Human body tumble detection method based on deep learning
By building a WA-transformer network, combining 2D lidar and millimeter wave radar data, the existing fall detection technology has solved the problems in complex environments and diversified action recognition, and achieved real-time and personalized fall detection effects.
Patent Information
- Application Number
- CN202510136460.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-07
- Publication Date
- 2025-04-11
AI Technical Summary
The existing fall detection technology has many technical difficulties in complex environments, diversified action recognition, real-time processing and personalized needs.
Using deep learning-based human fall detection method, data is obtained by using 2D lidar and millimeter wave radar, data fusion, feature embedding and position encoding are carried out, WA-transformer network is built for training, and fall posture is detected in real time and alarm signals are sent.
The recognition and real-time processing of diversified actions are realized in complex environments, meeting personalized needs and improving the accuracy and efficiency of fall detection.
Smart Images
Figure CN119810924A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of sensor radar technology, and in particular to a human fall detection method based on deep learning. Background Art
[0002] As China's population aging continues to intensify, my country is experiencing an important stage of transformation from a populous country to an aging society. Population aging has become a basic problem that my country will face for a long period of time in the future. According to the Ministry of Human Resources and Social Security, the number of people aged 60 and above in China will exceed 300 million, indicating that we are about to enter a moderately aging society. This trend is characterized by large scale, rapid growth, and obvious aging trend, which will have an all-round impact on my country's industrial structure, economic structure, labor force structure, social structure and other aspects. In the face of the challenges brought by the aging population, the care of the elderly has become a problem that the society needs to solve urgently.
[0003] At present, the existing technology generally uses accelerometers and gyroscopes. These sensors can monitor the user's movement status in real time and identify fall events by analyzing the acceleration and angular velocity data, but they rely on the user to wear them. If the user does not wear the device, the detection effect will be affected; or the ground pressure sensor is installed under the floor to monitor the user's gait and position changes, but the installation is complicated and multiple sensors need to be installed in the home or monitoring area; or image processing and computer vision technology are used to analyze the user's posture and movement pattern to identify fall events, but in an environment with insufficient or excessive light, the detection effect may be affected. Summary of the invention
[0004] Purpose of the invention: The purpose of the present invention is to provide a human fall detection method based on deep learning, which solves many technical problems of traditional fall detection technology in complex environments, diversified motion recognition, real-time processing and personalized needs.
[0005] Technical solution: A human fall detection method based on deep learning described in the present invention comprises the following steps: (1) Use 2D lidar and millimeter-wave radar to obtain lidar data and millimeter-wave radar data and perform preprocessing; (2) Fusing the preprocessed lidar data and millimeter-wave radar data; (3) Convert the fused data into a data representation; (4) Feature embedding and position encoding of the data obtained in step (3); (5) Construct the WA-transformer network and input the data obtained in step (4) into the network for training; (6) Detect falling posture and send alarm signals in real time.
[0006] Furthermore, step (1) includes the following steps: (11) For LiDAR data: collect point cloud data in three-dimensional space, where each point contains spatial coordinates and reflection intensity; for millimeter-wave radar data: collect target lists, including distance, angle, speed, and reflection intensity; (12) Use a synchronous trigger signal to ensure that the sensors start collecting data at the same time; based on the timestamps of each sensor, use an interpolation or alignment algorithm to achieve time synchronization of the data; (13) Determine the relative position and attitude of the laser radar and millimeter wave radar in the spatial coordinate system and calculate the external parameter matrix; (14) Data cleaning for lidar data includes noise removal and downsampling; data cleaning for millimeter wave radar data includes clutter removal and signal enhancement.
[0007] Furthermore, step (2) includes the following steps: (21) Convert the millimeter-wave radar data to the lidar coordinate system as follows: For each target of the millimeter-wave radar, convert the polar coordinates to Cartesian coordinates and apply the extrinsic matrix: ; in, , , is the Cartesian coordinate in the millimeter wave radar coordinate system; The target distance measured by the millimeter-wave radar; is the target angle measured by the millimeter-wave radar; ; Then, using the external parameter matrix That is, the transformation matrix from the millimeter-wave radar coordinate system to the lidar coordinate system, transforming the points in the millimeter-wave radar coordinate system to the lidar coordinate system and applying the external parameter matrix: ; is the external parameter matrix, which represents the transformation relationship from the millimeter wave radar coordinate system to the lidar coordinate system; homogeneous coordinates mean adding a 1 at the end of the coordinate vector to facilitate matrix multiplication and unify rotation and translation operations; (22) The converted data is subjected to feature fusion to obtain the feature vector , the formula is as follows: ; in, : The three-dimensional spatial coordinates of the lidar point, is the reflection intensity of the laser radar point, To correspond to the speed information of the millimeter wave radar, is the signal strength or reflection intensity of the corresponding millimeter wave radar.
