Personnel tumble detection system based on millimeter wave radar

Abstract

The invention discloses a millimeter-wave radar-based personnel falling detection system, which is characterized in that in order to solve the problem that the noise of original data acquired by a millimeter-wave radar is relatively large, denoising processing is performed on radar echo signals by adopting wavelet transform, and noise interference is effectively suppressed and the signal quality is improved by setting a proper wavelet basis and a proper number of decomposition layers; the method comprises the following steps of: segmenting and aligning radar signals in a tumble process by adopting a dynamic time warping algorithm on the basis of extracting tumble features, and eliminating the influence of tumble speeds and durations of different people through nonlinear time normalization by the algorithm to obtain a normalized tumble feature sequence; a tumble attitude feature library is established, a support vector machine is adopted to carry out classification training on tumble features under different body types and attitudes, a self-adaptive tumble judgment model is obtained, and the generalization ability and the detection accuracy of an algorithm are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of human body detection, and in particular relates to a human fall detection system based on millimeter wave radar. Background Art

[0002] When using millimeter-wave radar to detect falls, there are some technical difficulties. First, the raw data obtained by the millimeter-wave radar has large noise, low signal quality, and is easily disturbed by the environment, resulting in a high false alarm rate for fall judgment. Secondly, different people have different body shapes and fall postures. A single judgment algorithm is difficult to cover all situations and is prone to missed detection. Furthermore, the fall occurs in a very short time, and how to extract effective fall features from radar data in a short time is also a major challenge. Finally, the actual application environment is complex and changeable, with interference factors such as occlusion and reflection, which places high demands on the robustness of the algorithm. In summary, how to quickly and accurately determine a person's fall event in noisy millimeter-wave radar data is a technical problem that needs to be solved urgently. Summary of the invention

[0003] In order to solve the above technical problems, the present invention proposes a personnel fall detection system based on millimeter wave radar to solve the problems existing in the above prior art.

[0004] To achieve the above object, the present invention provides a personnel fall detection system based on millimeter wave radar, comprising:

[0005] A signal preprocessing module is used to preprocess the original radar data to obtain preprocessed data;

[0006] A feature extraction module, used to extract the fall-sensitive features of the preprocessed data and perform dynamic time warping to obtain a fall feature sequence;

[0007] A classification training module, used for training a support vector machine according to the fall feature sequence to obtain an adaptive fall determination model;

[0008] A multi-view signal acquisition module, used to acquire multi-view radar echo signals through a multi-antenna array;

[0009] A deep learning module is used to train a convolutional neural network using multi-view radar echo signals to obtain a highly robust fall detection model;

[0010] The recognition module is used to detect the data to be detected by using the adaptive fall determination model and the high-robustness fall detection model, and to fuse the detection results of the adaptive fall determination model and the high-robustness fall detection model to obtain a final detection result.

[0011] Preferably, the signal preprocessing module comprises:

[0012] A wavelet transform unit, used to reconstruct the original radar data to obtain a denoised radar echo signal;

[0013] A detection unit is used to perform quality assessment on the denoised radar echo signal.

[0014] Preferably, the feature extraction module comprises:

[0015] A transform unit, used for obtaining a time-frequency spectrum of a signal by short-time Fourier transform;

[0016] An extraction unit extracts fall-sensitive features according to the energy distribution characteristics of the fall action in the time-frequency domain, wherein the fall-sensitive features include energy mutation and spectrum expansion;

[0017] The warping unit is used to segment and align the radar signal of the falling process using a dynamic time warping algorithm.

[0018] Preferably, the regularization unit comprises:

[0019] A mapping unit, used to map the falling processes of different durations to a unified time scale through a nonlinear time normalization method;

[0020] a difference elimination unit for resampling and interpolating the fall feature parameters according to the time scale;

[0021] The serialization processing unit is used to perform serialization processing on the resampled and interpolated fall feature parameters to obtain a normalized fall feature sequence.

[0022] Preferably, the classification training module includes:

[0023] A first construction unit is used to construct an initial adaptive fall determination model according to a support vector machine;

[0024] A first training unit, configured to train the initial adaptive fall determination model through a fall feature sequence;

[0025] The first verification unit is used to evaluate the trained initial adaptive fall determination model and optimize parameters according to the evaluation result to obtain the adaptive fall determination model.

