A fall detection method and system based on a MobileNetV2 model
By introducing the CBAM attention mechanism into the MobileNetV2 model, and combining it with channel and spatial attention mechanisms, the problems of high computational pressure and information loss in existing algorithms on mobile devices are solved, achieving a significant performance improvement in fall detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-25
- Publication Date
- 2026-03-20
AI Technical Summary
Existing CNN-based fall detection algorithms suffer from high computational burden and ineffective integration of global information when applied to mobile devices. In particular, the fusion algorithm of CNN and attention mechanism ignores the interaction of channel and spatial information, resulting in information loss and insufficient detection performance.
The CBAM attention mechanism is introduced into the MobileNetV2 model, which combines channel and spatial attention mechanisms to improve the network structure and increase the capture rate of key information. The classifier is also processed to solve the problem of information loss.
Without increasing network computation, the network's capture rate of key information and the effectiveness of feature extraction were improved, thus enhancing the performance and accuracy of fall detection.
Smart Images
Figure CN115471909B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fall detection technology, and in particular relates to a fall detection method and system based on the MobileNetV2 model. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] With the increasing prevalence of population aging, the health of elderly people living alone has become a key focus in the medical and health field. Accidental falls are a leading cause of death among the elderly due to accidental injuries. When they lose mobility after a fall and are unable to save themselves, they may face extremely critical situations. Therefore, fall detection for elderly people living alone is crucial.
[0004] Fall detection can be abstracted as a binary classification problem, where the goal is to use algorithms to determine whether a person's final posture indicates a fall or not. With the emergence of networks such as Res-Net, VGG-Net, and Google-Net, CNNs (Convolutional Neural Networks) have been widely applied in deep learning and computer vision for classification and recognition tasks. Many researchers have chosen to improve upon CNNs to more effectively detect falls. However, due to the large number of parameters in the main network used by existing algorithms, porting them to mobile devices places a heavy burden on the device and makes them difficult to adapt to intelligent fall detection environments.
[0005] Because CNNs inherently have limitations and cannot effectively integrate global information, attention mechanisms have become a hot research topic to address this issue. The rapid development of attention mechanisms has provided new possibilities for network optimization. While simple attention mechanisms can effectively integrate contextual information, they face difficulties in acquiring target location information in video images and cannot effectively solve the location encoding problem. Therefore, many algorithms combine CNNs with attention mechanisms for improvement.
[0006] However, existing algorithms have all demonstrated that fusing CNNs with attention mechanisms can effectively improve algorithm applicability, but they only focus on one direction: channel attention or spatial attention. Spatial attention ignores information interaction within channels, while channel attention ignores information interaction between channels. The hybrid attention mechanism CBAM, on the other hand, combines channel and spatial attention mechanisms and is a plug-and-play lightweight convolutional attention module. Therefore, this invention chooses to draw on the working principle of CBAM to improve the block structure of MobileNetV2. Summary of the Invention
[0007] To solve the technical problems in the background art, the application provides a fall detection method and system based on a MobileNetV2 model, which introduces a CBAM attention mechanism in the MobileNetV2 model, improves the capture rate of key information without increasing the additional calculation amount of the network, and processes the classifier, so as to solve the problem of excessive information loss and make the network more effectively use the extracted feature information for fall detection.
[0008] To achieve the above-mentioned purpose, the application adopts the following technical solutions:
[0009] The first aspect of the application provides a fall detection method based on a MobileNetV2 model, which comprises:
[0010] Obtaining a human motion image to be detected;
[0011] Inputting the human motion image into a MobileNetV2 model based on an attention mechanism to obtain a fall detection result;
[0012] The MobileNetV2 model based on the attention mechanism adds a channel attention mechanism and a spatial attention mechanism of the CBAM attention mechanism before or in the feature extraction part of the MobileNetV2 network, and sequentially connects an attention layer, a pooling layer, a flattening operation and a fully connected layer after the feature extraction part.