[0008] Furthermore, step (3) includes the following steps: (31) Convert the point cloud data into sequence form, the formula is as follows: Input sequence ; in, Indicates The feature vector of a point, is the total number of points in the point cloud, and the feature vectors of all points are arranged in order to form a length of sequence; (32) Normalize each feature dimension to the same scale, the formula is as follows: ; in, It is the standardized feature vectors, is the original i-th eigenvector, is the mean vector of the features, the mean of each dimension, is the standard deviation vector of the features, with the standard deviation in each dimension.
[0009] Furthermore, step (4) includes the following steps: (41) The standardized feature vector is mapped to a high-dimensional space through a linear layer, as follows: ; in, It is The embedding vector of a point, is a learnable weight matrix; is a learnable bias vector with dimension D; It is the standardized feature vectors; (42) Encode the three-dimensional spatial coordinates of the laser radar point by sine and cosine position; where for the coordinates , , The position encoding is: ; ; ; in, It is Point Positional encoding components; It is Points in The coordinates of the axes; is the total dimension of the position encoding; The value is 0, 1, 2, ..., -1; Will The encoding results are concatenated to form a complete position encoding vector : ; (43) Add the feature embedding and position encoding to form the input sequence, as follows: ; in, It is vectors input to the Transformer, It is feature embedding vector, It is Position encoding vector.
[0010] Furthermore, step (5) is as follows: based on the transformer network, the GCN module and discrete cosine transform DCT are introduced to improve the backbone and neck; then Mixup data enhancement is used in the training stage; finally, the positioning loss function CIoU is replaced by SioU.
[0011] Furthermore, the WA-transformer network includes: a graph convolutional network (GCN), a spatial Transformer block, a hybrid Transformer block, and a discrete cosine transform DCT in sequence; wherein the graph convolutional network GCN captures the dependency between nodes by defining the connection relationship between nodes; the spatial Transformer block is used to combine the self-attention mechanism and the multi-layer perceptron to enable adaptive aggregation and transformation of features, and stabilize the training process through residual connections and LayerNorm; the hybrid Transformer block is used to divide the features into two parts, and apply Mlp and FreqMlp respectively to fuse the time domain and frequency domain information; multi-scale features are fused by segmenting, processing and splicing the features; the discrete cosine transform DCT is used to convert the joint motion data in the time domain to the frequency domain to extract key frequency features.
[0012] Furthermore, the SioU loss function is as follows: it consists of angle cost, distance cost, shape cost, and IoU cost; the formula is as follows: SIoU_Loss=Angle cost+Distance cost+Shape cost+IoU cost Among them, Angle cos is the angle cost, which measures the difference in rotation angle between the predicted bounding box and the true bounding box; Distance cost is the distance cost, which measures the spatial distance between the center point of the predicted bounding box and the center point of the true bounding box; Shape cos is the shape cost, which measures the difference in size between the predicted bounding box and the true bounding box; IoU cost is the IoU cost based on the intersection-over-union ratio, which measures the degree of overlap between the predicted bounding box and the true bounding box.
[0013] An electronic device described in the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, any one of the methods for detecting a human fall based on deep learning is implemented.
[0014] A storage medium described in the present invention stores a computer program, and when the computer program is executed by a processor, any one of the methods for detecting a human fall based on deep learning is implemented.