[0026] Preferably, the training unit comprises a classification unit, which uses a kernel function to map feature samples in the fall feature sequence into a high-dimensional space, and classifies the feature samples by finding a maximum margin hyperplane.

[0027] Preferably, the multi-view signal acquisition module includes:

[0028] A signal acquisition unit, used for acquiring multi-view radar echo signals through a multi-antenna array;

[0029] The signal processing unit is used to process the multi-view radar echo signal by a spatial spectrum estimation method to obtain a denoised signal.

[0030] Preferably, the deep learning module includes:

[0031] A second construction unit is used to construct an initial high-robustness fall determination model according to a convolutional neural network;

[0032] A second training unit, configured to train the initial high-robustness fall determination model through a denoised signal;

[0033] The second verification unit is used to evaluate the trained initial high-robustness fall determination model through a cross-validation method, and optimize parameters according to the evaluation results to obtain an adaptive fall determination model.

[0034] Compared with the prior art, the present invention has the following advantages and technical effects:

[0035] The invention discloses a fall detection method based on millimeter wave radar. In view of the problem of large noise in radar raw data, wavelet transform is used for denoising. Time-frequency analysis is performed on the denoised signal to extract fall-sensitive features such as energy mutation and spectrum expansion. A dynamic time warping algorithm is used to segment and align the fall process to eliminate the differences in fall speed and duration of different people. A fall posture feature library is established, and a support vector machine is used for classification training to obtain an adaptive fall judgment model. In order to overcome the limitations of a single perspective, a multi-antenna array design is used to obtain multi-perspective radar signals, and a spatial spectrum estimation method is used to suppress environmental clutter. On this basis, multi-domain fall features are fused, and a convolutional neural network is used to determine fall events. An attention mechanism is introduced to adaptively weight sensitive features to highlight key information. The invention significantly improves the accuracy and robustness of fall detection through multi-perspective signal fusion and deep learning methods, and provides reliable technical support for fall monitoring for the elderly. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0037] Figure 1 4 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0039] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0040] Embodiment 1

[0041] like Figure 1 As shown, in this embodiment, a personnel fall detection system based on millimeter wave radar is provided, including:

[0042] A signal preprocessing module is used to preprocess the original radar data to obtain preprocessed data;

[0043] A feature extraction module is used to extract the fall-sensitive features of the preprocessed data and perform dynamic time warping to obtain a fall feature sequence;

[0044] A classification training module is used to train a support vector machine according to a fall feature sequence to obtain an adaptive fall determination model;

[0045] A multi-view signal acquisition module, used to acquire multi-view radar echo signals through a multi-antenna array;

[0046] A deep learning module is used to train a convolutional neural network using multi-view radar echo signals to obtain a highly robust fall detection model;

[0047] The recognition module is used to detect the data to be detected through the adaptive fall determination model and the highly robust fall detection model, and to fuse the detection results of the adaptive fall determination model and the highly robust fall detection model to obtain the final detection result.

[0048] Furthermore, the signal preprocessing module includes:

[0049] A wavelet transform unit is used to reconstruct the original radar data to obtain a denoised radar echo signal;

[0050] The detection unit is used to evaluate the quality of the radar echo signal after denoising.

[0051] Specifically, the original echo signal data of the millimeter-wave radar is obtained, and the wavelet transform algorithm is used to perform denoising to address the noise interference problem in the signal. According to the characteristics of the signal, a suitable wavelet basis function, such as Daubechies wavelet, Symlets wavelet, etc., is selected to match the time-frequency characteristics of the signal. The number of layers of wavelet decomposition is determined, and the signal is decomposed into sub-band signals of different frequency bands through multi-scale decomposition. The noise is mainly concentrated in the high-frequency sub-band. Threshold processing is performed on each sub-band signal, and a suitable threshold is set according to the statistical characteristics of the noise. The high-frequency coefficients less than the threshold are suppressed to remove the noise component. Through the wavelet reconstruction algorithm, the processed sub-band signal is inversely transformed and reconstructed to obtain the denoised radar echo signal. The quality of the denoised signal is evaluated, and indicators such as signal-to-noise ratio and mean square error are calculated to determine whether the denoising effect meets the requirements. If the denoising effect is not ideal, the wavelet basis function or the number of decomposition layers is adjusted, and the wavelet transform denoising is performed again until a high-quality radar echo signal is obtained.

