processing, posture detection, and communication hardware. The terminal devices connect to the host computer system of the surface monitoring platform via a base station (communication equipment). Therefore, the mine terminal devices adopt a host computer-base station (other communication intermediary layer)-terminal device approach. When the terminal device detects a fall, it effectively transmits the data to the surface via the data network, thus achieving an alarm effect. Local audible and visual alarms can alert local personnel working underground to provide immediate assistance and upload data. This embodiment uses a fall detection method based on a cascaded voting mechanism of HMM (Hidden Markov Model) and SVM (Support Vector Machine). HMM is a statistical model used to describe time-series data, effectively capturing the time-series features of worker movements. Support Vector Machine is a machine learning algorithm used for classification, possessing good generalization ability. By cascading HMM and SVM and combining them with a voting mechanism, the advantages of both can be combined to improve the accuracy and robustness of fall detection. The specific implementation of each module in the terminal device is as follows.

[0042] like Figures 1 to 7h As shown, the fall detection classification method based on HMM and SVM cascaded voting machines proposed in this invention includes the following steps:

[0043] Step 1, Acquisition Module: Acquire IMU data during human activity by using an inertial measurement unit (IMU) sensor. The IMU data includes triaxial acceleration and triaxial angular velocity.

[0044] Personnel perform normal daily activities and simulate abnormal movements such as falls. The data acquisition module of the worn terminal device collects the personnel's six-axis attitude data (IMU data) at a frequency of 100 Hz, and the collected data is stored in a buffer. Initially, the sampling frequency is set to 100 Hz, the accelerometer unit is g with a range of + / - 16 g, the angular velocity unit is ° / s with a range of + / - 2000 ° / s, and the IMU is configured in high-performance mode.

[0045] Step 2, Feature Processing Module: Read and parse IMU data using a fixed-length sliding window, convert it into a floating-point two-dimensional array, perform quaternion fusion on the floating-point two-dimensional array to generate three-dimensional attitude angles based on roll, yaw, and pitch, perform first-order difference on the three-dimensional attitude angles, vectorize the obtained first-order difference data with the three-dimensional attitude angle features to update the floating-point two-dimensional array, and perform data filtering on the updated floating-point two-dimensional array;

[0046] The feature processing module is triggered by an IMU interrupt and reads IMU data from the IMU's buffer. The feature processing module consists of three stages:

[0047] The first stage uses a fixed-length FIFO interrupt to read IMU data from the IMU's FIFO. The FIFO interrupt threshold is set to 20 packets of data, with a cycle of 200ms calculated at 100Hz. A sliding window method is used. When the set 1s window period is met, all data in the window is parsed and converted into a floating-point two-dimensional array Sliding_data_fifo

[100] [6], which is a 100*6 matrix data.

[0048] In the second stage, quaternion fusion is performed on Sliding_data_fifo

[100] [6]. Quaternion fusion can be directly implemented using existing technologies to generate three-dimensional attitude angles based on roll, yaw, and pitch. First-order difference is performed on the three-dimensional attitude angles to obtain pitch_diff1, yaw_diff1, and roll_diff1. The obtained difference data and attitude angle features are vectorized and Sliding_data_fifo

[100] [6] is updated so that its six-dimensional features are attitude angles and corresponding first-order difference values.

[0049] The third stage involves filtering the updated Sliding_data_fifo

[100] [6] data using a first-order Butterworth filter. First-order filtering is highly efficient and can address sensor data noise issues. Accelerometers and angular velocity sensors are susceptible to various noise interferences during measurement, such as electromagnetic noise and mechanical vibration noise. The Butterworth filter has a very steep stopband attenuation characteristic, which can effectively filter out noise above the required frequency range and improve the signal-to-noise ratio of the data.

[0050] This embodiment creatively uses first-order difference features of attitude angles as input to the Hidden Markov Model (HMM), thereby improving the accuracy of the model in terms of state representation, observation time series, and definition of state transition probabilities. First-order difference features of attitude angles can more sensitively capture the dynamic changes of the system, improving the HMM's ability to describe complex dynamic systems. Furthermore, these features are robust to small-amplitude noise because they focus on the trend of angle changes rather than instantaneous values.

