Human fall behavior detection method, device, equipment, storage medium and program product

Abstract

This application relates to the field of blockchain technology, and discloses a method, apparatus, device, storage medium, and program product for detecting human fall behavior. The method includes: acquiring raw video through a camera; correcting the measurement matrix based on a chaotic pseudo-random algorithm; using the measurement matrix to perform compressed sensing processing on the raw video to obtain CS data; extracting features from the CS data to identify foreground moving targets; inputting the foreground moving targets into a BWACGAN network; embedding and compensating information from the foreground moving targets based on the network to detect whether human fall behavior exists in the raw video; generating abnormal data if detected; and sending the abnormal data to a blockchain node to initiate a transaction for alarm processing. This application provides an architecture for processing human behavior information, and blockchain technology ensures the immutability of home video surveillance information. Furthermore, the improved network used to embed and compensate information from the extracted features solves the problem of insufficient sample size and protects user privacy.

Description

Technical Field

[0001] This application relates to the field of blockchain technology, specifically to methods, devices, equipment, storage media, and program products for detecting human fall behavior. Background Technology

[0002] The core idea of ​​compressed sensing (CS) theory is based on the sparsity of sampled signals, integrating the sampling and compression processes into the signal acquisition process. In existing research, commonly used measurement matrices mainly include deterministic matrices, partially orthogonal matrices, and random matrices. Deterministic measurement matrices have fixed generation methods and structures, resulting in low complexity and rapid construction. Partially orthogonal matrices are obtained by extracting certain behaviors from orthogonal matrices and then normalizing them, thus exhibiting approximately orthogonal characteristics and good reconstruction performance. Random matrices contain elements that satisfy a certain probability distribution and are independent of each other. A common characteristic of these measurement matrices is their low coherence with most sparse signals, allowing for reconstruction of the original signal with only small measurement values ​​to achieve the desired effect.

[0003] Generative Adversarial Networks (GANs) utilize two networks—a generator and a discriminator—for adversarial training, updating the network weights through backpropagation during training. As a generative model, GANs and their derivatives are frequently used for data augmentation and preprocessing to generate sample data, making them widely applicable in fields such as image processing and biomedicine.

[0004] However, the measurement matrices commonly used in current CS processing, such as deterministic measurement matrices, suffer from poor reconstruction results due to their "fixedness." Furthermore, in practical applications, the matrix needs to be matched with the problem rather than being flexibly adjustable. Partially orthogonal matrices have the disadvantage of having a relatively narrow scope of application and significant limitations on dimensions, making them unsuitable for use in real-world problem scenarios. As for the currently widely used random matrices, they can cause instability in experimental results, requiring numerous experiments to take the average to reduce the impact of the random matrix's uncertainty on the experiments, increasing computational complexity and making them difficult to implement in hardware.

[0005] Traditional Generative Adversarial Networks (GANs) are primarily used for dataset expansion and sample construction. On one hand, current GAN variants almost entirely lack architectures that process human behavioral information, thus failing to consider temporal information between video sequences. However, in fall detection research, behavioral information between adjacent video frames is crucial for fall detection. On the other hand, some GAN variants are often suited to large datasets, while currently available fall detection datasets have relatively few samples. Therefore, the network architecture should not be too deep, and the stability of training under limited sample conditions must also be considered.

[0006] Existing blockchain-based video surveillance technology mainly considers the distributed storage and immutability of data. However, in the current context, when detecting falls among elderly people living alone, it often involves some of their private behaviors (such as bathing and using the toilet). In this case, it is not only necessary to consider the security and immutability of the data, but more importantly, the personal privacy of the elderly. Summary of the Invention

[0007] This application provides a method, apparatus, device, storage medium, and program product for detecting human fall behavior, in order to solve the technical problems of insufficient sample size, inability of GAN models to be applied in real-world scenarios of elderly falls, and exposure of user privacy in the prior art.

[0008] In a first aspect, embodiments of this application provide a method for detecting human fall behavior, comprising: acquiring an original video through a camera, and sending the original video, shooting location, and shooting time to a blockchain; correcting a non-negative Gaussian measurement matrix based on a chaotic pseudo-random algorithm, and performing compressed sensing processing on the original video using the non-negative Gaussian measurement matrix to obtain CS data; extracting features from the CS data using a low-rank sparse decomposition algorithm to extract foreground moving targets; inputting the foreground moving targets into a BWACGAN network, and performing information embedding compensation on the foreground moving targets based on the BWACGAN network to detect whether human fall behavior exists in the original video; generating abnormal data when human fall behavior is detected in the original video; sending the abnormal data to a blockchain node to initiate a transaction for alarm processing; wherein, the transaction objects include emergency centers and pre-set emergency contacts.