[0052] Furthermore, the feature extraction module includes:

[0053] A transform unit, used for obtaining a time-frequency spectrum of a signal by short-time Fourier transform;

[0054] An extraction unit extracts fall-sensitive features according to the energy distribution characteristics of the fall action in the time-frequency domain, wherein the fall-sensitive features include energy mutation and spectrum expansion;

[0055] The warping unit is used to segment and align the radar signal of the falling process using a dynamic time warping algorithm.

[0056] Specifically, a short-time Fourier transform is performed on the signal to obtain the time-frequency spectrum matrix of the signal. The rows of the matrix represent time, the columns represent frequency, and the matrix elements represent the energy of the corresponding time-frequency points. According to the energy distribution characteristics of the falling action in the time-frequency domain, an energy mutation detection algorithm is designed. If the energy value of the current time-frequency point is greater than T times the energy average of the adjacent time-frequency points, it is judged as an energy mutation point. According to the spectrum expansion characteristics of the falling action in the time-frequency domain, a spectrum expansion detection algorithm is designed. If the frequency range of the current time-frequency point exceeds F1 Hz and the duration exceeds t1 seconds, it is judged as a spectrum expansion interval. The number of energy mutation points and the length of the spectrum expansion interval of each frame signal are counted and used as the fall-sensitive feature of the frame.

[0057] Furthermore, the regularization unit includes:

[0058] A mapping unit, used to map the falling processes of different durations to a unified time scale through a nonlinear time normalization method;

[0059] a difference elimination unit for resampling and interpolating the fall feature parameters according to the time scale;

[0060] The serialization processing unit is used to perform serialization processing on the resampled and interpolated fall feature parameters to obtain a normalized fall feature sequence.

[0061] Specifically, the dynamic time warping algorithm is used to segment and align the extracted fall feature parameters. In the dynamic time warping algorithm, the fall processes of different durations are mapped to a unified time scale through a nonlinear time normalization method. According to the normalized time scale, the fall feature parameters are resampled and interpolated to eliminate the influence of the differences in fall speed and duration of different people. The resampled and interpolated fall feature parameters are serialized to obtain a standardized fall feature sequence.

[0062] Furthermore, the classification training module includes:

[0063] A first construction unit is used to construct an initial adaptive fall determination model according to a support vector machine;

[0064] A first training unit is used to train an initial adaptive fall determination model through a fall feature sequence;

[0065] The first verification unit is used to evaluate the trained initial adaptive fall determination model and optimize parameters according to the evaluation result to obtain the adaptive fall determination model.

[0066] Specifically, the extracted feature vector is input into a pre-trained support vector machine classification model, and the classification model is used to determine whether the posture of the target person is in a falling state; if the classification model determines that the target person is in a falling state, a fall alarm is triggered, and if the judgment result is that the target person is not in a falling state, the target person's posture image data is returned to be obtained; in the training process of the support vector machine classification model, according to the fall feature sample data under different body shapes and postures, the kernel function is used to map the feature samples into a high-dimensional space, and the feature samples are classified by finding the maximum interval hyperplane to obtain a fall judgment classification model with adaptive ability; the cross-validation method is used to evaluate the generalization ability of the trained classification model, and the parameters of the classification model are tuned according to the evaluation results to improve the generalization ability of the model; test sample data is obtained to test the trained classification model, and the accuracy, recall rate and other indicators of the model are calculated, and the classification model is further optimized according to the indicator results, and finally a fall judgment model with strong generalization ability and high detection accuracy is obtained.

[0067] Furthermore, the training unit includes a classification unit, which uses a kernel function to map feature samples in the fall feature sequence to a high-dimensional space, and classifies the feature samples by finding a maximum margin hyperplane.

