Fall detection method based on OEF-YOLO model

Abstract

The invention discloses a fall detection method based on an OEF-YOLO model, and is applied to the technical field of computer vision. Comprising the following steps: constructing a data set for training and detecting the performance of an OEF-YOLO model; the YOLOv8 model is improved, and an OEF-YOLO model used for tumble detection is obtained; s2, training an OEF-YOLO model through the data set constructed in the step S1; and the detection effect of the OEF-YOLO model is evaluated through the data set constructed in the step S1. According to the method, the YOLOv8 model is improved through the EMA attention mechanism module, the C2f-ODConv module and the loss function integrated with the Focal Loss, the reliability of fall detection in the indoor scene is improved, and reasonable balance is achieved in the aspects of light weight and high precision.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically to a fall detection method based on the OEF-YOLO model. Background Technology

[0002] With the accelerating pace of aging and the increasing size of the elderly population, the phenomena of "empty nest syndrome" and "population cliff" are becoming increasingly apparent. Many elderly people face numerous risks and hidden dangers in their lives due to declining physical fitness, insufficient balance, first aid skills, and reaction time, as well as complex environments and chronic illnesses. If any inconvenience experienced by the elderly is not effectively addressed and promptly assisted, even minor accidents can have a "magnifying effect," leading to serious consequences and even endangering their lives. Falls are a significant threat to the health and safety of the elderly in their daily lives, not only easily causing serious physical injuries but also triggering other more serious complications.

[0003] Given the significant harm falls cause in daily life, and the fact that the systematic development of elderly care is not yet fully mature, designing a reliable fall detection method to promptly detect falls and minimize their adverse effects is essential for improving the quality of life for vulnerable groups, particularly the elderly. Fall detection can be implemented using wearable sensors or environmental sensors. While these methods offer advantages in data acquisition and detection accuracy, they also have significant drawbacks: they rely on wearable and environmental sensor hardware, limiting application scenarios and restricting user movement. In contrast, fall detection algorithms based on computer vision or deep learning only require image acquisition from a camera, do not restrict user movement, are highly scalable, easily extend to other application scenarios, and exhibit excellent detection accuracy.

[0004] Deep learning-based fall detection methods are currently the mainstream approach. Deep learning algorithms construct deep networks, learn from data, and use backpropagation to adjust network parameters and correct errors, ultimately obtaining the optimal network output. However, existing fall detection methods still have limitations in terms of robustness, real-time performance, and accuracy. Specifically: fall detection accuracy still has room for improvement; the model architecture is too complex, resulting in a large number of network parameters and computational overhead, making it difficult to balance detection performance and lightweight deployment; and the detector's performance or stability remains to be verified under changing environmental conditions (including changes in brightness, angle changes, and partial occlusion of the human body). Therefore, how to provide a fall detection method based on the OEF-YOLO model is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a fall detection method based on the OEF-YOLO model, which improves the performance of the fall detection algorithm through design improvements, and achieves a balance between network lightweighting and detection performance.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A fall detection method based on the OEF-YOLO model includes the following steps:

[0008] S1. Construct a dataset for training and testing the performance of the OEF-YOLO model;

[0009] S2. Improve the YOLOv8 model to obtain the OEF-YOLO model for fall detection;

[0010] S3. Train the OEF-YOLO model using the dataset constructed in S1;

[0011] S4. Evaluate the detection performance of the OEF-YOLO model using the dataset constructed in S1.

[0012] Optionally, S1 is as follows:

[0013] The dataset, which contains fall behavior and daily life behaviors, was manually labeled and four categories were set: standing, bending over, sitting, and falling. The images in the dataset include people of different heights and body types, which helps the model learn the features of various human body shapes. The images are taken from multiple angles, which helps the model learn the appearance features of the human body from different angles. The dataset also includes images with varying lighting and partial occlusion, which allows the model to better adapt to different lighting conditions and environments with partial occlusion.

[0014] Optionally, S2 is as follows:

[0015] S21. The ODConv module is introduced to improve the C2f layer in the YOLOv8 model, dynamically adjusting the weights of the convolution kernels to adapt to features of different scales and complexities, thereby enhancing the model's feature extraction capabilities.