[0013] Further, the channel attention mechanism first performs maximum pooling and average pooling on the input to obtain two vectors respectively; then inputs the two vectors into a weight-shared network; adds the two vectors after the weight-shared network, and activates them by using an activation function to obtain a channel attention correction vector; multiplies the channel attention correction vector with the input to obtain a feature map after the channel attention effect.
[0014] Further, the spatial attention mechanism performs maximum pooling and average pooling on the input, concatenates them according to the channel dimension, and performs convolution by using a convolution kernel to obtain a spatial attention vector after an activation function; multiplies the spatial attention vector with the input to obtain an output after the spatial attention effect.
[0015] Further, if the CBAM attention mechanism is added in the feature extraction part, the channel attention mechanism is added before the channel-by-channel convolution, and the spatial attention mechanism is added before the point-by-point convolution.
[0016] Further, if the CBAM attention mechanism is added before the feature extraction part, the channel attention mechanism and the spatial attention mechanism of the CBAM attention mechanism are added after the first convolutional layer of the MobileNetV2 model.
[0017] The second aspect of the present application provides a fall detection system based on a MobileNetV2 model, comprising:
[0018] an image acquisition module configured to acquire a human motion image to be detected;
[0019] a fall detection module configured to input the human motion image into the MobileNetV2 model based on an attention mechanism to obtain a fall detection result;
[0020] The MobileNetV2 model based on the attention mechanism adds the channel attention mechanism and the spatial attention mechanism of the CBAM attention mechanism before or in the feature extraction part of the MobileNetV2 network, and sequentially connects an attention layer, a pooling layer, a flattening operation and a fully connected layer after the feature extraction part.
[0021] Further, if the CBAM attention mechanism is added in the feature extraction part, the channel attention mechanism is added before the channel-wise convolution, and the spatial attention mechanism is added before the point-wise convolution.
[0022] Further, if the CBAM attention mechanism is added before the feature extraction part, the channel attention mechanism and the spatial attention mechanism of the CBAM attention mechanism are added after the first convolutional layer of the MobileNetV2 model.
[0023] The third aspect of the present application provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the fall detection method based on the MobileNetV2 model as described above.
[0024] The fourth aspect of the present application provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the steps of the fall detection method based on the MobileNetV2 model as described above.
[0025] Compared with the prior art, the present application has the following beneficial effects:
[0026] The application provides a fall detection method based on a MobileNetV2 model, which adds a CBAM attention mechanism before a MobileNetV2 network block structure framework, or adds a channel attention mechanism and a spatial attention mechanism inside the block, improves the capture rate of key information without increasing the additional calculation amount of the network, improves the feature extraction effectiveness of the network, and further improves the network detection performance.
[0027] The application provides a fall detection method based on a MobileNetV2 model, which processes a classifier, can solve the problem of excessive information loss caused by sudden reduction of feature dimension in the original framework, and makes the network more effectively use the extracted feature information for fall detection. BRIEF DESCRIPTION OF DRAWINGS
[0028] The drawings constituting a part of the specification of the application are used to provide further understanding of the application, the illustrative embodiments of the application and the description thereof are used to explain the application, and do not constitute improper limitation on the application.
[0029] Figure 1 is a schematic diagram of adding a CBAM attention mechanism before MobileNetV2 network feature extraction of the first embodiment of the application;
[0030] Figure 2 is a CBAM attention mechanism framework diagram of the first embodiment of the application;
[0031] Figure 3 is a block structure diagram when the step is 1 of the first embodiment of the application;
[0032] Figure 4 is a block structure diagram when the step is 2 of the first embodiment of the application;
[0033] Figure 5 is a block structure diagram after adding a channel attention and a spatial attention of the first embodiment of the application;
[0034] Figure 6 is a classifier part framework diagram of the first embodiment of the application;
[0035] Figure 7 is a model overall framework diagram after adding a CBAM attention mechanism in the network initially of the first embodiment of the application;
[0036] Figure 8 is a network model overall framework diagram after modifying the block structure of the network of the first embodiment of the application. DETAILED DESCRIPTION
[0037] The application will be further described below in combination with the drawings and embodiments.
