A fall detection method based on lightweight LMBW-YOLO

Abstract

The application provides a fall detection method based on a lightweight LMBW-YOLO, which comprises the following steps: constructing a fall detection dataset; improving a YOLOv5 model to form an LMBW-YOLO model, replacing an original C3 module with a CSRG module and introducing an EMA attention mechanism to reduce the loss in the information transmission process and enhance the feature expression capability; using a WFPN network to optimize a feature fusion network and replacing a standard convolution with a GSConv to reduce the model calculation complexity and enhance the multi-scale feature fusion capability; replacing an original loss function with an Inner-WIoU loss function to accelerate the model convergence and improve the generalization capability; training the LMBW-YOLO model by using the fall detection dataset; and applying the trained LMBW-YOLO model to a fall detection task in an actual scene; and the application can realize model lightweight and improve the detection precision and performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of human body fall detection, and in particular to a fall detection method based on light-weight LMBW-YOLO. BACKGROUND

[0002] Fall is one of the main safety hazards for the elderly. With the increase of age, the movement and balance ability of the elderly gradually weakens, increasing the risk of falling. Therefore, the research of human body fall detection technology is particularly important. This technology can detect and quickly respond to abnormal falling behavior of the elderly in real time, effectively reducing the harm caused by falling, and providing more comprehensive safety protection for the elderly.

[0003] Human body fall detection technology has attracted widespread attention due to its broad application prospects and important significance. Researchers have explored various fall detection methods, among which computer vision fall detection technology based on human pose estimation and target detection algorithm has gradually become a research hotspot. This kind of technology captures real-time images through a camera, combines a deep learning model to extract human features, and thus realizes fall detection. The fall detection algorithm based on human pose estimation mainly focuses on the change of human posture, which can reduce environmental interference; but this method involves multiple computing stages such as key point extraction and analysis, which will lead to high calculation error and high calculation cost. In contrast, the target detection algorithm detects fall events by directly detecting human behavior, simplifying the detection process, reducing calculation error and network complexity, including two-stage algorithm and one-stage algorithm. The two-stage algorithm has an advantage in detection accuracy, but its inference speed is slow, making it difficult to meet the real-time fall detection demand.

[0004] In contrast, the one-stage algorithm, especially the YOLO series algorithm, detects human behavior in an end-to-end manner, with the characteristics of fast, efficient and flexible. At the same time, the fall detection model needs to be deployed on resource-constrained embedded devices, with higher requirements for computational complexity and detection performance. Therefore, the research of light-weight fall detection method based on YOLO is crucial. Current research mainly focuses on improving the detection ability of the model in complex scenes, such as dealing with human occlusion; although these improvements improve the detection accuracy, they also significantly increase the computational burden, limiting the deployment of the model on resource-limited embedded devices. In addition, although the light-weight improvement of the model reduces the computational demand, it often sacrifices the detection accuracy, thereby affecting the accuracy and stability of the fall detection. SUMMARY

[0005] In view of the shortcomings of the prior art, the present application provides a fall detection method based on light-weight LMBW-YOLO, which realizes model light-weight and improves the detection accuracy and performance of the model through optimized structure.

[0006] The technical scheme of the present application is: a fall detection method based on lightweight LMBW-YOLO, comprising the following steps:

[0007] S1), collect and process a plurality of images containing human normal behavior and fall behavior, and construct a fall detection dataset;

[0008] S2), improve the YOLOv5 model to obtain the LMBW-YOLO model;

[0009] S3), train the LMBW-YOLO model using the fall detection dataset;

[0010] S4), apply the trained LMBW-YOLO model to the fall detection task in the actual scene to obtain the detection result of the fall.

[0011] As preferred, in step S1), the fall detection dataset includes images of human normal behavior and fall behavior under various angles, different light conditions and different occlusion degrees.

[0012] As preferred, in step S2), the improvement of the YOLOv5 model includes:

[0013] First, for the backbone network Backbone of the YOLOv5 model; use the CSRG feature extraction module to replace the C3 module of the YOLOv5 model, and introduce the EMA attention mechanism to reduce the interference of environment-independent information and enhance the model's ability to capture human posture details;

[0014] Second, use the WFPN network and GSConv convolution layer to improve the neck network Neck;

[0015] Third, use the Inner-WIoU loss function to replace the original loss function of YOLOv5.