[0068] Furthermore, the multi-view signal acquisition module includes:

[0069] A signal acquisition unit, used for acquiring multi-view radar echo signals through a multi-antenna array;

[0070] The signal processing unit is used to process the multi-view radar echo signal by a spatial spectrum estimation method to obtain a denoised signal.

[0071] Specifically, a multi-antenna array design, such as a uniform linear array, with 8 receiving antenna units arranged at half-wavelength intervals, can obtain radar echo signals from different perspectives of the target. The echo signal received by each antenna unit contains information about the target at that perspective. Using this multi-perspective information, the scattering characteristics of the target can be more comprehensively characterized, thereby improving the accuracy of fall detection. Spatial spectrum estimation methods, such as the MUSIC algorithm, can effectively suppress environmental clutter. Assuming that environmental clutter mainly comes from ground reflection, its signal energy is mainly concentrated in the low-frequency region. Through the MUSIC algorithm, the spatial spectrum of the clutter signal can be estimated and removed from the received radar echo signal, thereby enhancing the radar scattering characteristics of the fallen target. For example, when a fall occurs, changes in the posture of the human body will cause the Doppler shift of the radar echo signal to change. By analyzing the changing law of the Doppler shift, the motion characteristic parameters of the fallen target, such as velocity and acceleration, can be extracted.

[0072] Furthermore, the deep learning module includes:

[0073] A second construction unit is used to construct an initial high-robustness fall determination model according to a convolutional neural network;

[0074] A second training unit is used to train an initial high-robustness fall determination model through a denoised signal;

[0075] The second verification unit is used to evaluate the trained initial high-robustness fall determination model through a cross-validation method, and optimize parameters according to the evaluation results to obtain an adaptive fall determination model.

[0076] Specifically, the time-frequency spectrum of each perspective is subjected to image preprocessing, including denoising, normalization and other operations, to obtain a normalized time-frequency spectrum; the feature sequence related to falls is extracted from the normalized time-frequency spectrum, and a multi-domain fall feature vector is constructed; a convolutional neural network model is constructed based on the feature vectors of multiple perspectives, and deep fall features are extracted through convolutional layers and pooling layers; the convolutional neural network is trained with a large amount of measured fall data, and the network parameters are optimized through the back propagation algorithm to improve the generalization ability of the model; during the training process, the model performance is evaluated through methods such as cross-validation, and the network structure and hyperparameters are dynamically adjusted according to the evaluation results to obtain a robust fall detection model.

[0077] In this embodiment, there are the following examples:

[0078] Radar signal acquisition is the first step in fall detection. Continuous wave radar sensors can be used to obtain reflected signals generated by human motion. For example, a radar sensor with an operating frequency of 24 GHz is used to collect data at the target area to obtain the original radar signal containing human motion information. These signals are usually complex signals containing amplitude, phase and frequency information. The next step is to preprocess the radar signal to remove noise and interference information and improve signal quality. Radar signals often contain interference information such as environmental noise and human breathing. In order to remove these noises, filtering methods, such as bandpass filters, can be used to filter out unnecessary frequency components in the signal. For example, setting the passband range to 0.5 Hz to 5 Hz can effectively retain the frequency components related to falls and remove high-frequency noise and low-frequency breathing interference. Wavelet transform and other methods can also be used for denoising to improve the signal-to-noise ratio and highlight the characteristics of fall signals. Extracting key information reflecting the characteristics of falls from the preprocessed signal is an important step. Feature parameters such as signal amplitude, frequency, and phase can be extracted. For example, when a fall occurs, the amplitude of the radar signal usually changes dramatically, and the amplitude change rate can be extracted as a feature parameter. In addition, the speed of human movement will change during the fall, which is reflected in the frequency of the radar signal, so Doppler shift is also an important characteristic parameter. Phase information can reflect the trajectory of human movement and can also be used as one of the characteristic parameters. The dynamic time warping algorithm (DTW) is used to segment and align the extracted fall characteristic parameters. Since the speed and duration of each person's fall are different, directly comparing the fall characteristic parameters of different people may lead to misjudgment. The DTW algorithm can solve this problem. It maps fall processes of different durations to a unified time scale through a nonlinear time normalization method. For example, aligning a fall process with a duration of 2 seconds and a fall process with a duration of 3 seconds can eliminate the impact of time differences and compare their features more accurately. According to the normalized time scale obtained by the DTW algorithm, the fall characteristic parameters are resampled and interpolated to eliminate the impact of the differences in the speed and duration of different people's falls. For example, after two fall signals of different lengths are aligned through the DTW algorithm, they are sampled into 100 points and can be directly compared. Interpolation can be used to fill in missing data to make the characteristic parameters more complete. The resampled and interpolated fall feature parameters are serialized to obtain a normalized fall feature sequence. For example, the three feature parameters of amplitude, frequency, and phase are arranged in chronological order to form a feature vector sequence. In this way, different types of feature parameters can be integrated together to facilitate subsequent machine learning algorithm processing. Finally, the normalized fall feature sequence is used as input, and a machine learning algorithm such as support vector machine (SVM) or neural network is used for fall detection and recognition. For example, SVM is used to classify the fall feature sequence to distinguish between fall and non-fall actions.Training the SVM model requires a large amount of sample data, including fall data and non-fall data. Through training, the SVM can learn the pattern of fall features, thereby realizing fall detection.