[0038] It should be noted that the following detailed description is illustrative only, and is intended to further provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
[0039] Example one
[0040] The embodiment provides a fall detection method based on a MobileNetV2 model.
[0041] The fall detection task needs to consider the detection accuracy and the detection rate at the same time, and the algorithm needs to be embedded into a mobile device more, and be integrated into an intelligent home ecological environment, so that children can perform emergency treatment on the fallen old people in the first time. In order to consider the detection rate and the accuracy at the same time, the embodiment selects a lightweight neural network MobileNetV2 for improvement. MobileNetV2 has high detection rate and accuracy, but lacks attention mechanism in the framework, and cannot effectively improve the attention of the algorithm to the target that greatly influences the classification result. In order to solve the above problem, the embodiment proposes two schemes of adding attention mechanism in the MobileNetV2 network, so as to improve the classification performance of the algorithm.
[0042] The embodiment provides a fall detection method based on a MobileNetV2 model.
[0043] Step 1, obtaining a human motion image to be detected;
[0044] Step 2, inputting the human motion image into a MobileNetV2 model based on an attention mechanism to obtain a fall detection result, fall / no fall.
[0045] The MobileNetV2 model based on the attention mechanism adds a channel attention mechanism and a spatial attention mechanism of the CBAM attention mechanism in front of or in the feature extraction part of the MobileNetV2 network, and sequentially connects an attention layer, a pooling layer, a flattening operation and a full connection layer after the feature extraction part.
[0046] The MobileNetV2 model based on the attention mechanism adds the CBAM attention mechanism before the block structure of the MobileNetV2 network performs feature extraction, or adds the channel attention mechanism and the spatial attention mechanism inside the block of the MobileNetV2 network. The original block structure of the MobileNetV2 network is first a dimension increasing layer composed of a common convolution with a convolution kernel of 1, then a channel-by-channel convolution with a convolution kernel of 3, and finally a point-by-point convolution layer with a convolution kernel of 1 to fuse cross-channel information.
[0047] (1) Adding the CBAM attention mechanism before the block structure of the MobileNetV2 network.
[0048] The deep neural network has good learning ability in high-level features, but the learning ability of the shallow network is poor. In order to make the MobileNetV2 network pay more attention to the target that has a greater impact on the detection result, the attention mechanism is added before the feature extraction stage of the MobileNetV2 network framework.
[0049] If the CBAM attention mechanism is added before the feature extraction part, the channel attention mechanism and the spatial attention mechanism of the CBAM attention mechanism are added after the first convolution layer of the MobileNetV2 model. Specifically, the MobileNetV2 model based on the attention mechanism has the framework as shown in Figure 1 The beginning of the framework is a common convolution layer with a convolution kernel of 3, that is, the first convolution layer, which performs preliminary feature extraction on the input picture. Since the learning ability of the shallow network is poor at this time, the ability to pay attention to the target that has a greater impact on the final classification is weak. Therefore, in order to make the subsequent operation obtain more features that have a greater impact on the result, the CBAM attention mechanism is added after the first convolution layer, and the spatial attention and the channel attention are introduced into the network framework to improve the ability of the shallow network to capture key information. The output of the CBAM attention mechanism is sequentially subjected to the second convolution layer for preliminary dimension reduction, seven modules stacked with different numbers of block structures for sufficient feature extraction, and finally a common convolution layer for the last dimension adjustment of the extracted features. The adjusted data will be input into the classifier for fall detection.
[0050] The CBAM framework is shown in Figure 2 The overall network has a series structure. The CBAM attention mechanism takes the feature map F output by the upper convolution layer as input, first enters the channel attention module in order, obtains F' after the channel attention effect, and obtains F'' after the spatial attention module. The calculation formula is:
[0051] F′=MCA (F) (1)
[0052] F'' = M SA (F′) (2)
[0053] wherein, M CA represents a channel attention vector obtained after passing through a channel attention mechanism, M SA represents a spatial attention vector obtained after passing through a spatial attention mechanism.