[0016] As preferred, in step S2), the CSRG feature extraction module includes a CBS convolution layer and an SRGBotteleneck module, the SRGBotteleneck module is composed of a RepGhost Bottleneck module and a Split branch, and the Split branch is arranged at the input end of the RepGhost Bottleneck module.

[0017] As preferred, in step S2), the EMA attention mechanism module includes a multi-scale attention module and a cross-space information aggregation module, the channel dimension of the input feature map is adjusted to the batch dimension, the channels are divided into a plurality of sub-feature groups, and then the feature interaction is performed on the two parallel branches to integrate the channel and spatial information.

[0018] As preferred, in step S2), the WFPN network adds a weighted branch between non-adjacent layers and replaces the standard convolution with a GSConv convolution layer; during training, an Inner-WIoU loss function is used.

[0019] The beneficial effects of the present application are:

[0020] 1. The present application constructs a fall detection dataset covering a wide range of scenarios, preventing model overfitting and enhancing its robustness;

[0021] 2. The present application uses a CSRG feature extraction module to replace the original C3 module and introduces an EMA attention mechanism, reducing information loss during transmission and enhancing feature expression ability to capture complex human posture changes;

[0022] 3. The present application uses a WFPN network to optimize the feature fusion network and uses GSConv to replace the standard convolution to reduce the computational complexity of the model and enhance the multi-scale feature fusion capability;

[0023] 4. The present application uses an Inner-WloU loss function to replace the original loss function, accelerating model convergence and improving generalization ability;

[0024] 5. The LMBW-YOLO model of the present application realizes lightweight while improving detection accuracy and performance, which can better meet the needs of fall detection for the elderly. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is a flowchart of the method of the present application;

[0026] Figure 2 is a framework structure diagram of the LMBW-YOLO model of the present application;

[0027] Figure 3 is a framework structure diagram of the CSRG feature extraction module and the GBS convolution layer of the present application;

[0028] Figure 4 is a framework structure diagram of the SRG Bottleneck module of the present application, wherein (a) is a framework diagram of the RepGhostBottleneck module; (b) is a flowchart of the training process of the SRG Bottleneck module; (c) is a flowchart of the inference process of the SRGBottleneck module;

[0029] Figure 5 is a structure diagram of the EMA attention mechanism module of the present application;

[0030] Figure 6 is a structure diagram of the WFPN network of the present application;

[0031] Figure 7 Structure diagram of the GSConv convolutional layer of the application;

[0032] Figure 8 Diagram of the auxiliary bounding box of different scales of the application;

[0033] Figure 9 Diagram of the detection results of the YOLOv5s and LMBW-YOLO models under different illumination conditions of the application;

[0034] Figure 10 Diagram of the detection results of the YOLOv5s and LMBW-YOLO models under different occlusion degrees of the application. DETAILED DESCRIPTION

[0035] The specific embodiments of the application will be further described below in conjunction with the accompanying drawings:

[0036] Embodiment 1

[0037] As shown in the drawings, the embodiment provides a fall detection method based on a lightweight LMBW-YOLO, which comprises the following steps: Figure 1 S1), collect and process a plurality of images containing human normal behavior and fall behavior, and construct a fall detection dataset;

[0038] The embodiment has a total of 4615 images, covering scenes of various angles, different illumination conditions and different occlusion degrees; and the images of the fall detection dataset are labeled and corresponding labels are generated using a labeling tool LabelImg, the behavior categories are divided into normal behavior and fall behavior, wherein the normal behavior is marked as "Person", and the fall behavior is marked as "Down"; and the fall detection dataset is divided into a training set, a validation set and a test set according to a certain proportion;

[0039] In order to improve the generalization ability of the model and prevent overfitting, the training set is subjected to data enhancement processing, including image flipping, adding Gaussian noise and the like. These measures not only expand the training sample size, but also make the dataset more close to various life scenes, thereby ensuring the authenticity and diversity of the dataset.

[0040] S2), improve the YOLOv5 model to obtain an LMBW-YOLO model;

[0041] As shown in the drawings, the embodiment improves the YOLOv5 model in three aspects; specifically:

[0042] Figure 2

[0043] ​​Firstly, for the backbone network Backbone of the YOLOv5 model; the C3 module of the YOLOv5 model is replaced with a CSRG feature extraction module, and an EMA attention mechanism is introduced to reduce the interference of environment-independent information and enhance the model's ability to capture human posture details;

[0044] Secondly, in order to reduce the computational burden and effectively capture multi-level context information, a WFPN network and a GSConv convolution are used to improve the neck network Neck;

[0045] In addition, in order to accelerate the convergence speed of the model and improve the generalization ability of the model, an Inner-WIoU loss function is used to replace the original loss function of YOLOv5.