[0016] S22. An EMA attention mechanism module is introduced to encode global information, divide multiple sub-features in the channel dimension, so that spatial semantic features are evenly distributed in each feature group, and capture pixel-level information through cross-dimensional interactive operations to improve the expressive power of the model.

[0017] S23. To address the issue of imbalanced training samples during bounding box training, the CIoU loss function in YOLOv8 is improved by incorporating the Focal Loss concept to increase the weight of difficult-to-classify samples, making the algorithm pay more attention to these samples and further improving detection accuracy.

[0018] Optionally, S21 specifically refers to:

[0019] The C2f-ODConv module consists of two convolutional layers and n Bottleneck-ODConv modules. The Bottleneck-ODConv module improves the standard convolution in the Bottleneck module into an ODConv module. During the forward propagation of the C2f-ODConv module, the input feature map is first subjected to a standard convolution operation, and then it is divided into two parts. One part is calculated with each of the n Bottleneck-ODConv modules, while the other part remains unchanged. Finally, the two parts are concatenated to form the output feature map.

[0020] Optionally, S22 specifically includes:

[0021] The EMA attention mechanism module encodes the input feature map to recalibrate the channel weights in each parallel branch. For any given input feature map, the EMA attention mechanism module divides the input X into G sub-features across the channel dimension to learn different semantic information, where each group consists of X = [X0, X...]. i …X G-1 ], The EMA attention mechanism is described by extracting attention weights from feature maps through three parallel branches: two are 1×1 branches and the third is a 3×3 branch. In the 1×1 branch, two 1D global average pooling operations are used to encode the channels along two spatial directions, respectively. In the 3×3 branch, single 3×3 convolutional kernels are stacked to capture multi-scale feature representations. Then, the EMA attention mechanism module continues to aggregate cross-spatial information in different spatial dimensions: firstly, the output of the 1×1 branch; secondly, the output of the 3×3 branch. The output of the 1×1 branch is encoded using 2D global average pooling, which can be represented as:

[0022]

[0023] In the formula, z is the output of 2D global average pooling, H is the height, W is the width, and x is the height. c Number of channels;

[0024] The output of 2D global average pooling is multiplied by the output of the 3×3 branch to obtain the first spatial attention map. The global spatial information in the 3×3 branch is then subjected to 2D global average pooling and multiplied by the output of the 1×1 branch to obtain the second spatial attention map that retains the entire precise spatial location information. The two generated spatial attention weight values ​​are aggregated and then reweighted with the input to capture pixel-level pairwise relationships and highlight the global context of all pixels.

[0025] Optionally, S23 specifically refers to:

[0026] Focal Loss adjusts sample weights, decreasing the weight of easily classified samples and increasing the weight of difficult-to-classify samples, thus making the model focus more on the difficult-to-classify samples. It combines the idea of ​​Focal Loss with CIoU, and the formula is:

[0027] L Focal-CIoU =IoU γ L CIoU

[0028]

[0029] In the formula, L Focal-CIoU To combine the loss functions of Focal Loss and CIoU, L CIoU Let CIoU be the loss function, γ be the curvature hyperparameter, IoU be the crossover ratio, and b be the center point of the predicted bounding box. gt α is the center point of the ground truth bounding box, β is the Euclidean distance between the bounding boxes, c represents the distance between the diagonals of the closure regions of the two bounding boxes, α is the weight, and v is used to measure the similarity of aspect ratio.

[0030] Optionally, S4 specifically refers to:

[0031] The performance of a model is measured using recall, precision, average precision, number of parameters, model computational power, and frame rate as evaluation criteria. Precision refers to the proportion of samples that the model predicts as positive, but which are actually positive. Recall refers to the model's ability to identify all positive samples; it is calculated as the sum of the average precision of all labels divided by the total number of categories.

[0032]

[0033] In the formula, R is recall, P is precision, mAP is mean precision, AP is accuracy, FN represents false negatives, TP represents true positives, FP represents false positives, and k classes The number of classes is denoted as , the number of model parameters is the total number of all trainable weight parameters in the model, the computational cost of the model is the theoretical number of billions of floating-point operations per second required when the model performs forward or backward propagation, and the frame rate is the number of image frames processed by the model per second.