[0054] The channel attention mechanism firstly simultaneously performs two pooling operations on the input, and obtains two vectors pool max and pool avg through maximum value pooling and average pooling respectively; then the two vectors are input into a weight-shared MLP (Multilayer Perceptron); after passing through the weight-shared network, the two vectors are added, and an activation function is used to activate them, and finally a channel attention correction vector M CA is obtained; M CA is multiplied with the original input image, and a feature map F′ after the channel attention effect is obtained.
[0055] The spatial attention mechanism performs maximum value pooling and average pooling on the input F′, and then concatenates according to the channel dimension, performs convolution through a 7x7 convolution kernel, and obtains a spatial attention vector M SA after passing through an activation function; M SA is multiplied with the input F′, and an output F'' after the spatial attention effect is obtained.
[0056] The addition of the CBAM attention mechanism can improve the attention degree of the MobileNetV2 network to the key target area at the beginning of learning, so as to promote the subsequent convolution operation to extract more key information, and thus improve the feature extraction efficiency of the MobileNetV2 network.
[0057] (2) The attention mechanism is added inside the block structure of the MobileNetV2.
[0058] In order to perfect the network structure of the feature extraction part and find the optimal position of adding the attention mechanism in the MobileNetV2 network in the application scene of fall detection, the block structure of the MobileNetV2 network is improved, and the attention mechanism is added inside the block.
[0059] The convolution part in the original block structure of the MobileNetV2 network is a depth separable convolution, which firstly performs convolution calculation on each channel through a convolution kernel of 3, and then fuses cross-channel information through point-by-point convolution with a convolution kernel of 1. However, the target affecting the classification result cannot improve the attention, so the original block structure is changed, and a new block structure fusing channel and spatial attention mechanisms is proposed.
[0060] As shown in Figure 3 , Figure 4 , the present application learns from the working principle of the CBAM attention mechanism, but changes the original attention mechanism module in series placement mode. Here, the channel attention and spatial attention mechanisms are placed independently and added to the block structure of MobileNetV2. That is, if the CBAM attention mechanism is added in the feature extraction part, the channel attention is added before the channel-by-channel convolution, and the spatial attention is added before the point-by-point convolution. If the step of the block is 1, the input of the block is added to the output of the point-by-point convolution to obtain the output of the block; if the step of the block is 2, the output of the point-by-point convolution is taken as the output of the block.
[0061] The detailed structure of the block after adding attention is shown in Figure 5 . After the data is input into the CBAM framework, two pooling operations are performed at the same time, and two vectors pool max and pool avg are obtained through maximum value pooling and average pooling respectively. Then the two vectors are input into the weight-shared MLP (Multilayer Perceptron), and after passing through the weight-shared network, the two vectors are added and activated by an activation function, and finally a channel attention correction vector M CA is obtained. CA M CA is multiplied by the original input image to obtain a feature map F CA after the channel attention effect. Then a 3x3 channel-by-channel convolution is used to convolve F CA , and each convolution kernel corresponds to a channel. After convolution, a feature map F1 is obtained.
[0062] F CA = M CA (input′) (3)
[0063] F1=Conv(F CA ) (4)
[0064] Then, spatial attention is applied to F1. The feature maps after max pooling and average pooling are concatenated along the channel dimension, and then convolved using a 7×7 convolution kernel. After passing through the activation function, the spatial attention vector M is obtained. SA M SA Multiplying by F1 yields the output F after spatial attention. SA Finally, pointwise convolution pairs with a 1×1 kernel are used to perform convolution on F. SA Perform convolution to fuse channel-dimensional features, and obtain the final output F of the pointwise convolution operation. out . Figure 5 In this context, K represents the kernel size, N represents the number of kernels, t represents the dimensionality factor, and H, W, and C represent the height, width, and number of channels of the input feature map, respectively.
[0065] F SA =M SA (F1) (5)
[0066] F out =Conv(F SA (6)
[0067] Since each position and channel of the input image has a different impact on the classification result, adding an attention mechanism assigns weights (degree of attention) based on the magnitude of the impact of each position on the detection result. This makes the network more efficient in processing features, enabling it to focus on key points and learn features more effectively.