[0009] In one embodiment, a non-negative Gaussian measurement matrix is ​​corrected based on a chaotic pseudo-random algorithm, and compressed sensing processing is performed on the original video using the non-negative Gaussian measurement matrix to obtain CS data. This includes: generating pseudo-random numbers using the Mason tween algorithm; inputting the pseudo-random numbers into a preset chaotic model to generate a chaotic pseudo-random sequence; mapping the chaotic pseudo-random sequence with a sign function to obtain a binary spread spectrum sequence; processing the binary spread spectrum sequence to obtain a sequence of corresponding correlation distributions, and integrating the sequences of correlation distributions into a chaotic pseudo-random Gaussian measurement matrix; and performing a linear transformation on the chaotic pseudo-random Gaussian measurement matrix to obtain CS data.

[0010] In one embodiment, the BWACGAN network includes an accuracy discriminator and a category discriminator. The accuracy discriminator outputs the detection accuracy and is used to determine the probability information that the current behavior is judged as a fall or not a fall. The category discriminator outputs the category to which the behavior belongs and is used to determine whether the current behavior is a fall or not a fall.

[0011] In one embodiment, the foreground moving target includes multiple video frame sequences corresponding to the original video; the foreground moving target is input into a BWACGAN network, and information embedding compensation is performed on the foreground moving target based on the BWACGAN network to detect whether there is a person falling in the original video, including: the BWACGAN network scores the human behavior columns in the multiple video frame sequences corresponding to the original video according to the characteristics of the fall behavior, and obtains a classification result based on the scoring result; wherein, the BWACGAN network includes tensor vectors, and the scoring result is used as a parameter in the tensor vector; the classification result is used to determine whether there is a person falling in the original video.

[0012] In one embodiment, embedding compensation of information for a foreground moving target based on the BWACGAN network includes: extracting features from the original video using a low-rank sparse decomposition algorithm to extract a second foreground moving target; and learning the BWACGAN network based on the second foreground moving target to complete the embedding compensation of information.

[0013] In one embodiment, abnormal data is sent to a blockchain node to initiate a transaction for alarm processing, including: sending the abnormal data to a gateway; the gateway performing secondary confirmation of the person's fall behavior; if the confirmation is correct, the gateway sends the abnormal data to an oracle to verify its reliability; if the oracle verification is successful, the abnormal data is sent to a blockchain node to initiate a transaction for alarm processing.

[0014] Secondly, embodiments of this application provide a device for detecting human fall behavior, comprising: a shooting module for acquiring original video through a camera and sending the original video, shooting location, and shooting time to a blockchain; a CS data module for correcting a non-negative Gaussian measurement matrix based on a chaotic pseudo-random algorithm and performing compressed sensing processing on the original video using the non-negative Gaussian measurement matrix to obtain CS data; a feature extraction module for extracting features from the CS data using a low-rank sparse decomposition algorithm to extract foreground moving targets; a network processing module for inputting the foreground moving targets into a BWACGAN network and performing information embedding compensation on the foreground moving targets based on the BWACGAN network to detect whether human fall behavior exists in the original video; and a blockchain transaction module for generating abnormal data when human fall behavior is detected in the original video; the abnormal data is sent to a blockchain node to initiate a transaction for alarm processing; wherein the transaction objects include emergency centers and pre-set emergency contacts.

[0015] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the program to implement the steps of the method for detecting human fall behavior of the first aspect.

[0016] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method for detecting human fall behavior as described in the first aspect.

[0017] Fifthly, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, implements the steps of the method for detecting human fall behavior according to the first aspect.

[0018] This application provides a method, apparatus, device, storage medium, and program product for detecting human fall behavior, comprising: acquiring original video through a camera; sending the original video, shooting location, and shooting time to a blockchain; correcting a non-negative Gaussian measurement matrix based on a chaotic pseudo-random algorithm, and performing compressed sensing processing on the original video using the non-negative Gaussian measurement matrix to obtain CS data; extracting features from the CS data using a low-rank sparse decomposition algorithm to extract foreground moving targets; inputting the foreground moving targets into a BWACGAN network, and performing information embedding compensation on the foreground moving targets based on the BWACGAN network to detect whether human fall behavior exists in the original video; generating abnormal data when human fall behavior is detected in the original video; sending the abnormal data to a blockchain node to initiate a transaction for alarm processing; wherein, the transaction objects include emergency centers and pre-set emergency contacts. Through the above method, this application provides an architecture for processing human behavior information, and blockchain technology ensures that home video surveillance information is tamper-proof. Simultaneously, the improved network is used to embed and compensate the extracted features, solving the problem of insufficient sample size and protecting user privacy. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the fall detection method of this application;