[0034] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a fall detection method based on the OEF-YOLO model, which has the following beneficial effects: The present invention improves the YOLOv8 model by using the EMA attention mechanism module, the C2f-ODConv module and the loss function incorporating Focal Loss, which has higher accuracy and generalization ability in different environmental scenarios, significantly improves the reliability of fall detection in indoor scenarios, and achieves a reasonable balance between lightweight and high precision, laying a solid foundation for the deployment of fall detection on terminal devices. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0036] Figure 1 This is a flowchart of the fall detection method based on the OEF-YOLO model of the present invention;

[0037] Figure 2 This is a schematic diagram of the OEF-YOLO model structure of the present invention;

[0038] Figure 3 This is a schematic diagram of the C2f-ODConv module structure of the present invention;

[0039] Figure 4 This is a schematic diagram of the neck network portion of the OEF-YOLO model of the present invention;

[0040] Figure 5 This is a schematic diagram illustrating the function of the EMA attention mechanism module of the present invention;

[0041] Figure 6 This is a schematic diagram showing the comparison of detection performance of different models in the embodiments of the present invention;

[0042] Figure 7 This is a schematic diagram of the detection results of the OEF-YOLO model in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] This invention discloses a fall detection method based on the OEF-YOLO model, such as... Figure 1 As shown, it includes the following steps:

[0045] S1. Construct a dataset for training and testing the performance of the OEF-YOLO model;

[0046] S2. Improve the YOLOv8 model to obtain the OEF-YOLO model for fall detection. The improved OEF-YOLO model is as follows: Figure 2 As shown;

[0047] S3. Train the OEF-YOLO model using the dataset constructed in S1;

[0048] S4. Evaluate the detection performance of the OEF-YOLO model using the dataset constructed in S1.

[0049] Furthermore, S1 specifically refers to:

[0050] The dataset, which contains fall behavior and daily life behaviors, was manually labeled and four categories were set: standing, bending over, sitting, and falling. The images in the dataset include people of different heights and body types, which helps the model learn the features of various human body shapes. The images are taken from multiple angles, which helps the model learn the appearance features of the human body from different angles. The dataset also includes images with varying lighting and partial occlusion, which allows the model to better adapt to different lighting conditions and environments with partial occlusion.

[0051] In this embodiment of the invention, the dataset is based on the following three publicly available datasets: URFD dataset (UR FallDetection Dataset), FDD dataset (Fall detection Dataset), and MCF dataset (MultipleCameras Fall Dataset);

[0052] The URFD dataset provides data on 70 different activities, including 30 fall behaviors and 40 daily life behaviors, captured by two cameras mounted on the floor and ceiling, providing the model with image data from different angles. The FDD dataset was captured in five different rooms, covering images of five different participants. Participant postures included standing, sitting, lying down, bending, and crawling. Some images suffer from insufficient or highly variable lighting, as well as blurry images, which helps improve the model's adaptability to complex lighting and blurry scenes. The MCF dataset was captured by rooftop cameras in eight different directions, containing 24 videos, including 22 fall scenes and 2 normal activity scenes. These images of multiple people help the model achieve better detection results in different angles and multi-person environments.

[0053] The three datasets were manually labeled using LabelImg software. The training set contained 3559 images, and the validation set contained 396 images. During training, 300 training epochs were set, and the number of images processed per epoch was set to 32 to balance computational efficiency and memory usage. The number of CPU threads used during data loading was 8, and the learning rate was 0.01.

[0054] Furthermore, S2 specifically refers to:

[0055] S21. The ODConv module is introduced to improve the C2f layer in the YOLOv8 model, dynamically adjusting the weights of the convolution kernels to adapt to features of different scales and complexities, thereby enhancing the model's feature extraction capabilities.

[0056] S22. An EMA attention mechanism module is introduced to encode global information, divide multiple sub-features in the channel dimension, so that spatial semantic features are evenly distributed in each feature group, and capture pixel-level information through cross-dimensional interactive operations to improve the expressive power of the model.

[0057] S23. To address the issue of imbalanced training samples during bounding box training, the CIoU loss function in YOLOv8 is improved by incorporating the Focal Loss concept to increase the weight of difficult-to-classify samples, making the algorithm pay more attention to these samples and further improving detection accuracy.