[0068] (3) Classifier reconstruction.
[0069] The original MobileNetV2 classifier consisted of a single average pooling layer and a fully connected layer, with the output directly through the fully connected layer. Fall detection was abstracted into a binary classification problem. This approach caused the feature map dimension to drop drastically from 1×1×1280 to 1×1×2 in the original network framework. This rapid dimensionality reduction resulted in significant information loss and led to non-convergence during training. To address these issues, this invention reconstructs the model's classifier.
[0070] Classifier framework such as Figure 6 (The diagram shows the dimensionality change process using a batch size of 32 as an example.) As shown, the classifier module receives the features extracted by the feature extraction part. First, it adds an attention layer to correct the depth feature bias that may exist in the feature extraction part, so that the network focuses on key feature information. Then, it uses a pooling layer to reduce the dimensionality from 32×1280×7×7 to 32×1280×1×1. Subsequently, it performs a flattening operation to flatten the features for input into the fully connected layer. Finally, it goes through two fully connected layers to reduce the dimensionality in sequence, and obtains the binary classification result.
[0071] The processing of the classifier can solve the problem of excessive information loss caused by the sudden reduction of feature dimension in the original framework, so that the network can more effectively utilize the extracted feature information for fall detection.
[0072] The present application proposes two improvement ideas for MobileNetV2, and the overall method frameworks are as shown in Figure 7 、 Figure 8 . Figure 7 The model overall framework after adding CBAM attention mechanism to the network at the beginning, in addition to the newly built classifier, the present application adds CBAM attention mechanism before feature extraction in the block structure, to improve the learning ability of the shallow network, so that the subsequent layers can more efficiently extract features. Figure 8 The network model framework after modifying the network block structure. The network block structure is modified, and channel attention and spatial attention mechanisms are added, so that the block structure is more complete in function, and the network is more effective in feature extraction. The input and output of the two frameworks are completely consistent, the input is a frame of human motion image, and the improved network framework is processed, and two detection results: fall / no fall are output.
[0073] Since CBAM is a lightweight attention mechanism, the parameter amount does not increase significantly after introducing it into the MobileNetV2 network. The two improvement schemes proposed in the present application improve the network function of the algorithm while keeping the parameter amount of the network framework small, so that the algorithm can pay more attention to the target that has a greater impact on the final detection result during the working process, thereby achieving higher performance.
[0074] With the intensification of the current social aging problem, the health problems of the elderly living alone have become the focus of the medical and health field. Since the current fall detection algorithm based on deep learning has a large network parameter amount, it requires high training equipment and takes a long time to train. In order to make the algorithm meet the requirements of lightweight and high accuracy at the same time, the present application proposes two improvement schemes for MobileNetV2: adding CBAM attention mechanism before the network block structure framework; adding channel attention mechanism and spatial attention mechanism inside the block. Both schemes improve the capture rate of key information without increasing the additional calculation amount of the network, thereby improving the feature extraction effectiveness of the network and further improving the network detection performance.
[0075] Embodiment Two
[0076] The present embodiment provides a fall detection system based on MobileNetV2 model, which specifically comprises the following modules:
[0077] an image acquisition module configured to acquire a human motion image to be detected;
[0078] a fall detection module configured to input the human motion image into a MobileNetV2 model based on an attention mechanism to obtain a fall detection result;
[0079] The MobileNetV2 model based on the attention mechanism adds a channel attention mechanism and a spatial attention mechanism of the CBAM attention mechanism before or in a feature extraction part of a MobileNetV2 network, and sequentially connects an attention layer, a pooling layer, a flattening operation and a fully connected layer after the feature extraction part.
[0080] If the CBAM attention mechanism is added in the feature extraction part, the channel attention mechanism is added before a channel-by-channel convolution, and the spatial attention mechanism is added before a point-by-point convolution.