[0021] Figure 2 This is a simplified network structure diagram of an embodiment of the BWACGAN network of this application;

[0022] Figure 3 This is a flowchart illustrating another embodiment of the fall detection method of this application;

[0023] Figure 4 This is a schematic diagram of the structure of an embodiment of the fall detection device of this application;

[0024] Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] This application provides a method for detecting a person's fall behavior. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the method for detecting fall behavior according to this application. In this embodiment, the method for detecting human fall behavior includes steps S110 to S140, each step of which is as follows:

[0027] S110: Obtain the original video through the camera and send the original video, shooting location and shooting time to the blockchain.

[0028] The method for detecting fall behavior in this application can be applied to detect fall behavior in elderly people living alone. When a fall is detected, a timer can be set up to issue an alarm, providing timely medical assistance to the elderly person living alone. The camera can be installed in areas where the elderly person living alone frequently goes.

[0029] S120: The non-negative Gaussian measurement matrix is ​​corrected based on a chaotic pseudo-random algorithm, and the original video is compressed and sensed using the non-negative Gaussian measurement matrix to obtain CS data.

[0030] In one embodiment, the steps of correcting the non-negative Gaussian measurement matrix based on a chaotic pseudo-random algorithm and using the non-negative Gaussian measurement matrix to perform compressed sensing processing on the original video to obtain CS data include:

[0031] The first step is to use the Mason twitch algorithm to generate pseudo-random numbers.

[0032]

[0033] The second step is to obtain the chaotic sequence. The aforementioned random numbers are input as the original signal into the formula.

[0034] (2) The corresponding chaotic model generates a chaotic pseudo-random sequence. and

[0035] The third step is to generate the spread spectrum sequence. The chaotic sequence obtained above... and Through the symbol function mapping in equation (2), the chaotic sequence is mapped into a binary spread spectrum sequence. and

[0036] a n =sgn(x n )……(3)

[0037] The fourth step is to generate a chaotic pseudo-random Gaussian measurement matrix. The generated binary spread spectrum sequence is processed by equation (3) to obtain the corresponding correlation distribution sequence. Then, the correlation distribution sequences are integrated into an m×n dimensional chaotic pseudo-random Gaussian measurement matrix.

[0038] The fifth step is to measure the matrix. Perform a linear transformation. Based on the measurement matrix obtained above... Non-negative values ​​will exist, therefore, to prevent feature loss, we first select... The minimum value in the matrix is ​​obtained by performing a global translation transformation on all elements of the matrix, shifting the minimum value to the range of 0, to obtain the measurement matrix. This processing method ensures that all elements in the matrix are greater than or equal to 0. Furthermore, considering practical application, the shifted matrix is ​​subjected to peak normalization to ensure that all elements are distributed within the range [0,1]. The above process can be expressed mathematically as follows:

[0039]

[0040] The chaotic pseudo-random modified non-negative Gaussian measurement matrix not only performs better in video dimensionality reduction, but also introduces less noise and reduces the loss of behavioral information in the video during the dimensionality reduction process.

[0041] S130: Use a low-rank sparse decomposition algorithm to extract features from the CS data and extract the foreground moving targets.

[0042] While compressive sensing processing of video data achieves visual privacy protection, it also introduces a lot of noise and causes the video to lose some information. Therefore, this application uses a low-rank sparse decomposition algorithm to extract the foreground information of the video and feeds the foreground information into an improved GAN network.

[0043] S140: Input the foreground moving target into the BWACGAN network, and perform information embedding compensation on the foreground moving target based on the BWACGAN network to detect whether there is a person falling in the original video.

[0044] Optionally, the BWACGAN network includes an accuracy discriminator and a category discriminator. The accuracy discriminator outputs the detection accuracy and is used to determine the probability information of whether the current behavior is judged as a fall or not. The category discriminator outputs the category to which the behavior belongs and is used to determine whether the current behavior is a fall or not.

[0045] In one embodiment, the foreground moving target includes multiple video frame sequences corresponding to the original video; the step of inputting the foreground moving target into the BWACGAN network and performing information embedding compensation on the foreground moving target based on the BWACGAN network to detect whether there is a person falling in the original video specifically includes:

[0046] The BWACGAN network scores human behavior columns in multiple video frames corresponding to the original video based on the characteristics of fall behavior, and obtains a classification result based on the scoring results. The BWACGAN network includes tensor vectors, and the scoring results are used as parameters in the tensor vectors. The classification results are used to determine whether there is a human fall behavior in the original video.