[0058] Furthermore, S21 specifically refers to:

[0059] like Figure 3 As shown, the C2f-ODConv module consists of two convolutional layers and n Bottleneck-ODConv modules. The Bottleneck-ODConv module improves upon the standard convolution in the Bottleneck module by replacing it with an ODConv module. During the forward pass of the C2f-ODConv module, a standard convolution operation is first performed on the input feature map. Then, it is divided into two parts: one part is computed sequentially with each of the n Bottleneck-ODConv modules, while the other part remains unchanged. Finally, these two parts are concatenated to form the output feature map. The C2f-ODConv module retains the structure of the C2f module and inherits its advantage of obtaining rich gradient flow information. Furthermore, it benefits from the parallel learning of the ODConv module across the four dimensions of the kernel space, effectively focusing on the characteristics of different convolutional kernels. The C2f-ODConv module not only enhances feature extraction capabilities but also effectively reduces the computational load of the model by decreasing redundant parameters and ineffective computations, thereby improving the model's detection performance and efficiency.

[0060] In this embodiment of the invention, the convolutional performance of the C2f module in YOLOv8 is limited by the fixed number and size of the convolutional kernels in the convolutional layer, making it unable to express richer and more abstract information. However, research on dynamic convolution shows that by learning a linear combination of n convolutional kernels and weighting their inputs, the accuracy of the convolutional network can be significantly improved. Compared to dynamic convolution, which only considers assigning dynamic characteristics to the number of convolutional kernels, the ODConv module considers the number of convolutional kernels, the spatial size of the convolutional kernels, the number of input channels, and the number of output channels, providing more comprehensive dynamic characteristics. The ODConv module can be seen as a new multi-dimensional attention mechanism, learning four types of attention in parallel along the four dimensions of the convolutional kernel space. These four types of attention learned by the ODConv module are complementary, greatly enhancing the feature extraction capability of Convolutional Neural Networks (CNNs). Even if the ODConv module contains only one kernel, it can compete with dynamic convolution with multiple kernels, and even outperform them in performance. The formula for the ODConv module is:

[0061] y=(α w1 ⊙α f1 ⊙α c1 ⊙α s1 ⊙W1+…+α wn ⊙α fn ⊙α cn ⊙α sn ⊙W n )*x

[0062] In the formula, and W represents the input features and output features, respectively. i Indicated by C out The i-th convolutional kernel is composed of filters. m = 1, c out Let α represent the m-th filter of the i-th convolutional kernel, ⊙ represent the multiplication operation along different dimensions of the kernel space, and α represent the multiplication operation along different dimensions of the kernel space. si ∈R k×k This represents the different attention scalars assigned to the convolution parameters (each filter of the convolution kernel) at a K×K spatial location; c represents the value of each convolutional filter. in Each channel is assigned a different attention scalar. Represented as C out Each convolutional filter is assigned a different attention scalar, αW. i Indicates the use of weighted W i The attention scalar is distributed across the entire convolution kernel, which is the same as the attention scalar in dynamic convolution. It is multiplied by the kernel W in the following order: position, channel, filter, and kernel weights. iBy taking into account four types of attention in this way, the ODConv module can significantly enhance the feature extraction capability of CNN convolution operations and capture rich contextual information.

[0063] Furthermore, S22 specifically refers to:

[0064] like Figure 4 and Figure 5 As shown, the EMA attention mechanism module encodes the input feature map to recalibrate the channel weights in each parallel branch. For any given input feature map, the EMA attention mechanism module divides the input X into G sub-features across the channel dimension to learn different semantic information, where each group consists of X = [X0, X...]. i …X G-1 ], The EMA attention mechanism is described by extracting attention weights from feature maps through three parallel branches: two are 1×1 branches and the third is a 3×3 branch. In the 1×1 branch, two 1D global average pooling operations are used to encode the channels along two spatial directions, respectively. In the 3×3 branch, single 3×3 convolutional kernels are stacked to capture multi-scale feature representations. Then, the EMA attention mechanism module continues to aggregate cross-spatial information in different spatial dimensions: firstly, the output of the 1×1 branch; secondly, the output of the 3×3 branch. The output of the 1×1 branch is encoded using 2D global average pooling, which can be represented as:

[0065]

[0066] In the formula, z is the output of 2D global average pooling, H is the height, W is the width, and x is the height. c Number of channels;

[0067] The output of 2D global average pooling is multiplied by the output of the 3×3 branch to obtain the first spatial attention map. The global spatial information in the 3×3 branch is then subjected to 2D global average pooling and multiplied by the output of the 1×1 branch to obtain the second spatial attention map that retains the entire precise spatial location information. The two generated spatial attention weight values ​​are aggregated and then reweighted with the input to capture pixel-level pairwise relationships and highlight the global context of all pixels.

[0068] The neck network is the intermediate part connecting different layers of the network, located between the backbone network and the detection head. The main functions of the neck network include: fusing feature information from each layer of the backbone network to enhance the model's understanding and representation of the target; reducing the dimensionality or number of channels of features, reducing computation while retaining key feature information; and controlling the transmission of features at different resolutions to the detection head to adapt to targets of various scales, improving the model's generalization ability. This invention introduces the EMA attention mechanism into the neck network, enabling the network to more effectively integrate and utilize attention information. By generating higher-quality pixel-level attention for the feature map and better modeling long-range dependencies, the subsequent C2f-ODConv module can more effectively extract features based on the current input context, significantly improving the network's processing capability in fall scenarios. The EMA attention mechanism module can effectively model long-range dependencies and retain accurate positional information, while fusing contextual information at different scales, thereby generating finer pixel-level attention for the feature map. By introducing the EMA attention mechanism module into the neck structure of the YOLOv8 network, the neck network can automatically adjust attention weights according to the current input context.

[0069] Furthermore, S23 specifically refers to:

[0070] Focal Loss adjusts sample weights, decreasing the weight of easily classified samples and increasing the weight of difficult-to-classify samples, thus making the model focus more on the difficult-to-classify samples. It combines the idea of ​​Focal Loss with CIoU, and the formula is:

[0071] L Focal-CIoU =IoU γ L CIoU

[0072]

[0073] In the formula, L Focal-CIoU To combine the loss functions of Focal Loss and CIoU, L CIoU Let CIoU be the loss function, γ be the curvature hyperparameter, IoU be the crossover ratio, and b be the center point of the predicted bounding box. gt α is the center point of the ground truth bounding box, β is the Euclidean distance between the bounding boxes, c represents the distance between the diagonals of the closure regions of the two bounding boxes, α is the weight, and v is used to measure the similarity of aspect ratio.

[0074] YOLOv8 uses the CIoU Loss function for bounding boxes. Intersection over Union (IoU) is a concept in object detection that represents the ratio of the intersection area to the union area between the predicted and ground truth bounding boxes. Traditional IoU loss only focuses on the overlap between bounding boxes, neglecting the positional and shape differences. This leads to inaccurate gradient information when dealing with bounding boxes with large aspect ratios, thus affecting the training performance of the object detection model. The CIoU loss function addresses this issue by combining three parameters: the overlap area, aspect ratio, and center distance. The overlap area represents the ratio of the intersection area to the union area between the detected and ground truth bounding boxes; the aspect ratio is the ratio of the width to the height of the bounding box, helping the model adapt to the shape characteristics of the target; and the center distance is the distance between the center points of the bounding boxes, used to measure the positional offset of the bounding boxes.

[0075] In this embodiment of the invention, during actual model training, the accuracy of easily classified samples can reach very high levels, while the accuracy of difficult-to-classify samples is very low. For example, in an image, the number of high-quality anchor boxes is relatively small compared to low-quality anchor boxes, and the regression error of high-quality anchor boxes is small. Therefore, low-quality anchor boxes will generate excessive gradients, affecting the training process. Considering this sample imbalance problem, the idea of ​​Focal Loss is combined to solve the sample imbalance problem and improve model performance. Focal Loss aims to address the problem of uneven classification difficulty in samples, that is, during training, easily classified samples often achieve high accuracy, while difficult-to-classify samples have low accuracy. To this end, Focal Loss adjusts the sample weights, reducing the weight of easily classified samples and increasing the weight of difficult-to-classify samples, making the model pay more attention to difficult-to-classify samples, which is similar to increasing the proportion of difficult-to-classify samples in the training set. The size of IoU represents the difficulty of sample classification, where a larger IoU indicates that the sample is easier to classify, and a smaller IoU means that the sample is difficult to classify.