[0046] The LMBW-YOLO model reduces the computational complexity through the above improvements, more accurately captures the complex changes of human actions, reduces the occurrence of missed detection and false detection, and thus more accurately judges the occurrence of a fall event.

[0047] As preferred in the embodiment, the backbone network Backbone of the LMBW-YOLO model includes a plurality of CBS convolution layers, a plurality of CSRG feature extraction modules, an EMA attention mechanism module and an SPPF pooling layer, as shown in Figure 2 The backbone network Backbone of the LMBW-YOLO model processes the picture as follows:

[0048] The input picture first undergoes channel number transformation and image size reduction through the first convolution layer CBS of the backbone network Backbone, and then enters the second convolution layer CBS for the same operation; then the output of the second convolution layer CBS is used as the input of the first CSRG feature extraction module for feature extraction.

[0049] After the above operations are completed, the third convolution layer CBS and the second CSRG feature extraction module, the fourth convolution layer CBS and the third CSRG feature extraction module, and the fifth convolution layer CBS and the fourth CSRG feature extraction module are used in turn to repeat the operation three times; and the output of the fourth CSRG feature extraction module is used as the input of the EMA attention mechanism module, and the output of the EMA attention mechanism module is used as the input of the SPPF pooling layer.

[0050] In the embodiment, as shown in Figure 3 The CSRG feature extraction module includes a plurality of CBS convolution layers and an SRGBotteleneck module.

[0051] As shown in Figure 4As shown, the SRG Bottleneck module is composed of a RepGhost Bottleneck module and a Split branch, by adding a Split branch at the input end of the RepGhost Bottleneck module; the improved feature extraction module not only reduces the dependence on computing resources, but also reduces the loss in the information transmission process, so that the backbone network Backbone can more accurately capture the detailed features of the target.

[0052] As shown in Figure 4 As shown in (a), the RepGhost Bottleneck module includes two 1x1 standard convolution layers Conv1x1, two depth separable convolution layers DWConv, and two BN layers.

[0053] As shown in Figure 4 As shown in (b), in the training process of the SRG Bottleneck, the original input feature map first passes through the first 1x1 standard convolution layer Conv1x1, and the output of the first 1x1 standard convolution layer Conv1x1 is divided into two branches: the first branch passes through the first depth separable convolution layer DWConv, and the second branch passes through the first BN layer. The outputs of the two branches are element-wise added, and the added output passes through the SiLU activation function. The output processed by the SiLU activation function is concatenated with the original input feature map in the channel dimension. The concatenated output is again passed through the second 1x1 standard convolution layer Conv1x1, and the output of the second 1x1 standard convolution layer Conv1x1 is divided into two branches: the first branch passes through the second depth separable convolution layer DWConv, and the second branch passes through the second BN layer. The outputs of the two branches are element-wise added, and the added output is element-wise added with the original input feature map to obtain the final output, which is output to the next level.

[0054] As shown in Figure 4As shown in (c), during the inference process of SRG Bottleneck, the original input feature map first passes through the first 1×1 standard convolutional layer Conv1x1. The output of the first 1×1 standard convolutional layer Conv1x1 is used as the input of the first depthwise separable convolutional layer DWConv. Then, the output of the first depthwise separable convolutional layer DWConv is passed through the SiLU activation function. The output after the SiLU activation function is then concatenated with the original input feature map in the channel dimension (Concat). The concatenated output is then passed through the second 1×1 standard convolutional layer Conv1x1. The output of the second 1×1 standard convolutional layer Conv1x1 is used as the input of the second depthwise separable convolutional layer DWConv. The output of the second depthwise separable convolutional layer DWConv is added element-wise with the original input feature map to obtain the final output, which is then output to the next layer.

[0055] In this embodiment, the expression for the activation function SiLU of the CSRG feature extraction module is:

[0056]

[0057] In the formula, x represents the input feature map, and sigmoid(x) represents the Sigmoid function;

[0058] The SiLU activation function exhibits smoother characteristics when processing low-value inputs, while maintaining a higher gradient in the high-value input region, effectively avoiding the problems of gradient vanishing and overlearning, thereby helping to improve the efficiency of gradient propagation and accelerate network training.