[0047] In one embodiment, the step of embedding and compensating for information on the foreground moving target based on the BWACGAN network specifically includes:

[0048] The original video is used to extract features using a low-rank sparse decomposition algorithm to extract the second foreground moving target; the BWACGAN network is then used to learn based on the second foreground moving target to complete the information embedding compensation.

[0049] S150: If a person is detected falling in the original video, abnormal data is generated; the abnormal data is sent to the blockchain node to initiate a transaction for alarm processing.

[0050] The transaction partners can include emergency centers and pre-set emergency contacts.

[0051] In one embodiment, abnormal data is sent to a blockchain node to initiate a transaction for alarm processing, including: sending the abnormal data to a gateway; the gateway performing secondary confirmation of the person's fall behavior; if the confirmation is correct, the gateway sends the abnormal data to an oracle to verify its reliability; if the oracle verification is successful, the abnormal data is sent to a blockchain node to initiate a transaction for alarm processing.

[0052] This application provides a method for detecting human fall behavior. The method involves acquiring raw video data via a camera and sending the raw video, shooting location, and shooting time to a blockchain. A non-negative Gaussian measurement matrix is ​​corrected using a chaotic pseudo-random algorithm, and the raw video is then compressed using this non-negative Gaussian measurement matrix to obtain CS data. Feature extraction is performed on the CS data using a low-rank sparse decomposition algorithm to extract foreground moving targets. These foreground moving targets are then input into a BWACGAN network, and information embedding compensation is performed based on the BWACGAN network to detect whether a human fall behavior exists in the raw video. If a human fall behavior is detected in the raw video, abnormal data is generated. This abnormal data is sent to a blockchain node to initiate a transaction for alarm processing. The transaction recipients include emergency centers and pre-set emergency contacts. Through this method, this embodiment provides an architecture for processing human behavior information, and blockchain technology ensures the immutability of home video surveillance information. Furthermore, the improved network used for information embedding compensation of the extracted features solves the problem of insufficient sample size.

[0053] The BWACGAN network is one of the key improvements in this application, and further explanation of the BWACGAN network follows. Please refer to [link / reference]. Figure 2 , Figure 2 This is a simplified network structure diagram of an embodiment of the BWACGAN network in this application.

[0054] The existing BWACGAN network also consists of two parts: a generator network and a discriminator network. The model's input is CS state data and type information C, and the output is the detection accuracy R of the current behavior and the category Y to which the behavior belongs. The main difference between the BWACGAN network proposed in this application and the existing BWACGAN network lies in the fact that, for the basic neural networks corresponding to the existing generator G and discriminator D, this model has the problem of the discriminator misclassifying the generated data. Specifically, in the field of computer vision, existing GAN networks are mainly used for processing image data. In the process of image processing, there is no need to consider the temporal information between images. Therefore, using the current GAN network for optimization cannot accurately classify fall behavior. At the same time, the features of fall video data are diverse, and the data range of the features is inconsistent. The number of video frames corresponding to each video sequence, the specific time of the fall in each video sequence, and other information are different. Unlike the pixels of an image, the features of video sequence data have different importance. Each pixel of an image may play an important role in the result, but in the specific video sequence detection process, different video sequences play different roles in training and learning. Therefore, to address the uniqueness of video sequences and resolve the misjudgment problem inherent in discriminators, this application proposes... Figure 2The intermediate architecture in the model has been improved in three aspects based on the basic model:

[0055] First, the discriminator is divided into detection accuracy and category information classification. Behavior detection accuracy only requires the probability information of whether the current behavior is judged as a fall or not, while category information classification only requires classifying the behaviors in the current video sequence, that is, giving whether the behavior in the current video is a fall or not.

[0056] Secondly, batch discrimination is incorporated. In video surveillance for fall detection, traditional GAN ​​networks often only classify individual images, which undoubtedly affects the accuracy of fall detection. Therefore, this application introduces the concept of batch discrimination into the BWACGAN network. In practical use, this can be understood as classifying a batch of video frames corresponding to a video, rather than a single image. The entire video sequence is then comprehensively scored to obtain a final classification result. Algorithmically, this can be represented as follows: a tensor vector is multiplied in the intermediate layer of the discriminator. Based on the characteristics of the fall, the algorithm scores the human behavior features in the current video sequence, and this score is used as the parameter in the tensor vector. Pre-training reveals that almost all fall-related video sequences exhibit a Gaussian-like distribution in their scores; that is, the first half of the video always contains normal behavior, followed by a fall at some point until the entire body collapses.