[0015] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: The introduction of discrete cosine transform (DCT) and inverse discrete cosine transform (IDCT) to process data in the frequency domain helps the model capture the frequency characteristics of the input signal and enhances the modeling ability of time series information. The input features are divided into two parts, which are processed by a standard multi-layer perceptron (MLP1) and a multi-layer perceptron (MLP2) in the frequency domain. This hybrid processing method enables the model to capture spatial and frequency information at the same time. By fusing the graph convolutional network and the self-attention mechanism, mixer_small (GCN) and mixer_large (attention mechanism) are used to process data respectively, thereby capturing the local relationship and global dependency between joints and improving the model's ability to understand complex postures. A self-attention mechanism for four-dimensional input tensors is implemented, which is suitable for the processing of spatiotemporal features. This customized self-attention module can more effectively process data formats unique to posture estimation tasks. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 is a flow chart of the present invention; Figure 2 It is the model principle diagram of the present invention. DETAILED DESCRIPTION
[0017] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0018] like Figure 1 As shown, an embodiment of the present invention provides a human fall detection method based on deep learning, comprising the following steps: (1) Using 2D laser radar and millimeter wave radar to obtain laser radar data and millimeter wave radar data, and preprocessing them; including the following steps: (11) For LiDAR data: collect point cloud data in three-dimensional space, where each point contains spatial coordinates and reflection intensity; for millimeter-wave radar data: collect target lists, including distance, angle, speed, and reflection intensity; (12) Use a synchronous trigger signal to ensure that the sensors start collecting data at the same time; based on the timestamps of each sensor, use an interpolation or alignment algorithm to achieve time synchronization of the data; (13) Determine the relative position and attitude of the laser radar and millimeter wave radar in the spatial coordinate system and calculate the external parameter matrix; (14) Data cleaning for lidar data includes noise removal and downsampling; data cleaning for millimeter wave radar data includes clutter removal and signal enhancement.
[0019] (2) Fusing the preprocessed lidar data and millimeter-wave radar data; including the following steps: (21) Convert the millimeter-wave radar data to the lidar coordinate system as follows: For each target of the millimeter-wave radar, convert the polar coordinates to Cartesian coordinates and apply the extrinsic matrix: ; in, , , is the Cartesian coordinate in the millimeter wave radar coordinate system; The target distance measured by the millimeter-wave radar; is the target angle measured by the millimeter-wave radar; ; Then, using the external parameter matrix That is, the transformation matrix from the millimeter-wave radar coordinate system to the lidar coordinate system, transforming the points in the millimeter-wave radar coordinate system to the lidar coordinate system and applying the external parameter matrix: ; is the external parameter matrix, which represents the transformation relationship from the millimeter wave radar coordinate system to the lidar coordinate system; homogeneous coordinates mean adding a 1 at the end of the coordinate vector to facilitate matrix multiplication and unify rotation and translation operations; (22) The converted data is subjected to feature fusion to obtain the feature vector , the formula is as follows: ; in, : The three-dimensional spatial coordinates of the lidar point, is the reflection intensity of the laser radar point, To correspond to the speed information of the millimeter wave radar, is the signal strength or reflection intensity of the corresponding millimeter wave radar.
[0020] (3) Convert the fused data into a data representation; this includes the following steps: (31) Convert the point cloud data into sequence form, the formula is as follows: Input sequence ; in, Indicates The feature vector of a point, is the total number of points in the point cloud, and the feature vectors of all points are arranged in order to form a length of sequence; (32) Normalize each feature dimension to the same scale, the formula is as follows: ; in, It is the standardized feature vectors, is the original i-th eigenvector, is the mean vector of the features, the mean of each dimension, is the standard deviation vector of the features, with the standard deviation in each dimension.
[0021] (4) Feature embedding and position encoding of the data obtained in step (3); including the following steps: The following steps are involved: (41) The standardized feature vector is mapped to a high-dimensional space through a linear layer, as follows: ; in, It is The embedding vector of each point, is a learnable weight matrix; is a learnable bias vector with dimension D; It is the standardized feature vectors; (42) Encode the three-dimensional spatial coordinates of the laser radar point by sine and cosine position; where for the coordinates , , The position encoding is: ; ; ; in, It is Point Positional encoding components; It is Points in The coordinates of the axes; is the total dimension of the position encoding; The value is 0, 1, 2, ..., -1; Will The encoding results are concatenated to form a complete position encoding vector : ; (43) Add the feature embedding and position encoding to form the input sequence, as follows: ; in, It is vectors input to the Transformer, It is feature embedding vector, It is Position encoding vector.