[0079] Acquire radar signal data from multiple perspectives. For example, three radar sensors can be arranged in different corners of the room to collect radar echo signals of human motion, so that radar signal data from three different perspectives can be obtained. These data contain the motion information of the human body at different perspectives, which can more comprehensively reflect the changes in the posture of the human body. Perform time-frequency analysis on the signal data from each perspective. For example, short-time Fourier transform (STFT) can be used to convert the radar signal from each perspective into a time-frequency spectrum. The time-frequency spectrum can reflect the changes of the frequency components of the signal over time. Assuming that the radar signal from a certain perspective first increases and then decreases during a fall, and the high-frequency energy first decreases and then increases, these change characteristics can be clearly reflected in the time-frequency spectrum. Perform image preprocessing on the time-frequency spectrum of each perspective, including denoising, normalization and other operations. For example, median filtering can be used to remove noise in the time-frequency spectrum, and then the pixel values ​​of the time-frequency spectrum can be normalized to between 0 and 1. This can eliminate the influence of the difference in radar signal strength from different perspectives, which is convenient for subsequent feature extraction and model training. Assuming that the radar signal of a perspective has a weak signal strength due to the long distance, after normalization, the features of its time-frequency spectrum can be compared and fused with the time-frequency spectrum of other perspectives. Extract the feature sequence related to the fall from the normalized time-frequency spectrum. For example, the change features of the energy distribution, the change features of the frequency center, and the change features of the spectral entropy in the time-frequency spectrum can be extracted, and these features can be combined into a multi-domain fall feature vector. Assuming that during the fall process, the energy of the time-frequency spectrum is mainly concentrated in the low-frequency area, the frequency center will move to the low frequency, and the spectral entropy will increase first and then decrease. These features can be used to characterize the fall action. According to the feature vectors of multiple perspectives, a convolutional neural network model is constructed. For example, a convolutional neural network model containing three convolutional layers and two pooling layers can be constructed. The feature vector of each perspective can be used as the input of the convolutional neural network. The convolutional layer and the pooling layer can extract and fuse the fall features of different perspectives. In this way, the information of multiple perspectives can be fully utilized to improve the accuracy of fall detection. A large amount of measured fall data is used to train the convolutional neural network. For example, 1,000 fall samples and 1,000 non-fall samples can be collected, and these samples can be divided into training sets and test sets. The training set is used to train the convolutional neural network, and the test set is used to evaluate the performance of the model. During the training process, the network parameters are optimized by the back propagation algorithm. For example, the stochastic gradient descent algorithm can be used to adjust the weights and biases of the convolution kernel to minimize the prediction error of the model. During the training process, the model performance is evaluated by methods such as cross-validation. For example, five-fold cross-validation can be used to evaluate the generalization ability of the model. The network structure and hyperparameters are dynamically adjusted according to the evaluation results. For example, if the accuracy of the model on the test set is low, the number of convolutional layers can be increased or hyperparameters such as the learning rate can be adjusted to improve the performance of the model.Through continuous adjustment and optimization, a robust fall detection model is finally obtained. The optimized convolutional neural network model is used to perform fall detection on newly acquired multi-view radar signals. For example, when a new radar signal is acquired, time-frequency analysis and preprocessing are first performed, and then multi-domain fall feature vectors are extracted. Finally, the feature vectors are input into the trained convolutional neural network model for prediction.