[0051] Step 3, the first stage of the attitude detection module: The SMV calculates the synthetic motion vector value based on the filtered floating-point two-dimensional array. If the synthetic motion vector value is less than the set SMV violent motion threshold, the already collected IMU data is buffered and collection continues. If the synthetic motion vector value is greater than or equal to the set SMV violent motion threshold, the SMV outputs the binary classification of violent motion and the timestamp. When the binary classification of violent motion is determined to be true, the second stage of the attitude detection module is entered. The SMV is a support vector machine.

[0052] In order to reduce power consumption in the terminal device, this embodiment designs a cascaded algorithm. The first stage is the SMV threshold determination. When the conditions of the first stage are met, the second stage voter model is determined. This can effectively avoid unnecessary computational resource consumption in the terminal device. The design of the algorithm effectively reduces the power consumption of the device.

[0053] The attitude detection module receives data (i.e., a filtered two-dimensional array of floating-point numbers) from the feature processing module. This data is first input into the first stage of the algorithm cascade to calculate the SMV (Synthetic Motion Vector) value. If the SMV value reaches the SMV violent motion threshold Threshold_smv, the second stage of the attitude detection module is entered. If it does not reach the threshold, the intermediate data is cached and acquisition continues. The goal of the first stage of the attitude detection module is to infer from prior knowledge that when no large SMV value is found, it is basically determined that no large movement has occurred, and simultaneously it is determined that no fall has occurred.

[0054] In the first stage of the attitude detection module, the input is the most recent data in the filtered Sliding_data_fifo

[100] [6]. The triaxial acceleration of this data is taken and SMV is calculated to obtain the synthetic motion vector value SMVreal. The threshold Threshold_smv is used for classification. If the standard is not met, the remaining steps are skipped and the data acquisition continues. The output of SMV in the first stage is: 1. Flag_smv_beyond, for binary classification of violent motion, 2. Current timestamp, corresponding to the Index, used to mark the index of the current fifo, that is, the index of the current SMV maximum value in the fifo Smv_index.

[0055] Step 4, the second stage of the attitude detection module: The filtered floating-point two-dimensional array is used to determine the explicit and implicit states through the emission matrix and transition matrix of the HMM. When the explicit and implicit state is the fall state, the HMM outputs the second fall probability. At the same time, the filtered floating-point two-dimensional array is calculated by linear SVM. The SVM outputs the first fall probability. The HMM is a Hidden Markov Model.

[0056] When Flag_smv_beyond is determined to be true in step three, the cascade algorithm is started. When the HMM model determines that it is a fall (Fall) based on the explicit and implicit states, a linear SVM with a weighting of 0.8 and a weighting of 0.2 is used to determine that if the value of Fall exceeds 0.5, it is determined to be a fall; if it is below 0.5, it is determined to be a non-fall.

[0057] The corresponding HMM models are constructed as (a1) to (a3):

[0058] (a1) The hidden states of the HMM are set to falling and non-falling, which are two features; the explicit states of the HMM are set to six computable features, namely triaxial acceleration and triaxial angular velocity, such as... Figure 3 As shown;

[0059] (a2) The parameters of the HMM are three parts: initialization matrix B, emission matrix S (2*6), and transition matrix Q (6*6). The three parts B, S, and Q are obtained by training and converging the hidden Markov model of the host computer.

[0060] (a3) The Viterbi algorithm is used, which is a dynamic programming algorithm for finding the most likely sequence of hidden states (called the Viterbi path) in a Hidden Markov Model (HMM) given a set of observed events. The algorithm operates on the assumption that future states depend only on the current state (a Markov property).