[0022] (5) Construct the WA-transformer network and input the data obtained in step (4) into the network for training; Figure 2 As shown in the figure, the details are as follows: based on the transformer network, the GCN module and discrete cosine transform DCT are introduced to improve the backbone and neck; then Mixup data enhancement is used in the training stage; finally, the positioning loss function CIoU is replaced by SioU.
[0023] The WA-transformer network includes: graph convolutional network (GCN), spatial Transformer block, hybrid Transformer block, discrete cosine transform DCT; among them, the graph convolutional network GCN captures the dependency between nodes by defining the connection relationship between nodes; the spatial Transformer block is used to combine the self-attention mechanism and the multi-layer perceptron to enable adaptive aggregation and transformation of features, and stabilize the training process through residual connection and LayerNorm; the hybrid Transformer block is used to divide the features into two parts, and apply Mlp and FreqMlp respectively to fuse time domain and frequency domain information; multi-scale features are fused by segmenting, processing and splicing features; discrete cosine transform DCT is used to convert joint motion data in the time domain to the frequency domain to extract key frequency features. The specific process is as follows: S1. Graph Convolutional Network (GCN) Location: Located in the first half of the model, it defines the specific implementation of the graph convolutional network.
[0024] Function: Graph structure modeling: Capture the dependencies between joints by defining the connection relationships (spatial or temporal patterns) between nodes.
[0025] Feature aggregation: Use graph convolution operations to aggregate the features of neighboring nodes and enhance node feature representation.
[0026] Key components: Adjacency Matrix: defines the connection relationship between nodes. For spatial patterns, predefined joint connections are used; for temporal patterns, similarity or fixed connection lengths are used.
[0027] Linear transformation (U and V): used for linear mapping of features.
[0028] Normalization: Normalize the adjacency matrix to stabilize the training process.
[0029] Workflow: Input processing: Depending on the mode (spatial or temporal), the shape of the input tensor is reshaped.
[0030] Similarity calculation (temporal mode only): If similarity is used, calculate the similarity between nodes and generate an adjacency matrix.
[0031] Feature aggregation: Features are weighted aggregated through the normalized adjacency matrix.
[0032] Non-linear activation: Apply the ReLU activation function to enhance feature expression capabilities.
[0033] Multilayer Perceptron (Mlp): Non-linear transformation: Through two layers of linear transformation and activation function (GELU), the input features are non-linearly transformed.
[0034] Feature expansion and compression: can be used to increase or decrease feature dimensions.
[0035] Frequency Domain Multilayer Perceptron (FreqMlp): Frequency domain feature processing: Features are processed in the frequency domain through discrete cosine transform (DCT) and inverse discrete cosine transform (IDCT).
[0036] Feature enhancement: Capture frequency information in time series and extract important motion patterns.
[0037] S2, Spatial Transformer Block Function: Standard Transformer layer: combines the self-attention mechanism and the multi-layer perceptron to achieve adaptive aggregation and transformation of features.
[0038] Residual connection and layer normalization: Through residual connection and LayerNorm, the training process is stabilized to prevent gradient disappearance.
[0039] Key components: LayerNorm: Normalize the features.
[0040] selfAttention: Self-attention module.
[0041] Mlp: Multi-layer perceptron module.
[0042] DropPath: Stochastic Depth regularization to prevent overfitting.
[0043] Workflow: Self-attention processing: Process input features through LayerNorm and self-attention mechanism. Global dependency modeling: Capture the dependency between any positions in the sequence through the self-attention mechanism.
[0044] Feature enhancement: According to the attention weight, important features are enhanced and irrelevant features are suppressed.
[0045] Residual connection: add the self-attention output to the input.
[0046] MLP processing: Features are further processed by LayerNorm and MLP.
[0047] Residual connection: Add the MLP output to the features after self-attention processing.
[0048] S3, MixedBlock Function: Hybrid feature processing: Divide the features into two parts, apply Mlp and FreqMlp respectively, and fuse the time domain and frequency domain information.
[0049] Feature separation and splicing: Multi-scale feature fusion is achieved by segmenting, processing and splicing features.
[0050] Workflow: Self-attention processing: Similar to the standard Transformer block.
[0051] Feature segmentation: Divide the features into two parts along the time dimension or other dimensions.