[0076] In this embodiment of the invention, setting the curvature hyperparameter γ = 0.5 helps to provide a larger weight increment for difficult-to-classify samples (i.e., samples with small IoU), thereby effectively improving the classification performance of difficult-to-classify samples.

[0077] Furthermore, S4 specifically refers to:

[0078] The performance of a model is measured using recall, precision, average precision, number of parameters, model computational power, and frame rate as evaluation criteria. Precision refers to the proportion of samples that the model predicts as positive, but which are actually positive. Recall refers to the model's ability to identify all positive samples; it is calculated as the sum of the average precision of all labels divided by the total number of categories.

[0079]

[0080] In the formula, R is recall, { is precision, mAP is mean precision, AP is accuracy, the curve plotted with recall on the horizontal axis and precision on the vertical axis is the PR curve, the integral average below the curve is the AP value for that class, FN represents false negatives, TP represents true positives, FP represents false positives, and k classes The number of classes is denoted as , the number of model parameters is the total number of all trainable weight parameters in the model, the computational cost of the model is the theoretical number of billions of floating-point operations per second required when the model performs forward or backward propagation, and the frame rate is the number of image frames processed by the model per second.

[0081] In this embodiment of the invention, mAP can be divided into mAP@0.5 and mAP@0.5:0.95. mAP@0.5 is the mAP value when the IoU threshold is 0.5; mAP@0.5:0.95 is the average mAP value when the IoU starts from 0.5 and increases to 0.95 in steps of 0.05.

[0082] Model parameters are continuously adjusted during training to adapt to the data through optimization algorithms. The number of parameters directly affects the complexity of the model and the demand for hardware resources. A smaller number of parameters means lower model complexity, thereby reducing the demand for computing resources and time during training and inference.

[0083] The computational cost of a model refers to the theoretical number of billions of floating-point operations per second required for forward or backward propagation. This metric is often used to evaluate the computational complexity of a model and its hardware computing power requirements. A lower FLOPs value means that the model performs fewer calculations per unit of time, thus resulting in higher overall computational efficiency. Under the same hardware conditions, a low FLOPs model is easier to deploy than a high FLOPs model.

[0084] Frame rate refers to the number of image frames a model processes per second. A high frame rate indicates a stronger real-time response capability. It's important to note that computer system hardware resources (such as CPU and GPU) can be affected by other background processes or system load, leading to random variations in available computing power. Therefore, in practical applications, the obtained FPS value has a certain degree of randomness. When evaluating the model's frame rate, a rigorous experimental design was employed to minimize the impact of external factors on the results. For example, system resource utilization was dynamically monitored, and resources were prioritized for model execution whenever possible to ensure the accuracy and reliability of the evaluation results. Repeated trials were used to obtain more robust and reliable FPS values.

[0085] In one embodiment of the present invention, an experimental comparison is conducted on the improved portion to detect the improvement effect, specifically:

[0086] The ordinary convolutional module in Bottleneck is improved based on the FasterNetBlock module, the DiverseBranchBlock module in DBBNe, and the ODConv2d module in ODConv; the improvement results are shown in Table 1.

[0087] Table 1. Improvement results of the C2f module

[0088]

[0089] Experimental results show that the improved C2f module not only improves performance, but also reduces the computational load and improves the running efficiency of the model by reducing redundant parameters and invalid calculations in the model.

[0090] Experiments were conducted by introducing EMA attention mechanism modules before and after the C2f-ODConv modules in the backbone network and before and after the last three C2f-ODConv modules in the neck network. The experimental results are shown in Table 2.

[0091] Table 2. Experimental Results of Improved MA Attention Mechanism Module

[0092]

[0093] It can be seen that introducing the C2f-ODConv module before the neck network yields the best performance. Compared with the YOLOv8n model, the improved model's mAP@0.5 increased from 0.964 to 0.976, mAP@0.5:0.95 increased from 0.725 to 0.736, GFLOP decreased from 8.1 to 6.5, and the number of parameters increased slightly (from 3.0M to 3.1M).