[0061] The Viterbi path is calculated based on the B, S, and Q matrices, specifically from (a3-1) to (a3-3):

[0062] (a3-1) Initialization: Initialize the probability of a given first observation being in each state, where the observation is a value in the filtered floating-point two-dimensional array of HMM inputs;

[0063] (a3-2) Recursion: For each subsequent observation, calculate the probability that each state becomes the next state in the hidden state sequence. The calculation process uses the HMM model's own calculation method, but the matrices and features involved are specifically set in this embodiment. The matrices are: initialization matrix B, emission matrix S (2*6), and transition matrix Q (6*6). The features are: the explicit states of the HMM are set to six computable features, namely triaxial acceleration and triaxial angular velocity. This step involves checking the probability of transitioning from each previous state to the current state and selecting the highest probability.

[0064] (a3-3) Termination: After processing all observations, backtrack to find the hidden state sequence with the highest probability. The highest hidden state sequence is taken as the probability of the second fall, and the probability of the second fall is determined to be a fall state if it exceeds 0.5.

[0065] Through (a3-1) to (a3-3), the recursive process of this embodiment effectively decomposes the problem into simpler, overlapping subproblems and stores their solutions to avoid redundant computation. In (a1) to (a3), the iteration of the dimensional bit algorithm based on actual IMU data is used, where the data type is still consistent with the floating-point two-dimensional array Sliding_data_fifo

[100] [6].

[0066] After the HMM outputs the probability of the second fall, a linear SVM is performed on the filtered floating-point two-dimensional array to obtain the probability of the first fall. The linear one-dimensional operator T in the SVM also uses the HMM algorithm, but the difference between this HMM algorithm and the one mentioned above lies in the explicit state parameters used. The explicit state parameters used here are: Pitch_diff1_cnt, corresponding to the number of bounces under Std_time_windows; Pitch_std, corresponding to the standard deviation of pitch Ptich under Std_time_windows; Roll_diff1_std, corresponding to the standard deviation of the first difference of the roll parameter under Std_time_windows; and Std_time_windows, the time window used to calculate the standard deviation. Currently, considering the duration after a fall, it is set at 1.5 seconds, and at 100 Hz, 150 points are used, i.e., 150 time steps. (This parameter can be used as an extended window to adjust sensitivity).

[0067] Step 5: Determine the final output based on the set weight ratio of SVM and HMM. If the final output is greater than or equal to the set value of 0.5, the voting device output is "falling down". If the final output is less than the set value of 0.5, the voting device output is "not falling down".

[0068] The HMM is assigned a weight of 0.8, and the SVM is assigned a weight of 0.2. The final output is the sum of the weights of the second fall probability output by the HMM and the first fall probability output by the SVM. If the final output is greater than or equal to the set value of 0.5, the voting device outputs a fall, which is determined to be an abnormal posture. The abnormal posture terminal issues an audible and visual alarm and sends the detection result to the base station. The base station transmits the abnormal posture to the well monitoring platform, and the well staff will then proceed with the next step.

[0069] According to steps one through five, this embodiment captures the time-series features of worker movements using a Hidden Markov Model (HMM), enabling more accurate differentiation between falls and other similar movements (such as squatting or bending over). Combining the classification capabilities of Support Vector Machines (SVMs) with a voting mechanism further improves the accuracy of fall detection and reduces false positives and false negatives. Moreover, this embodiment leverages the advantages of both HMMs and SVMs, maintaining high detection performance in the complex and dynamic coal mine environment. The voting mechanism effectively reduces misjudgments by a single model in special circumstances, improving the overall robustness and stability of the system. Furthermore, timely and accurate fall detection can trigger early warnings and rescue measures immediately, reducing injuries to workers from falls. This helps coal mining enterprises establish more intelligent safety monitoring systems, improve overall safety management, and protect worker safety.

[0070] As an example:

[0071] Step 1: Personnel should correctly wear the terminal device around their waist.

[0072] Step 2: Personnel perform normal daily activities and simulate abnormal movements such as falls. The terminal device's data acquisition module collects the personnel's six-axis posture data at a frequency of 100Hz, and the collected data is stored in a cache.

[0073] Step 3: The feature processing module is triggered by an IMU interrupt and reads IMU data from the cache. The IMU data is converted into attitude angles and first-order difference data, totaling six-dimensional features. After passing through a Badiworth low-order filter, it serves as the input for the next step.