[0057] Therefore, when a fall occurs, a larger parameter value is assigned when operating on the intermediate layer f(x), resulting in the current video frame's judgment result having a larger weight in the overall result. For non-fall video sequences, the tensor vector's numerical distribution exhibits a relatively flat curve, and the parameter values ​​assigned when operating on the intermediate layer f(x) are not significantly different. Next, multiplying this vector by the intermediate layer f(x) yields a matrix M. The L1 distance between each row of matrix M is calculated and used as the input to the next layer of f(x). A schematic diagram of this process is shown below. Figure 2 As shown. Its mathematical process can be described as follows:

[0058]

[0059] 0(x i )=[o(x i )1, o(x i )2,…,o(x i ) B ]∈R B ……(6)

[0060] o(X)∈Rn×B ……(7)

[0061] Where, o(x i ) represents the L1 distance corresponding to each row, o(x i )1 represents the L1 distance of the first row, o(x i ) b Let f(x) represent the L1 distance of the b-th row. If x represents a non-falling behavior, then the corresponding f(x) must be very similar, and the values ​​in the tensor vectors will be relatively flat. Therefore, the resulting matrix M must also be very similar. The L1 distance C(x) obtained from each row of matrix M is... i ,x j This is very close to 0, making o(x) i The vector is approximately close to zero. Conversely, if x represents a falling action, the corresponding f(x) will differ, and the distribution of the tensor vectors will also satisfy a Gaussian-like distribution, making o(x)... i It can never approach the zero vector. At this point, the discriminator can accurately distinguish between falling behavior and non-falling behavior.

[0062] The problem with intermediate models is that the training speed differs between actual results and classification results, leading to instability in the training process. Based on this, the final model architecture was developed, primarily by fusing the first layer of the classification network with the accuracy detection network, thereby improving the stability and speed of the training process. Furthermore, considering the influence of indoor lighting intensity and camera placement on video shooting in reality, this application further modifies this tensor vector. We first set a baseline, and by calculating the mean and variance of the current video frame sequence, if the mean and variance deviate from the baseline, it indicates that the current video sequence is in an environment with poor exposure or lighting. In this case, equalization processing is required with each video frame, and the weights are adjusted accordingly.

[0063] Optionally, considering the small amount of fall detection sample data, this application adjusts the convolution depth of the two networks to 6 layers.

[0064] The main modification of the network training process in this application is that the discriminator is divided into detection accuracy and category information categories, and trained separately.

[0065] This application improves the training method, wherein the generator training process is as follows: Figure 5 As shown, by modifying the network weights of the blank parts, the generated data G satisfies the objective functions represented by formulas (8) and (9).

[0066] L S =E[P(S=CS|X cs )]......(8)

[0067] L C =E[logP(C=c|X fake )]......(9)

[0068] Among them, L S L indicates the case where the generated data is judged as false. C This indicates that G can be correctly identified based on the input category information. The goal of training G is to make L... C -L S maximum.

[0069] The goal of training the detection accuracy module (R(D)) corresponding to the discriminator is to modify the network weights in the blank parts so that R(D) can distinguish between real data and generated data. The corresponding objective function is:

[0070] L S =E[P(S=real|X real )]+E[P(S=CS|X cs )]+E[P(grad=1|grad(D(X t )))]......(10)

[0071] X t =tX real +(1-t)X cs , 0≤t≤1......(11)

[0072] Wherein, E[P(S=real|X real )] represents the expected recognition of real data, E[P(S=CS|X cs )] represents the expectation of recognition of the generated data, E[P(grad=1|grad(D(X)] t )))] represents gradient penalty.

[0073] The training objective of the category information classification module (C(D)) corresponding to the training discriminator is to enable C(D) to distinguish the category Y corresponding to the real data. At the same time, in order to meet the adjustment of the weights of R(D), the overlapping part between the two also needs to be adjusted after each training.

[0074] L C =E[logP(C=c|X real )]......(12)

[0075] Where C represents category label information, X real Represents real data.

[0076] In summary, in actual training, all three aspects are trained simultaneously. Therefore, the above training method can be summarized as follows:

[0077]

[0078] Among them, D R and D C Let P represent the two overlapping parts in the discriminator network, and G represent the generator. r and P g represents the distribution of real data and generated data, respectively, and Y represents the category to which the classification belongs.

[0079] The first row of formula (13) It is a process of "game" between the generator and the discriminator, with the goal of making the generated data as close as possible to the real data. In this process, since the discriminator network D is trained by splitting it, compared with the original conditional generative adversarial network, there is no need to perform logarithmic function operations. Only the expectation needs to be processed, which simplifies the computational complexity.