[0059] like Figure 5 As shown, the EMA attention mechanism module includes a multi-scale attention module and a cross-spatial information aggregation module, which effectively integrates channel and spatial dimension information. It adjusts some channel dimensions of the input feature map to the batch dimension and divides the channels into multiple sub-feature groups to ensure the uniform distribution of spatial semantic features in each feature group. Then, through the cross-spatial information aggregation module, feature interaction is performed on two parallel branches to integrate channel and spatial information, further enhancing feature representation capabilities. The specific operations are as follows:

[0060] The input feature map is first grouped into multiple sub-feature groups by a group operation, each sub-feature group containing specific spatial and semantic information; each sub-feature group is input into a multi-scale attention module through two parallel branches to perform a 3*3 standard convolution operation and horizontal X-direction and vertical Y-direction average pooling X Avg Pool, Y Avg Pool operations; global features along the horizontal and vertical directions are extracted, and then the outputs of the X-direction and Y-direction average pooling operations are concatenated and subjected to a 1*1 standard convolution operation, and then the outputs of the two directions are respectively processed by a Sigmoid function to generate weight information across the X and Y directions, and then the results processed by the Sigmoid function of the two directions are re-weighted with the output of the group operation; then the input is aggregated into the cross-space information aggregation module;

[0061] In the cross-space information aggregation module, the result of the re-weighting operation is first normalized by a group normalization operation GroupNorm, and then an average pooling operation Avg Pool is performed to capture long-distance dependency information, and then the output of the average pooling operation Avg Pool is processed by a Softmax function, and then the result of the Softmax function processing is matrix multiplied with the result of the 3*3 standard convolution operation of the multi-scale attention module Matmul operation;

[0062] Meanwhile, the cross-space information aggregation module also performs an average pooling operation Avg Pool, a Softmax function, and a matrix multiplication Matmul operation on the result of the 3*3 standard convolution operation of the multi-scale attention module;

[0063] Then the results of the two matrix multiplication Matmul operations are added and fused, and then processed by a Sigmoid function, and then re-weighted with the output of the group operation Re-weight to generate the final result.

[0064] As preferred in this embodiment, as shown in Figure 2 The neck network Neck of the improved LMBW-YOLO model is improved by using a WFPN network, and the standard convolution is replaced by a GSConv convolution; the improved neck network Neck of the LMBW-YOLO model includes multiple C3_2 modules, multiple WF_Concat weighted splicing modules, and multiple GBS convolution layers.

[0065] As preferred in this embodiment, as shown in Figure 3 The GBS convolution layer includes a GSConv convolution layer and a BN layer, and the BN layer is connected with a SiLU activation function.

[0066] As preferred in this embodiment, asFigure 2 As shown, the LMBW-YOLO model has three prediction heads P3, P4, P5, wherein P3, P4, P5 represent output results of different feature map scales.

[0067] As shown Figure 6 As shown, the WFPN network introduces a weighted fusion strategy, which aims to assign a unique weight value to each fused feature map, so that the network can learn and emphasize the relative importance of different input features; the weighted fusion strategy adjusts the contribution of each layer feature map dynamically to more effectively fuse multi-level information, and the mathematical expression of the weighted fusion strategy is as follows:

[0068]

[0069] In the formula, I in , I out respectively represent input feature maps and output feature maps, w i , w j are learnable weights used to represent the importance of each input feature in the fusion process, and ε is a small constant to ensure the stability of the overall calculation;

[0070] As shown Figure 7 As shown, the GSConv convolutional layer performs standard convolution operation on the input feature map to generate feature map M1. Then, each channel is respectively convolved to produce M2. Then M1 and M2 are spliced and shuffled by shuffle operation to realize the exchange of information in M1 among different channels and make it penetrate into each part of M2, thereby enhancing the nonlinear expression ability of the model.

[0071] In this embodiment, the neck network Neck of the LMBW-YOLO model processes the image as follows:

[0072] After the operation of the SPPF pooling layer is completed, the SPPF pooling layer output feature map is taken as the input of the first GBS convolutional layer; the output of the first GBS convolutional layer is divided into three branches; wherein the first branch and the feature map output by the second CSRG feature extraction layer are subjected to WF_Concat weighted fusion; the second branch and the output of the fourth GBS convolutional layer are subjected to WF_Concat weighted fusion, and the third branch is subjected to up sampling operation Unsample to expand the size of the feature map; and the third branch and the feature map output by the third CSRG feature extraction layer are subjected to WF_Concat weighted fusion;