[0074] Step 4: First, determine whether the filtered floating-point two-dimensional array meets the SMV (Severe Motion) threshold. If it does, proceed to the next stage. At this point, the binary classification of severe motion, Flag_smv_beyond, is True. Calculate the Viterbi path using the given B, S, and Q matrices, and recursively obtain the hidden state under the HMM algorithm, i.e., the probability of the second fall. If the probability exceeds 0.5, it is determined as a fall. The training samples used for the B, S, and Q matrices are the motion recordings shown in Table 1.

[0075] Next, linear SVM is performed on the feature vectors in the filtered floating-point two-dimensional array. The linear one-dimensional operator T in the SVM is also the fitted SVM model obtained from Table 1. The fall probability can also be obtained. The HMM algorithm is assigned a weight of 0.8 and the SVM is assigned a weight of 0.2 to obtain the final output. If the final output is greater than 0.5, a fall is determined; if it is less than 0.5, no processing is done.

[0076] Step 5: In the training samples above, the dataset used for training the posture detection module was autonomously collected. The action paradigms of the collected data are shown in Table 1. Developers wore a device around their waist to simulate corresponding actions according to the action paradigms. There were 20 action paradigms in total, divided into daily activity action paradigms and fall action paradigms. A total of 15 people with different body types were collected as samples, and each action paradigm was simulated 30 times, for a total of 9000 sets of data.

[0077] Table 1 Human Posture and Movement Paradigms

[0078]

[0079]

[0080] Step 6: When an abnormal attitude is detected, the terminal will send the abnormal status to the base station (or other communication intermediate device) through the communication module, and the base station will transmit it to the well monitoring platform.

[0081] This embodiment utilizes the sensitivity of Hidden Markov Models to binary states and a segmented cascade approach to reduce power consumption. It also employs a voting mechanism based on the SVM linear machine learning algorithm to perform fall detection on wearable devices in mines. By integrating the advantages of different algorithms, it improves detection accuracy and system robustness. Furthermore, this embodiment exhibits strong data filtering capabilities, adaptability, and robustness, making it particularly suitable for complex environments such as underground coal mines. It can operate stably under harsh conditions, providing reliable protection for miners' safety.

[0082] like Figures 4a to 6c The trends of attitude angles Roll, Pitch, and Yaw over time were collected in three fall paradigms. Figures 4a to 4c As a group, Figures 5a to 5c As a group, Figures 6a to 6c (as a group), Figures 4a to 6c The horizontal axis represents the time step, with a single time step of 1 / 100 s. The vertical axis indicates different units depending on the angle. Figures 4a to 6c From these three groups, we can see that: 1. The Roll angle drops sharply at the moment of falling; 2. The Pitch fluctuates periodically before falling, drops sharply at the moment of falling, and then changes relatively steadily; 3. The Yaw changes similarly to the Pitch.

[0083] like Figures 7a to 7h For different data characteristics under a single set of data, the horizontal axis also represents the time step, with the unit being 1 / 100s, and the vertical axis represents the corresponding scale unit.

[0084] from Figures 4a to 7h The following conclusions can be drawn: By fusing attitude angles, more obvious fall characteristics can be obtained under visualization. Different differencing methods can improve data stability, filter time trend terms, and better extract features.