[0080] In the second row of formula (13) Some of the proposed changes to D R The weight parameters are set to minimize the gradient penalty for this part.

[0081] In the second row of formula (13) And the third row of formula (13) All correspond to the classification section, where D C The goal of G is to optimize the classification result, and d C Equivalent to D C However, it will not participate in updates.

[0082] Furthermore, the modifications to the discriminator D are also reflected in: Figure 2 The middle slash padding also incorporates a portion of the first-layer discriminator network, thus the upper and lower parts of the network "grow" their own distinct network structures. Since the R(D) part affects all nodes in the first layer, the batch decision mentioned above will be implemented across the entire first layer.

[0083] To address the temporal nature of fall detection, this application adjusts the input to the Generative Adversarial Network (GAN). Traditional GANs are primarily used for data augmentation, i.e., generating data from scratch. However, the GAN in this application is mainly used for classification, thus requiring adjustments to the network structure. Specifically, to address the current scarcity of fall detection datasets, the number of layers in the generator and discriminator networks is increased to six, and the generator's input is replaced with CS-mode video frame data instead of noisy data.

[0084] Please see Figure 3 , Figure 3 This is a flowchart illustrating another embodiment of the fall detection method of this application. A camera captures an original image, and the original image, shooting location, and shooting time are uploaded to the blockchain. Then, compressed sensing processing is performed on the original video based on a chaotic pseudo-random modified non-negative Gaussian measurement matrix. The data from the compressed sensing processing is used to extract foreground moving targets using a low-rank sparse decomposition algorithm. Next, a test dataset is input, and compressed sensing processing is performed on the videos in the test dataset based on a chaotic pseudo-random modified non-negative Gaussian measurement matrix. The data from the compressed sensing processing is used to extract foreground moving targets using a low-rank sparse decomposition algorithm. Figure 3 (As shown by the dashed line). Next, the original video was also processed using a low-rank sparse decomposition algorithm to extract the second foreground moving target.

[0085] The data extracted using the low-rank sparse decomposition algorithm is input into the BWACGAN network. Based on the BWACGAN network, feature embedding learning compensation and classifier training are implemented to detect whether elderly people in the video exhibit fall behavior. Finally, blockchain technology is used to send the processed feature information to the gateway. The blockchain uses oracles for secondary verification and alerts regarding fall behavior. Once verification is successful, the abnormal data is sent to blockchain nodes. Nodes initiate transactions based on the abnormal data, with the other party primarily including emergency centers and the elderly person's emergency contacts. Simultaneously, all blockchain nodes forward the transaction.

[0086] In this application, blockchain technology serves two main purposes. First, it ensures the immutability of home video surveillance information. Since the adversarial network used in this application needs to learn the original video features when embedding and compensating for information, this application aims to upload the original video to the blockchain with the highest access level, disallowing viewing and tampering. Second, this application utilizes blockchain technology for secondary behavior verification. The behavior classification results obtained by the adversarial network, along with the camera's location when the original video is uploaded to the blockchain, are sent to the gateway. The gateway then sends detected ethically abnormal data and information to an oracle for secondary reliability verification. Once the oracle verifies successfully, the abnormal data is sent to the blockchain nodes. The blockchain nodes initiate transactions based on the abnormal data to trigger alerts, with the other party in the transaction being the emergency center and the elderly person's designated emergency contact.

[0087] The optimized apparatus for detecting human fall behavior provided in this application will be described below. The apparatus for detecting human fall behavior described below can be referred to in correspondence with the method for detecting human fall behavior described above.

[0088] Please see Figure 4 , Figure 4This is a schematic diagram of an embodiment of the fall detection device of this application. The fall detection device includes:

[0089] The shooting module 410 is used to obtain the original video through the camera and send the original video, shooting location and shooting time to the blockchain.

[0090] The CS data module 420 is used to correct the non-negative Gaussian measurement matrix based on a chaotic pseudo-random algorithm, and to perform compressed sensing processing on the original video using the non-negative Gaussian measurement matrix to obtain CS data.

[0091] The feature extraction module 430 is used to extract features from CS data using a low-rank sparse decomposition algorithm to extract foreground moving targets.

[0092] The network processing module 420 is used to input the foreground moving target into the BWACGAN network, and to perform information embedding compensation on the foreground moving target based on the BWACGAN network in order to detect whether there is a person falling in the original video.