[0073] The third branch of the first GBS convolutional layer is input to the first C3_2 module after WF_Concat weighted fusion operation, and the output of the first C3_2 module is taken as the input of the second GBS convolutional layer; the output of the second GBS convolutional layer is divided into two branches: the first branch is subjected to WF_Concat weighted fusion with the output of the third GBS convolutional layer, and the second branch is subjected to WF_Concat weighted fusion with the feature map output by the second CSRG feature extraction layer and the first branch of the first GBS convolutional layer; the second branch of the second GBS convolutional layer is input to the second C3_2 module after WF_Concat weighted fusion operation;

[0074] After the second C3_2 module is completed, the output is divided into two branches: the first branch is input to the P3Head head to generate the output result of the P3 scale feature map after Conv2d standard convolution operation, and the second branch is taken as the input of the third GBS convolutional layer and subjected to channel number transformation and feature map size reduction; the output of the third GBS convolutional layer is divided into two branches: the first branch is subjected to WF_Concat weighted fusion with the outputs of the first and fourth GBS convolutional layers, and the second branch is subjected to WF_Concat weighted fusion with the output of the second GBS convolutional layer; the second branch of the output of the third GBS convolutional layer is subjected to WF_Concat weighted fusion, and the output is taken as the input of the third C3_2 module; the output of the third C3_2 module is divided into two branches: the first branch is input to the P4Head head to generate the output result of the P4 scale feature map after Conv2d standard convolution operation, and the second branch is taken as the input of the fourth GBS convolutional layer; the output of the fourth GBS convolutional layer is subjected to WF_Concat weighted fusion with the outputs of the first and third GBS convolutional layers; the output of WF_Concat weighted fusion is taken as the input of the fourth C3_2 module, and the output of the fourth C3_2 module is input to the P5Head head to generate the output result of the P5 scale feature map after Conv2d standard convolution operation, thereby finally obtaining the output results at three different feature map scales of P3, P4 and P5.

[0075] S3), training the LMBW-YOLO model using the fall detection dataset;

[0076] In this embodiment, the Inner-WIoU loss function is used to optimize network training to enhance the performance of the model, wherein the Inner-WIoU loss function L Inner-WIoU is:

[0077]

[0078] wherein:

[0079]

[0080] IoU = 1 - L IoU ,L IoU ∈[0,1];

[0081] IOU Inner =1-L Inner-IoU L Inner-IoU ∈[0,1];

[0082] Where IoU represents the degree of overlap between the predicted bounding box and the ground truth bounding box. Inner This indicates the degree of overlap between the auxiliary bounding boxes of the predicted bounding box and the ground truth bounding box. Loss function By constructing a metric for the distance between the center points of the predicted bounding box and the ground truth bounding box, a distance attention mechanism is formed, which focuses on the distance between the center points of the anchor box and the target box when they coincide; R WIoU The scaling factor for WIoU is a system that weights distance and the aspect ratio of the bounding box. IoU Indicates IoU loss; (x,y) represents the center point coordinates of the predicted bounding box; (x gt ,y gt ) represents the coordinates of the center point of the true bounding box; W g H g R represents the width and height of the minimum bounding box between the predicted and ground truth boxes. WIoU This allows for greater focus on the L-shape of ordinary mass anchor frames when calculating losses. IoU ;

[0083] loss function In the loss function Based on this, a gradient gain allocation strategy with a dynamic non-monotonic focusing mechanism is constructed by defining the outlier degree β. L represents dynamic updates IoU Mean normalization factor; δ and α are hyperparameters that adjust the anchor box dispersion; r represents the gradient gain; when the outlier β deviates from the set normal range, it means that the anchor box quality is extreme. By assigning a smaller gradient gain r, the model can focus the bounding box regression on anchor boxes with normal quality.

[0084] like Figure 8 The overlap between the predicted bounding boxes and the ground truth bounding boxes generated for different scaling factors is analyzed. The auxiliary bounding boxes differ from the actual bounding boxes only in scale. When the scaling factor ratio is less than 1, the auxiliary bounding box size is smaller than the actual bounding box, and the effective range of the regression is less than L. IoU However, the absolute value of its gradient is greater than L. IoUThe gradient is accelerated, thereby accelerating the bounding box regression convergence of high-quality samples; on the contrary, when the ratio is greater than 1, the auxiliary bounding box of a larger size expands the effective range of regression, thereby improving the regression gain of low-quality samples.

[0085] S4), applying the trained LMBW-YOLO model to the actual scene of the fall detection task to obtain the detection result of the fall.