[0085] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A fall detection classification method based on a cascaded HMM and SVM voting system, comprising the following steps: Step 1: Acquire IMU data during human activity collected by the IMU sensor. The IMU data includes triaxial acceleration and triaxial angular velocity. Step 2, Feature Processing Module: Read and parse IMU data using a fixed-length sliding window, convert it into a floating-point two-dimensional array, perform quaternion fusion on the floating-point two-dimensional array to generate three-dimensional attitude angles based on roll, yaw, and pitch, perform first-order difference on the three-dimensional attitude angles, vectorize the obtained first-order difference data with the three-dimensional attitude angle features to update the floating-point two-dimensional array, and perform data filtering on the updated floating-point two-dimensional array; Step 3, the first stage of the attitude detection module: The SMV calculates the synthetic motion vector value based on the filtered floating-point two-dimensional array. If the synthetic motion vector value is less than the set SMV violent motion threshold, the collected IMU data is buffered and collection continues. If the synthetic motion vector value is greater than or equal to the set SMV violent motion threshold, the SMV outputs the binary classification of violent motion and the timestamp. When the binary classification of violent motion is determined to be true, the second stage of the attitude detection module is entered. The SMV is a support vector machine. Step 4, the second stage of the attitude detection module: The filtered floating-point two-dimensional array is used to determine the explicit and implicit states through the emission matrix and transition matrix of the HMM. When the explicit and implicit state is the fall state, the HMM outputs the second fall probability. At the same time, the filtered floating-point two-dimensional array is calculated by linear SVM. The SVM outputs the first fall probability. The HMM is a Hidden Markov Model. Step 5: Determine the final output based on the set weight ratio of SVM and HMM. If the final output is greater than or equal to the set value of 0.5, the voting device outputs "fall down". If the final output is less than the set value of 0.5, the process loops back to Step 1 until the voting device outputs "fall down".

2. The fall detection classification method based on HMM and SVM cascaded voting machines according to claim 1, characterized in that, In step three, SMV calculates the synthetic motion vector value based on the filtered floating-point two-dimensional array. Specifically, it uses the most recent data in the filtered floating-point two-dimensional array to perform support vector machine calculations on the three-axis acceleration to obtain the synthetic motion vector value.

3. The fall detection classification method based on HMM and SVM cascaded voting machines according to claim 1, characterized in that, In step four, the Hidden Markov Model is constructed as follows: The hidden states of an HMM are set to "falling" and "not falling," which are two features. The explicit state settings of the HMM are set to six computable features: triaxial acceleration and triaxial angular velocity. The parameters of the HMM consist of three parts: the initialization matrix B, the emission matrix S (2*6), and the transition matrix Q (6*6). The three parts B, S, and Q are obtained by training and converging the Hidden Markov Model on the host computer. The Viterbi path is calculated based on the B, S, and Q matrices, specifically as follows: Initialization: Initialize the probability of a given first observation being in each state, where the observation is a value in a filtered two-dimensional array of floating-point numbers from the HMM input; Recursion: For each subsequent observation, calculate the probability that each state becomes the next state in the hidden state sequence; Termination: After processing all observations, backtrack to find the hidden state sequence with the highest probability. The highest hidden state sequence is taken as the probability of the second fall, and the second fall probability is judged as a fall state if it exceeds 0.

5.

4. The fall detection classification method based on HMM and SVM cascaded voters according to claim 3, characterized in that, In step five, where the final output is determined by the set weight ratio of SVM and HMM, the final output is as follows: The first fall probability is obtained by performing linear SVM on the filtered two-dimensional array of floating-point numbers. The HMM is assigned a weight of 0.8, and the SVM is assigned a weight of 0.

2. The final output is the sum of the weights of the second fall probability output by the HMM and the first fall probability output by the SVM.

5. The fall detection classification method based on HMM and SVM cascaded voting machines according to claim 1, characterized in that, In step two, the two-dimensional array of floating-point numbers is generated as follows: IMU data is read from the IMU's FIFO using a fixed-length FIFO interrupt method, where the FIFO interrupt threshold is set to 20 data packets, with a cycle of 200ms at 100Hz. Using a sliding window approach, when the set 1-second window period is met, all data in the window is parsed and converted into a floating-point two-dimensional array Sliding_data_fifo[100][6], which is a 100*6 matrix data.

6. The fall detection classification method based on HMM and SVM cascaded voting machines according to claim 1, characterized in that, In step two, the updated floating-point two-dimensional array is subjected to a first-order Butterworth filter to obtain the filtered floating-point two-dimensional array.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the fall detection and classification method as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of classification programs, which are used by a processor to call and execute the fall detection classification method as described in any one of claims 1-6.

Images