[0052] Independent processing: A portion is processed through Mlp.
[0053] The other part is processed by FreqMlp.
[0054] Feature splicing: Splice the two processed features in the feature dimension.
[0055] Mixer: Select Attention, GCN, or MultiScaleTCN as the feature mixer according to the configuration.
[0056] LayerNorm: Normalize the features.
[0057] MLP: Multi-layer Perceptron module.
[0058] DropPath and Layer Scale: further stabilize the training process and enhance feature expression.
[0059] Workflow: Feature normalization: Normalize the input features through LayerNorm.
[0060] Feature mixing: Process features through the selected Mixer (Attention, GCN).
[0061] Residual connection: Add the mixed features to the input.
[0062] MLP processing: The features are further processed by MLP.
[0063] Residual connection: Add the MLP output to the mixed features.
[0064] S4. Discrete cosine transform (DCT) is mainly used to convert joint motion data in the time domain to the frequency domain to extract key frequency features. This process achieves dimensionality reduction and selection of features by selecting and retaining important low-frequency coefficients, thereby reducing computational complexity and improving the operating efficiency of the model. Frequency domain features help to identify and understand periodic and dynamic patterns in motion, and enhance the model's ability to recognize motion patterns. At the same time, DCT has a certain inhibitory effect on high-frequency noise, which improves the robustness and stability of the model. By effectively fusing the features processed in the frequency domain with the spatial features, DCT promotes the comprehensive expression of spatiotemporal features, ultimately improving the accuracy and overall performance of posture estimation.
[0065] The SioU loss function is as follows: It consists of angle cost, distance cost, shape cost, and IoU cost; the formula is as follows: SIoU_Loss=Angle cost+Distance cost+Shape cost+IoU cost Among them, Angle cos is the angle cost, which measures the difference in rotation angle between the predicted bounding box and the true bounding box; Distance cost is the distance cost, which measures the spatial distance between the center point of the predicted bounding box and the center point of the true bounding box; Shape cos is the shape cost, which measures the difference in size between the predicted bounding box and the true bounding box; IoU cost is the IoU cost based on the intersection-over-union ratio, which measures the degree of overlap between the predicted bounding box and the true bounding box.
[0066] (6) Detect falling posture and send alarm signals in real time.
[0067] An embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the computer program implements any one of the deep learning-based human fall detection methods described above.
[0068] An embodiment of the present invention further provides a storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, any one of the methods for detecting a human fall based on deep learning is implemented.
Claims
1. A human fall detection method based on deep learning, characterized in that: The following steps are involved: (1) Use 2D lidar and millimeter-wave radar to obtain lidar data and millimeter-wave radar data and perform preprocessing; (2) Fusing the preprocessed lidar data and millimeter-wave radar data; (3) Convert the fused data into a data representation; (4) Feature embedding and position encoding of the data obtained in step (3); (5) Construct the WA-transformer network and input the data obtained in step (4) into the network for training; (6) Detect falling posture and send alarm signals in real time.
2. According to the method for human fall detection based on deep learning in claim 1, it is characterized in that: Step (1) includes the following steps: (11) For LiDAR data: collect point cloud data in three-dimensional space, where each point contains spatial coordinates and reflection intensity; for millimeter-wave radar data: collect target lists, including distance, angle, speed, and reflection intensity; (12) Use a synchronous trigger signal to ensure that the sensors start collecting data at the same time; based on the timestamps of each sensor, use an interpolation or alignment algorithm to achieve time synchronization of the data; (13) Determine the relative position and attitude of the laser radar and millimeter wave radar in the spatial coordinate system and calculate the external parameter matrix; (14) Data cleaning for lidar data includes noise removal and downsampling; data cleaning for millimeter wave radar data includes clutter removal and signal enhancement.