[0094] Based on the improvements described above, comparative experiments were conducted on the loss functions. Commonly used loss functions were selected: DIoU, GIoU, EIoU, SIoU, and Focal-CIoU. The experimental results are shown in Table 3.

[0095] Table 3 Experimental results for different loss functions

[0096]

[0097] Experimental results clearly demonstrate that, based on the combination of the C2f-ODConv module and the neck network with the EMA module, the Focal-CIoU loss function can improve the model's performance and accuracy. This proves the effectiveness of the Focal-CIoU loss function in object detection tasks.

[0098] To verify the advantages of the algorithm proposed in this embodiment of the invention in terms of detection accuracy and speed, comparative experiments were conducted with other excellent detection algorithms at present, while keeping the experimental environment unchanged. The results are shown in Table 4 and... Figure 6 As shown, the OEF-YOLO model proposed in this embodiment of the invention is significantly superior to other detectors in terms of detection performance; although it is slightly inferior to YOLOv5n in terms of detection speed, it is far superior to YOLOv5n in terms of accuracy.

[0099] Table 4. Experimental Results of Detection Algorithm Comparison

[0100]

[0101] In this embodiment of the invention, to further verify the generalization ability and anti-interference capability of the OEF-YOLO model, comparative experiments were conducted on different fall scenarios in the dataset, covering various complex conditions such as changes in lighting, perspective shifts, and multi-person scenes. The experimental results are as follows: Figure 7 As shown in the figure, the left side is the original image of the scene, the middle part is the detection result of the YOLOv8n model, and the right part is the detection result of the OEF-YOLO model. It can be seen that the confidence level of the OEF-YOLO model is 15% higher than that of YOLOv8n under normal lighting conditions, while it is 7% higher under low lighting conditions. OEF-YOLO can cope with the impact of lighting changes on detection performance.

[0102] Due to drastic changes in lighting, the outlines of the human body become blurred, making it more difficult to distinguish between the foreground and background. Experimental results show that the YOLOv8n model exhibited false positives in this scene, while the OEF-YOLO model not only avoided false positives but also achieved a confidence level of 71%. This indicates that the OEF-YOLO model, by enhancing its feature extraction capabilities, can effectively address the challenges posed by drastic changes in lighting.

[0103] The OEF-YOLO model achieves a 43% improvement in confidence level compared to YOLOv8n at the vertical viewpoint and a 35% improvement at the horizontal viewpoint. This improvement demonstrates that the OEF-YOLO model maintains high detection accuracy even under complex viewing conditions.

[0104] Due to occlusion between people, the YOLOv8n model missed some detections in the upper right corner of the results, while the OEF-YOLO model successfully detected partially occluded human targets.

[0105] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0106] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fall detection method based on the OEF-YOLO model, characterized in that, Includes the following steps: S1. Construct a dataset for training and testing the performance of the OEF-YOLO model; S2. Improve the YOLOv8 model to obtain the OEF-YOLO model for fall detection; S3. Train the OEF-YOLO model using the dataset constructed in S1; S4. Evaluate the detection performance of the OEF-YOLO model using the dataset constructed in S1.

2. The fall detection method based on the OEF-YOLO model according to claim 1, characterized in that, S1 specifically refers to: The dataset containing fall behavior and daily life behavior was manually labeled, and four categories were set: standing, bending over, sitting and falling. The images in the dataset include people of different heights and body types, which helps the model learn the features of various human body shapes. The dataset also includes images taken from multiple angles, which helps the model learn the appearance features of the human body from different angles. Images including those showing changes in lighting and partial occlusion allow the model to better adapt to different lighting conditions and environments with partial occlusion.