[0093] The blockchain transaction module 450 is used to generate abnormal data when a person falls in the original video. The abnormal data is sent to the blockchain node to initiate a transaction for alarm processing. The transaction objects include emergency centers and pre-set emergency contacts.

[0094] In one embodiment, the CS data module 420 is used for:

[0095] Pseudo-random numbers are generated using the Mason twitch algorithm; these pseudo-random numbers are input into a pre-defined chaotic model to generate a chaotic pseudo-random sequence; the chaotic pseudo-random sequence is mapped using a sign function to obtain a binary spread spectrum sequence; the binary spread spectrum sequence is processed to obtain a sequence of corresponding correlation distributions, and the sequences of correlation distributions are integrated into a chaotic pseudo-random Gaussian measurement matrix; the chaotic pseudo-random Gaussian measurement matrix is ​​linearly transformed to obtain CS data.

[0096] In one embodiment, the BWACGAN network includes an accuracy discriminator and a category discriminator. The accuracy discriminator outputs the detection accuracy and is used to determine the probability information that the current behavior is judged as a fall or not a fall. The category discriminator outputs the category to which the behavior belongs and is used to determine whether the current behavior is a fall or not a fall.

[0097] In one embodiment, the network processing module 420 is used for:

[0098] The BWACGAN network scores human behavior columns in multiple video frames corresponding to the original video based on the characteristics of fall behavior, and obtains a classification result based on the scoring results. The BWACGAN network includes tensor vectors, and the scoring results are used as parameters in the tensor vectors. The classification results are used to determine whether there is a human fall behavior in the original video.

[0099] In one embodiment, the network processing module 420 is used for:

[0100] The original video is used to extract features using a low-rank sparse decomposition algorithm to extract the second foreground moving target; the BWACGAN network is then used to learn based on the second foreground moving target to complete the information embedding compensation.

[0101] In one embodiment, the blockchain transaction module 450 is used for:

[0102] The abnormal data is sent to the gateway; the gateway performs secondary confirmation on the person's fall behavior; if the confirmation is correct, the gateway sends the abnormal data to the oracle for reliability verification; if the oracle verification is successful, the abnormal data is sent to the blockchain node to initiate a transaction for alarm processing.

[0103] This application also provides an electronic device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of an embodiment of the electronic device of this application. In this embodiment, the electronic device may include a memory 510, a processor 520, and a computer program stored in the memory 520 and executable on the processor 510. When the processor 510 executes the program, it implements the methods for detecting human fall behavior provided by the above-described methods.

[0104] Optionally, the electronic device may further include a communication bus 530 and a communication interface 540, wherein the processor 510, the communication interface 540, and the memory 520 communicate with each other via the communication bus 530. The processor 510 can call logical instructions in the memory 520 to execute a method for detecting human fall behavior, the method including:

[0105] The system acquires raw video footage via a camera and sends the raw video, shooting location, and shooting time to the blockchain. A non-negative Gaussian measurement matrix is ​​corrected using a chaotic pseudo-random algorithm, and the raw video is then compressed using this matrix to obtain CS data. Feature extraction is performed on the CS data using a low-rank sparse decomposition algorithm to extract foreground moving targets. These foreground moving targets are then input into a BWACGAN network, which performs information embedding compensation to detect whether a person is falling in the raw video. If a person is detected falling in the raw video, abnormal data is generated. This abnormal data is sent to blockchain nodes to initiate transactions for alarm processing. The transaction recipients include emergency centers and pre-set emergency contacts.

[0106] Furthermore, the logical instructions in the aforementioned memory 520 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0107] On the other hand, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the methods for detecting human fall behavior provided by the above methods. The steps and principles of the methods have been described in detail in the above methods and will not be repeated here.

[0108] In another aspect, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the methods for detecting human fall behavior provided by the above methods. The steps and principles of these methods have been described in detail in the above methods and will not be repeated here.