[0081] If the CBAM attention mechanism is added before the feature extraction part, the channel attention mechanism and the spatial attention mechanism of the CBAM attention mechanism are added after a first convolution layer of the MobileNetV2 model.
[0082] It should be noted that each module in the embodiment corresponds to each step in Embodiment One, and the specific implementation process is the same, which will not be repeated here.
[0083] Embodiment Three
[0084] The embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the steps in the fall detection method based on the MobileNetV2 model according to Embodiment One.
[0085] Embodiment Four
[0086] The embodiment provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor realizes the steps in the fall detection method based on the MobileNetV2 model according to Embodiment One when executing the program.
[0087] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer usable program code.
[0088] The embodiments of methods, apparatuses (systems) and computer program products according to the present application can be described in the general context of method steps and processes, which can be implemented in one embodiment by a program of instructions being executed in a computer system on a machine such as the computer systems 102 and 202. The program is preferably implemented in program code that can be applied to input data to perform the functions described and illustrated above, and then generated as suitable output. The program code, when executed, can be used to implement the processes and other functionality as discussed herein. The program code can be stored in a storage device, such as the system or volatile memory 206 and / or the non-volatile storage 208. Alternatively, the program code can be stored on a computer readable medium such as a floppy disk, a hard disk, a CD ROM, a DVD, a memory stick, or an optical disk stored in the storage device. The program code, when executed, can be used to implement the processes and other functionality as discussed herein. Figure 1 one or more functions specified in the flow or flows and / or blocks. Figure 1 one or more functions specified in the flow or flows and / or blocks.
[0089] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more functions specified in the flow or flows and / or blocks. Figure 1 one or more functions specified in the flow or flows and / or blocks.
[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more functions specified in the flow or flows and / or blocks. Figure 1 one or more functions specified in the flow or flows and / or blocks.
[0091] Those skilled in the art can understand that all or part of the flow of the above-mentioned embodiment method can be realized by computer program instructions instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the flow of the above-mentioned embodiment method. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.
[0092] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A fall detection method based on the MobileNetV2 model, characterized in that, include: Acquire images of the human motion to be detected; Human motion images are input into the attention-based MobileNetV2 model to obtain fall detection results; Among them, the MobileNetV2 model based on the attention mechanism adds the channel attention mechanism and spatial attention mechanism of CBAM attention mechanism before or in the feature extraction part of the MobileNetV2 network, and connects the attention layer, pooling layer, flattening operation and fully connected layer in sequence after the feature extraction part; Channel attention and spatial attention mechanisms are placed independently and added separately to the block structure of MobileNetV2. If the CBAM attention mechanism is added in the feature extraction part, the channel attention mechanism is added before the channel-wise convolution, and the spatial attention mechanism is added before the pointwise convolution. If the stride of the block is 1, the block input is added to the output of the pointwise convolution to obtain the block output. If the stride of the block is 2, the output of the pointwise convolution is used as the block output. After the data is input into the CBAM framework, two pooling operations are performed simultaneously, obtaining two vectors through max pooling and average pooling respectively. , The two vectors are then input into a weight-sharing MLP. After passing through the weight-sharing network, the two vectors are added together, and an activation function is applied to activate them, ultimately yielding the channel attention correction vector. ,Will Multiplying the original input image by the channel attention function yields the feature map. ; Then, a 3×3 channel-wise convolution pair was used. Perform a convolution operation, with each convolution kernel corresponding to one channel, and obtain a feature map after convolution. Then on Spatial attention is performed by concatenating the feature maps after max pooling and average pooling along the channel dimension, then convolving them using a 7×7 convolution kernel, and finally passing the convolution through an activation function to obtain the spatial attention vector. ,Will and Multiplying them together yields the output after spatial attention. ; Finally, pointwise convolution pairs with 1×1 kernels are used. Perform convolution to fuse channel-dimensional features, and obtain the final output of the pointwise convolution operation. The classifier module receives the features extracted by the feature extraction part. First, it adds an attention layer to correct the depth feature bias in the feature extraction part, so that the network focuses on key feature information. Then, it uses a pooling layer to reduce the dimension from 32×1280×7×7 to 32×1280×1×1. Then, it performs a flattening operation to flatten the features for input to the fully connected layer. Finally, it passes through two fully connected layers to reduce the dimension and obtain the binary classification result. If the CBAM attention mechanism is added before the feature extraction part, then the channel attention mechanism and spatial attention mechanism of the CBAM attention mechanism are added after the first convolutional layer of the MobileNetV2 model.