[0080] The two prediction results are fused to obtain the final result. The fusion method includes fusion by assigning weights to different detection results.

[0081] In this embodiment, an attention mechanism is also introduced to adaptively weight the fall-sensitive features. The attention mechanism learns the importance weights of different features, dynamically adjusts the importance of each feature component, highlights the key information of the fall event, suppresses the influence of background noise, and forms an efficient and reliable fall detection solution.

[0082] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A personnel fall detection system based on millimeter wave radar, characterized in that: include: A signal preprocessing module is used to preprocess the original radar data to obtain preprocessed data; A feature extraction module, used to extract the fall-sensitive features of the preprocessed data and perform dynamic time warping to obtain a fall feature sequence; A classification training module, used for training a support vector machine according to the fall feature sequence to obtain an adaptive fall determination model; A multi-view signal acquisition module, used to acquire multi-view radar echo signals through a multi-antenna array; A deep learning module is used to train a convolutional neural network using multi-view radar echo signals to obtain a highly robust fall detection model; The recognition module is used to detect the data to be detected by using the adaptive fall determination model and the high-robustness fall detection model, and to fuse the detection results of the adaptive fall determination model and the high-robustness fall detection model to obtain a final detection result.

2. The system according to claim 1, characterized in that The signal preprocessing module comprises: A wavelet transform unit, used to reconstruct the original radar data to obtain a denoised radar echo signal; A detection unit is used to perform quality assessment on the denoised radar echo signal.

3. The system according to claim 1, characterized in that The feature extraction module comprises: A transform unit, used for obtaining a time-frequency spectrum of a signal by short-time Fourier transform; An extraction unit extracts fall-sensitive features according to the energy distribution characteristics of the fall action in the time-frequency domain, wherein the fall-sensitive features include energy mutation and spectrum expansion; The warping unit is used to segment and align the radar signal of the falling process using a dynamic time warping algorithm.

4. The system according to claim 3, characterized in that The regularization unit comprises: A mapping unit, used to map the falling processes of different durations to a unified time scale through a nonlinear time normalization method; a difference elimination unit for resampling and interpolating the fall feature parameters according to the time scale; The serialization processing unit is used to perform serialization processing on the resampled and interpolated fall feature parameters to obtain a normalized fall feature sequence.

5. The system according to claim 1, characterized in that The classification training module includes: A first construction unit is used to construct an initial adaptive fall determination model according to a support vector machine; A first training unit, configured to train the initial adaptive fall determination model through a fall feature sequence; The first verification unit is used to evaluate the trained initial adaptive fall determination model and optimize parameters according to the evaluation result to obtain the adaptive fall determination model.

6. The system according to claim 5, characterized in that The training unit comprises a classification unit, which uses a kernel function to map feature samples in a fall feature sequence to a high-dimensional space, and classifies the feature samples by finding a maximum interval hyperplane.

7. The system according to claim 1, characterized in that The multi-view signal acquisition module comprises: A signal acquisition unit, used for acquiring multi-view radar echo signals through a multi-antenna array; The signal processing unit is used to process the multi-view radar echo signal by a spatial spectrum estimation method to obtain a denoised signal.

8. The system according to claim 1, characterized in that The deep learning module includes: A second construction unit is used to construct an initial high-robustness fall determination model according to a convolutional neural network; A second training unit, configured to train the initial high-robustness fall determination model through a denoised signal; The second verification unit is used to evaluate the trained initial high-robustness fall determination model through a cross-validation method, and optimize parameters according to the evaluation results to obtain an adaptive fall determination model.

Images