[0086] Embodiment 2

[0087] The experimental environment of this embodiment is shown in Table 1. The LMBW-YOLO model of this embodiment is trained for a total of 300 iterations, and the resolution of the input image is 640x640 pixels. During the training process, the batch size is set to 16, the initial learning rate is set to 0.01, and the cosine annealing strategy is used to gradually reduce the learning rate. The stochastic gradient descent (SGD) is used as the optimizer, the learning rate momentum is set to 0.937, and the weight decay coefficient is set to 0.0003.

[0088] Table 1 Experimental environment

[0089]

[0090]

[0091] This embodiment uses the precision, recall, average precision mAP0.5 and mAP0.5:0.95, floating point operation amount GFLOPS, and parameter amount Parameter to evaluate the trained LMBW-YOLO model, wherein,

[0092]

[0093] In the formula, TP represents the number of times that the model correctly detects the fall behavior as a fall; TN represents the number of times that the model correctly detects the normal behavior as normal; FP represents the number of times that the model incorrectly detects the normal behavior as a fall; and FN represents the number of times that the model incorrectly detects the fall behavior as normal.

[0094]

[0095] wherein m represents the total number of sample categories, AP i represents the average precision of the i-th sample, which is obtained from the area surrounded by the Precision-Recall curve and the coordinate axis.

[0096] The results of a series of ablation experiments on the LMBW-YOLO model are shown in Table 2. Specifically, model A represents the original YOLOv5s model; models B and C represent replacing the original C3 module with the improved CSRG module in the feature extraction network and introducing the EMA attention mechanism module, which improves the precision, recall rate, mAP0.5, and mAP0.5:0.95 indicators by 1.3%, 1.2%, 0.6%, and 2.0%, respectively, fully demonstrating that the proposed measures effectively enhance the model's attention to human targets. Models D and E represent optimizing the feature fusion network to the WFPN network and using GSConv convolution instead of standard convolution, which further improves the precision, recall rate, mAP0.5, and mAP0.5:0.95 indicators by 1.1%, 0.8%, 0.3%, and 0.6%, respectively, indicating the dual advantages of the improved measures in reducing model computational complexity and enhancing information fusion efficiency. In addition, model F uses Inner-WIoU Loss instead of CIoU Loss as the bounding box regression loss function, which improves the precision, recall rate, mAP0.5, and mAP0.5:0.9 by 0.8%, 0.7%, 0.6%, and 1.3%, respectively, showing the significant effect of Inner-WIoU Loss in improving the positioning accuracy of target bounding boxes. In terms of model parameter quantity and computational quantity, the LMBW-YOLO model reduces by 19.2% and 21.3% compared with YOLOv5s, respectively.

[0097] Table 2 Performance of different improved models

[0098]

[0099] Through these ablation experiments, it is verified that the proposed improvement measures not only effectively improve the detection performance of the model, but also reduce the computational complexity of the model, thereby confirming the effectiveness of the improvement strategy. From the results before and after optimization, it can be seen that the LMBW-YOLO model exhibits superior performance advantages in detecting human falling behaviors compared to the original YOLOv5s model.

[0100] This embodiment uses the same dataset and experimental settings to compare the LMBW-YOLO model with current mainstream object detection models (such as Faster R-CNN, SSD, RT-DETR, and the YOLO series). Table 3 shows the performance of different lightweight models in the fall detection task. Compared with two-stage detection models, the LMBW-YOLO model outperforms Faster R-CNN by 4.2% on the mAP0.5 metric. Similarly, compared with the transformer-based RT-DETR, the LMBW-YOLO model achieves a 2.6% advantage on the same metric. Compared with one-stage detection models, the LMBW-YOLO model outperforms SSD by 2.3% on the mAP0.5 metric, and outperforms YOLOv5s, YOLOX-s, YOLOv7-tiny by 1.5%, 1.1%, 2.2%, and 0.3%, respectively. In the YOLO lightweight series, although MBW-YOLO's mAP of 0.5:0.95 is slightly lower than YOLOv8s' 0.2%, its parameters and computational cost are reduced by 48.6% and 56.1%, respectively. In terms of model size and computational resources, LMBW-YOLO has fewer parameters than other models and requires fewer computational resources.