3. The fall detection method based on the OEF-YOLO model according to claim 1, characterized in that, S2 specifically refers to: S21. The ODConv module is introduced to improve the C2f layer in the YOLOv8 model, dynamically adjusting the weights of the convolution kernels to adapt to features of different scales and complexities, thereby enhancing the model's feature extraction capabilities. S22. An EMA attention mechanism module is introduced to encode global information, divide multiple sub-features in the channel dimension, so that spatial semantic features are evenly distributed in each feature group, and capture pixel-level information through cross-dimensional interactive operations to improve the expressive power of the model. S23. To address the issue of imbalanced training samples during bounding box training, the CIoU loss function in YOLOv8 is improved by incorporating the Focal Loss concept to increase the weight of difficult-to-classify samples, making the algorithm pay more attention to these samples and further improving detection accuracy.

4. The fall detection method based on the OEF-YOLO model according to claim 3, characterized in that, S21 specifically refers to: The C2f-ODConv module consists of two convolutional layers and n Bottleneck-ODConv modules. The Bottleneck-ODConv module improves the standard convolution in the Bottleneck module into an ODConv module. During the forward propagation of the C2f-ODConv module, the input feature map is first subjected to a standard convolution operation, and then it is divided into two parts. One part is calculated with each of the n Bottleneck-ODConv modules, while the other part remains unchanged. Finally, the two parts are concatenated to form the output feature map.

5. The fall detection method based on the OEF-YOLO model according to claim 3, characterized in that, S22 specifically refers to: The EMA attention mechanism module encodes the input feature map to recalibrate the channel weights in each parallel branch. For any given input feature map, the EMA attention mechanism module divides the input X into G sub-features across the channel dimension to learn different semantic information, where each group consists of... The EMA attention mechanism is described by extracting attention weights from feature maps through three parallel branches: two are 1×1 branches and the third is a 3×3 branch. In the 1×1 branch, two 1D global average pooling operations are used to encode the channels along two spatial directions, respectively. In the 3×3 branch, single 3×3 convolutional kernels are stacked to capture multi-scale feature representations. Then, the EMA attention mechanism module continues to aggregate cross-spatial information in different spatial dimensions: firstly, the output of the 1×1 branch; secondly, the output of the 3×3 branch. The output of the 1×1 branch is encoded using 2D global average pooling, which can be represented as: In the formula, z is the output of 2D global average pooling, H is the height, W is the width, and x is the height. c Number of channels; The output of 2D global average pooling is multiplied by the output of the 3×3 branch to obtain the first spatial attention map. The global spatial information in the 3×3 branch is then subjected to 2D global average pooling and multiplied by the output of the 1×1 branch to obtain the second spatial attention map that retains the entire precise spatial location information. The two generated spatial attention weight values ​​are aggregated and then reweighted with the input to capture pixel-level pairwise relationships and highlight the global context of all pixels.

6. The fall detection method based on the OEF-YOLO model according to claim 3, characterized in that, S23 specifically refers to: Focal Loss adjusts sample weights, decreasing the weight of easily classified samples and increasing the weight of difficult-to-classify samples, thus making the model focus more on the difficult-to-classify samples. It combines the idea of ​​Focal Loss with CIoU, and the formula is: L Focal-CIoU =IoU γ L CIoU In the formula, L Focal-CIoU To combine the loss functions of Focal Loss and CIoU, L CIoU Let CIoU be the loss function, γ be the curvature hyperparameter, IoU be the crossover ratio, and b be the center point of the predicted bounding box. gt α is the center point of the ground truth bounding box, β is the Euclidean distance between the bounding boxes, c represents the distance between the diagonals of the closure regions of the two bounding boxes, α is the weight, and v is used to measure the similarity of aspect ratio.

7. The fall detection method based on the OEF-YOLO model according to claim 1, characterized in that, S4 specifically refers to: The performance of a model is measured using recall, precision, average precision, number of parameters, model computational power, and frame rate as evaluation criteria. Precision refers to the proportion of samples that the model predicts as positive, but which are actually positive. Recall refers to the model's ability to identify all positive samples; it is calculated as the sum of the average precision of all labels divided by the total number of categories. In the formula, R is recall, P is precision, mAP is mean precision, AP is accuracy, FN represents false negatives, TP represents true positives, FP represents false positives, and k classes The number of classes is denoted as , the number of model parameters is the total number of all trainable weight parameters in the model, the computational cost of the model is the theoretical number of billions of floating-point operations per second required when the model performs forward or backward propagation, and the frame rate is the number of image frames processed by the model per second.

Images