3. The method for human fall detection based on deep learning according to claim 1, characterized in that: Step (2) includes the following steps: (21) Convert the millimeter-wave radar data to the lidar coordinate system as follows: For each target of the millimeter-wave radar, convert the polar coordinates to Cartesian coordinates and apply the extrinsic matrix: ； in, , , is the Cartesian coordinate in the millimeter wave radar coordinate system; The target distance measured by the millimeter-wave radar; is the target angle measured by the millimeter-wave radar; ; Then, using the external parameter matrix That is, the transformation matrix from the millimeter-wave radar coordinate system to the lidar coordinate system, transforming the points in the millimeter-wave radar coordinate system to the lidar coordinate system and applying the external parameter matrix: ； is the external parameter matrix, which represents the transformation relationship from the millimeter wave radar coordinate system to the lidar coordinate system; homogeneous coordinates mean adding a 1 at the end of the coordinate vector to facilitate matrix multiplication and unify rotation and translation operations; (22) The converted data is subjected to feature fusion to obtain the feature vector , the formula is as follows: ； in, : The three-dimensional spatial coordinates of the lidar point, is the reflection intensity of the laser radar point, To correspond to the speed information of the millimeter wave radar, is the signal strength or reflection intensity of the corresponding millimeter wave radar.
4. The method for human fall detection based on deep learning according to claim 1, characterized in that: Step (3) includes the following steps: (31) Convert the point cloud data into sequence form, the formula is as follows: Input sequence ; in, Indicates The feature vector of a point, is the total number of points in the point cloud, and the feature vectors of all points are arranged in order to form a length of sequence; (32) Normalize each feature dimension to the same scale, the formula is as follows: ； in, It is the standardized feature vectors, is the original i-th eigenvector, is the mean vector of the features, the mean of each dimension, is the standard deviation vector of the features, with the standard deviation in each dimension.
5. The method for human fall detection based on deep learning according to claim 1, characterized in that: Step (4) includes the following steps: (41) The standardized feature vector is mapped to a high-dimensional space through a linear layer, as follows: ； in, It is The embedding vector of each point, is a learnable weight matrix; is a learnable bias vector with dimension D; It is the standardized feature vectors; (42) Encode the three-dimensional spatial coordinates of the laser radar point by sine and cosine position; where for the coordinates , , The position encoding is: ； ； ； in, It is Point Positional encoding components; It is Points in The coordinates of the axes; is the total dimension of the position encoding; The value is 0, 1, 2, ..., -1; Will The encoding results are concatenated to form a complete position encoding vector : ； (43) Add the feature embedding and position encoding to form the input sequence, as follows: ； in, It is vectors input to the Transformer, It is feature embedding vector, It is Position encoding vector.
6. The method for human fall detection based on deep learning according to claim 1, characterized in that: Step (5) is as follows: Based on the transformer network, the GCN module and discrete cosine transform DCT are introduced to improve the backbone and neck; then Mixup data enhancement is used in the training stage; finally, the positioning loss function CIoU is replaced by SioU.
7. A human fall detection method based on deep learning according to claim 6, characterized in that: The WA-transformer network includes: graph convolutional network (GCN), spatial Transformer block, hybrid Transformer block, and discrete cosine transform DCT in sequence; among them, the graph convolutional network GCN captures the dependency between nodes by defining the connection relationship between nodes; the spatial Transformer block is used to combine the self-attention mechanism and the multi-layer perceptron to enable adaptive aggregation and transformation of features, and stabilize the training process through residual connection and LayerNorm; the hybrid Transformer block is used to divide the features into two parts, and apply Mlp and FreqMlp respectively to fuse time domain and frequency domain information; multi-scale features are fused by segmenting, processing and splicing features; discrete cosine transform DCT is used to convert joint motion data in the time domain to the frequency domain to extract key frequency features.
8. The method for human fall detection based on deep learning according to claim 6, characterized in that: The SioU loss function is as follows: It consists of angle cost, distance cost, shape cost, and IoU cost; the formula is as follows: SIoU_Loss=Angle cost+Distance cost+Shape cost+IoU cost Among them, Angle cos is the angle cost, which measures the difference in rotation angle between the predicted bounding box and the true bounding box; Distance cost is the distance cost, which measures the spatial distance between the center point of the predicted bounding box and the center point of the true bounding box; Shape cos is the shape cost, which measures the difference in size between the predicted bounding box and the true bounding box; IoU cost is the IoU cost based on the intersection-over-union ratio, which measures the degree of overlap between the predicted bounding box and the true bounding box.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into a processor, a human fall detection method based on deep learning according to any one of claims 1 to 8 is implemented.
10. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, a human fall detection method based on deep learning according to any one of claims 1 to 8 is implemented.