[0101] Table 3 Performance of different lightweight models

[0102] Algorithm mAP0.5 / % mAP0.5:0.95 / % Parameter / M GFLOPS Faster R-CNN 88.5 60.9 41.3 189 SSD 89.3 62.3 24.1 30.5 RT-DETR 90.1 64.6 32.0 110.2 YOLOv5s 91.2 66.1 7.0 16.0 YOLOX-s 91.6 68.8 8.9 26.6 YOLOv7-tiny 90.5 65.2 6.0 13.2 YOLOv8s 92.4 70.3 11.1 28.7 LMBW-YOLO 92.7 70.1 5.7 12.6

[0103] Comparative experimental results show that the LMBW-YOLO model maintains high detection accuracy while significantly reducing its dependence on computing resources. Its overall detection performance is superior to other models. Its excellent detection performance, low resource dependence, and real-time detection capability fully meet the requirements of fall detection tasks for high accuracy, lightweight design, and fast response.

[0104] This embodiment comprehensively analyzes and visualizes fall scenarios under various lighting conditions and degrees of occlusion from multiple angles, using the YOLOv5s model and the LMBW-YOLO model. For example... Figure 9 and Figure 10 As shown, the differences in performance between the two models in different scenarios are revealed. The left side is the original scene image, while the middle and right sides show the detection results of YOLOv5s and LMBW-YOLO, as well as the corresponding Grad-CAM heatmap visualizations.

[0105] Figure 9 This indicates that in dimly lit environments, YOLOv5s may misidentify environmental objects as people, while the LMBW-YOLO model can accurately detect falls under different lighting conditions. From Figure 10It can be seen from the detection results that when the human body is blocked, the detection accuracy of YOLOv5s decreases significantly, while LMBW-YOLO can effectively reduce the interference of environmental information by improving the attention to human targets and enhancing the learning of human features, thereby accurately detecting the blocked human targets. By comparing the detection results in different scenes, the LMBW-YOLO model has higher accuracy and stability in detecting normal behavior and falling behavior than the YOLOv5s model, significantly reducing the occurrence of missed detection and false detection. Therefore, the improvement measures proposed in this embodiment effectively improve the detection performance and generalization ability of the LMBW-YOLO model, and have significant performance advantages over the YOLOv5s model in the fall detection task.

[0106] In summary, the present application proposes a high-precision, lightweight and real-time fall detection model LMBW-YOLO to overcome the limitations of low detection accuracy and high computational complexity of existing fall detection models. By collecting and processing a number of images containing human normal behavior and falling behavior in different scenes, a fall detection dataset is constructed for model training. Based on the YOLOv5 model, the CSRG feature extraction module is used to replace the original C3 module, and the EMA attention mechanism is introduced to enhance the model's ability to capture key features and improve the attention to human bodies. Second, the feature fusion network is optimized to WFPN network and the GSconv convolution is used to replace the standard convolution, which reduces the computational complexity while improving the efficiency of multi-scale information fusion. In addition, Inner-WIoU Loss is used to replace CIoU Loss to improve the convergence speed and generalization ability of the model. The present application also builds an experimental platform to conduct in-depth experimental verification of the LMBW-YOLO model. The experimental results show that LMBW-YOLO improves the mAP0.5 index by 1.5% compared to YOLOv5s, and reduces the parameter amount and computational amount by 19.2% and 21.3% respectively. This indicates that the LMBW-YOLO model not only has a more streamlined network structure, but also significantly outperforms YOLOv5s in terms of detection accuracy and computational efficiency. Compared with the current mainstream lightweight target detection model, LMBW-YOLO has higher detection accuracy and lower computational complexity, with an mAP0.5 index of 0.3% higher than YOLOv8, and its overall detection performance is better than other models. It is fully proved that the LMBW-YOLO model proposed in the present application can better meet the needs of the fall detection task in actual scenarios, and in the future, the LMBW-YOLO model will be deployed on resource-limited embedded devices for further optimization to achieve higher application value and practicality.

[0107] The foregoing embodiments and description of the application only illustrate the principle and the best mode of the application, and various changes and modifications can be made to the application without departing from the spirit and scope of the application, and all these changes and modifications fall within the scope of the application.

Claims

1. A fall detection method based on lightweight LMBW-YOLO, comprising the following steps: S1) Collect and process a plurality of images containing human normal behavior and fall behavior to construct a fall detection dataset; S2), the YOLOv5 model is improved to obtain an LMBW-YOLO model; wherein Improvements to the YOLOv5 model include: First, for the backbone network Backbone of the YOLOv5 model; use the CSRG feature extraction module to replace the C3 module of the YOLOv5 model, and introduce the EMA attention mechanism; wherein the CSRG feature extraction module comprises a CBS convolution layer and an SRG Bottleneck module, the SRG Bottleneck module is composed of a RepGhost Bottleneck module and a Split branch, and the Split branch is arranged at the input end of the RepGhost Bottleneck module; Second, use WFPN network and GSConv convolution to improve the neck network Neck; Third, use the Inner-WIoU loss function to replace the original CIoU loss function of YOLOv5; S3) Train the LMBW-YOLO model using the fall detection dataset; S4) Apply the trained LMBW-YOLO model to the fall detection task in the actual scene to obtain the detection result of the fall.