2. The fall detection method based on the MobileNetV2 model as described in claim 1, characterized in that, The channel attention mechanism first performs max pooling and average pooling on the input simultaneously, resulting in two vectors; then, the two vectors are input into a weight-sharing network. After passing through a weight-sharing network, the two vectors are added together, and an activation function is used to activate them, resulting in a channel attention correction vector. Multiply the channel attention correction vector by the input to obtain the feature map after channel attention.
3. The fall detection method based on the MobileNetV2 model as described in claim 1, characterized in that, The spatial attention mechanism performs max pooling and average pooling on the input, concatenates them according to the channel dimension, performs convolution using a convolution kernel, and obtains a spatial attention vector after passing through an activation function. The spatial attention vector is then multiplied by the input to obtain the output after the spatial attention process.
4. A fall detection system based on the MobileNetV2 model, characterized in that, include: The image acquisition module is configured to acquire images of human motion to be detected. The fall detection module is configured to input human motion images into an attention-based MobileNetV2 model to obtain fall detection results. Among them, the MobileNetV2 model based on the attention mechanism adds the channel attention mechanism and spatial attention mechanism of CBAM attention mechanism before or in the feature extraction part of the MobileNetV2 network, and connects the attention layer, pooling layer, flattening operation and fully connected layer in sequence after the feature extraction part; Channel attention and spatial attention mechanisms are placed independently and added separately to the block structure of MobileNetV2. If the CBAM attention mechanism is added in the feature extraction part, the channel attention mechanism is added before the channel-wise convolution, and the spatial attention mechanism is added before the pointwise convolution. If the stride of the block is 1, the block input is added to the output of the pointwise convolution to obtain the block output. If the stride of the block is 2, the output of the pointwise convolution is used as the block output. After the data is input into the CBAM framework, two pooling operations are performed simultaneously, obtaining two vectors through max pooling and average pooling respectively. , The two vectors are then input into a weight-sharing MLP. After passing through the weight-sharing network, the two vectors are added together, and an activation function is applied to activate them, ultimately yielding the channel attention correction vector. ,Will Multiplying the original input image by the channel attention function yields the feature map. ; Then, a 3×3 channel-wise convolution pair was used. Perform a convolution operation, with each convolution kernel corresponding to one channel, and obtain a feature map after convolution. Then on Spatial attention is performed by concatenating the feature maps after max pooling and average pooling along the channel dimension, then convolving them using a 7×7 convolution kernel, and finally passing the convolution through an activation function to obtain the spatial attention vector. ,Will and Multiplying them together yields the output after spatial attention. ; Finally, pointwise convolution pairs with 1×1 kernels are used. Perform convolution to fuse channel-dimensional features, and obtain the final output of the pointwise convolution operation. The classifier module receives the features extracted by the feature extraction part. First, it adds an attention layer to correct the depth feature bias in the feature extraction part, so that the network focuses on key feature information. Then, it uses a pooling layer to reduce the dimension from 32×1280×7×7 to 32×1280×1×1. Then, it performs a flattening operation to flatten the features for input to the fully connected layer. Finally, it passes through two fully connected layers to reduce the dimension and obtain the binary classification result. If the CBAM attention mechanism is added before the feature extraction part, then the channel attention mechanism and spatial attention mechanism of the CBAM attention mechanism are added after the first convolutional layer of the MobileNetV2 model.
5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the fall detection method based on the MobileNetV2 model as described in any one of claims 1-3.
6. 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 program, it implements the steps in the fall detection method based on the MobileNetV2 model as described in any one of claims 1-3.