[0109] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0110] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method of detecting a fall behavior of a person, characterized by, The method comprises the following steps: obtaining an original video through a camera, and sending the original video, a shooting location and a shooting time to a blockchain; correcting a non-negative Gaussian measurement matrix based on a chaotic pseudo-random algorithm, and performing compressive sensing processing on the original video by using the non-negative Gaussian measurement matrix to obtain CS data; extracting a foreground moving target by using a low-rank sparse decomposition algorithm on the CS data; inputting the foreground moving target into a BWACGAN network, and embedding and compensating information of the foreground moving target based on the BWACGAN network to detect whether a falling behavior of a person exists in the original video; generating abnormal data in the case that the falling behavior of the person is detected in the original video; sending the abnormal data to a blockchain node to initiate a transaction for alarm processing; wherein the transaction object comprises an emergency center and a pre-set emergency contact person; the BWACGAN network comprises an accuracy discriminator and a category discriminator, the accuracy discriminator outputs detection accuracy, and the accuracy discriminator is used to determine the probability information that the current behavior is judged as falling or non-falling; the category discriminator outputs the category, and the category discriminator is used to determine whether the current behavior is falling or non-falling; the foreground moving target comprises a plurality of video frame sequences corresponding to the original video; the BWACGAN network is used to embed and compensate information of the foreground moving target based on the BWACGAN network to detect whether a falling behavior of a person exists in the original video, comprising: the BWACGAN network scores the human behavior column in the plurality of video frame sequences corresponding to the original video according to the characteristics of the falling behavior, and obtains a classification result based on the scoring result; wherein the BWACGAN network comprises a tensor vector, and the scoring result is used as a parameter in the tensor vector; the classification result is used to determine whether a falling behavior of a person exists in the original video; the BWACGAN network is used to embed and compensate information of the foreground moving target based on the BWACGAN network, comprising: extracting a second foreground moving target by using a low-rank sparse decomposition algorithm on the original video; the BWACGAN network learns based on the second foreground moving target to complete the embedding and compensation of information.

2. The method of detecting a fall behavior of a person according to claim 1, wherein The method comprises the following steps: generating a pseudo-random number by using a Meissner rotation algorithm; inputting the pseudo-random number into a pre-set chaotic model to generate a chaotic pseudo-random sequence; mapping the chaotic pseudo-random sequence by using a sign function to obtain a binary spread spectrum sequence; obtaining a corresponding correlation distribution sequence by processing the binary spread spectrum sequence, and integrating the correlation distribution sequence into a chaotic pseudo-random Gaussian measurement matrix; performing linear transformation on the chaotic pseudo-random Gaussian measurement matrix to obtain the CS data.

3. The method of detecting a fall behavior of a person according to claim 1, wherein The abnormal data is sent to the blockchain node to initiate a transaction for alarm processing, comprising: sending the abnormal data to a gateway; the gateway secondarily confirming the falling behavior of the person; if the confirmation is correct, the gateway sends the abnormal data to an oracle to verify the reliability by using the oracle; if the oracle verifies successfully, the abnormal data is sent to a blockchain node to initiate a transaction for alarm processing.

4. A device for detecting human fall behavior, characterized in that, comprising: a shooting module, configured to obtain an original video by a camera, and send the original video, a shooting location and a shooting time to a blockchain; a CS data module, configured to correct a non-negative Gaussian measurement matrix based on a chaos pseudo-random algorithm, and perform compressive sensing processing on the original video by using the non-negative Gaussian measurement matrix to obtain CS data; a feature extraction module, configured to perform feature extraction on the CS data by using a low-rank sparse decomposition algorithm to extract a foreground moving target; a network processing module, configured to input the foreground moving target into a BWACGAN network, and perform information embedding compensation on the foreground moving target based on the BWACGAN network to detect whether a falling behavior of the person exists in the original video; a blockchain transaction module, configured to generate abnormal data if it is detected that the falling behavior of the person exists in the original video; the abnormal data is sent to a blockchain node to initiate a transaction for alarm processing; wherein, transaction objects include an emergency center and pre-set emergency contacts; the BWACGAN network comprises an accuracy discriminator and a category discriminator, the accuracy discriminator outputs detection accuracy, and the accuracy discriminator is used to determine probability information that a current behavior is judged as falling or non-falling; the category discriminator outputs a category, and the category discriminator is used to determine whether the current behavior is falling or non-falling; the network processing module is specifically configured to: the BWACGAN network scores human behavior columns in a plurality of video frame sequences corresponding to the original video according to characteristics of the falling behavior, and obtains a classification result based on a scoring result; wherein, the BWACGAN network comprises a tensor vector, the scoring result is used as a parameter in the tensor vector, and the classification result is used to determine whether the falling behavior of the person exists in the original video; the network processing module is specifically configured to: perform feature extraction on the original video by using a low-rank sparse decomposition algorithm to extract a second foreground moving target; and the BWACGAN network learns based on the second foreground moving target to complete information embedding compensation.

5. An electronic device comprising a processor and a memory having stored therein a computer program, characterized in that, The processor executes the computer program to realize the steps of the method for detecting the falling behavior of the person according to any one of claims 1 to 3.

6. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method for detecting the falling behavior of the person according to any one of claims 1 to 3.

7. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method for detecting the falling behavior of the person according to any one of claims 1 to 3.

Images