2. The fall detection method based on lightweight LMBW-YOLO according to claim 1, characterized in that: In step S2), the backbone network Backbone of the LMBW-YOLO model comprises a plurality of CBS convolution layers, a plurality of CSRG feature extraction modules, an EMA attention mechanism module and an SPPF pooling layer.

3. The fall detection method based on lightweight LMBW-YOLO according to claim 2, characterized in that: The expression of the activation function SiLU of the CSRG feature extraction module is: Wherein, x represents the input feature map, and sigmoid(x) represents the Sigmoid function.

4. The fall detection method based on lightweight LMBW-YOLO according to claim 1, characterized in that: In step S2), the EMA attention mechanism module comprises a multi-scale attention module and a cross-space information aggregation module, adjusts the channel dimension of the input feature map to the batch dimension, divides the channels into a plurality of sub-feature groups, and then performs feature interaction on the two parallel branches to integrate the channel and spatial information.

5. The fall detection method based on lightweight LMBW-YOLO according to claim 1, characterized in that: In step S2), the neck network Neck of the LMBW-YOLO model comprises a plurality of C3_2 modules, a plurality of WF_Concat weighted splicing modules and a plurality of GBS convolution layers.

6. The fall detection method based on lightweight LMBW-YOLO according to claim 5, characterized in that: In step S2), the GBS convolution layer comprises a GSConv convolution layer and a BN layer, and the BN layer is connected with a SiLU activation function.

7. The fall detection method based on lightweight LMBW-YOLO according to claim 1, characterized in that: In step S2), the WFPN network adds a weighted branch between non-adjacent layers, and uses the Inner-WIoU loss function during training.

8. The fall detection method based on lightweight LMBW-YOLO according to claim 7, characterized in that: The WFPN network introduces a weighted fusion strategy, which aims to assign a unique weight value to each fused feature map, so that the network can learn and emphasize the relative importance of different input features; the weighted fusion strategy dynamically adjusts the contribution of each layer feature map to more effectively fuse multi-level information, and the mathematical expression of the weighted fusion strategy is as follows: In the formula, I in , I out respectively represent input feature maps and output feature maps, w i , w j are learnable weights used to represent the importance of each input feature in the fusion process, and ε is a small constant used to ensure the stability of the overall calculation.

9. The fall detection method based on lightweight LMBW-YOLO according to claim 7, characterized in that: The network training is optimized using an Inner-WIoU loss function to enhance the model performance, wherein the Inner-WIoU loss function L Inner-WIoU is: Wherein: IoU = 1 - L IoU L IoU ∈ [0, 1] IOU Inner = 1 - L Inner-IoU , L Inner-IoU ∈ [0,1] wherein, IoU represents the overlapping degree of the prediction box and the real box, IoU Inner represents the overlapping degree of the auxiliary bounding box of the prediction box and the real box; the loss function By constructing the distance between the center points of the prediction box and the real box, the distance attention mechanism is formed, which focuses on the distance between the center points of the anchor box and the target box when the anchor box and the target box coincide; R WIoU represents the proportion factor of WIoU, which is a system weighted by the distance and the aspect ratio of the target box, L IoU represents the IoU loss; (x, y) represents the center point coordinates of the prediction box; (x gt ,y gt ) represents the center point coordinates of the real box; W g ,H g is the width and height of the minimum bounding box of the prediction box and the real box; R WIoU , so that more attention is paid to the L IoU of the ordinary quality anchor box when calculating the loss. Loss function is based on the loss function , a gradient gain distribution strategy with dynamic non-monotonic focusing mechanism is constructed by defining the outlying degree β. represents the dynamically updated L IoU is the mean normalization factor; δ, α are hyperparameters that adjust the dispersion of anchor boxes; r represents the gradient gain; when the outlying degree β deviates from the set normal range, it means that the anchor box quality is extreme, and by assigning a smaller gradient gain r, the model can focus on the anchor box with ordinary quality.

Images