Search

CN-122024307-A - Indoor old people falling detection method based on improvement YOLOv8

CN122024307ACN 122024307 ACN122024307 ACN 122024307ACN-122024307-A

Abstract

The application relates to an indoor old people falling detection method based on an improvement YOLOv8, which comprises the steps of obtaining UMFD falling data sets, dividing the data sets into training data sets and test data sets, constructing an indoor old people falling detection model based on YOLOv n, replacing BottleNeck in a C2f module in a YOLOv n network backbone network with BottleNeck-ODConv modules based on a YOLOv n network, fusing the C2f module in a neck network by using FASTERNET modules, fusing a LSKA attention mechanism in an SPPF, embedding the SEAM attention mechanism at the front end of a detection head, training and optimizing the model by using the training set, and testing the testing set to obtain the indoor old people falling detection model. The model of the application obviously reduces false detection in a complex environment and realizes balance between accuracy and light weight.

Inventors

  • LIU YUNTING
  • LIU XINRAN

Assignees

  • 沈阳理工大学

Dates

Publication Date
20260512
Application Date
20251028

Claims (10)

  1. 1. The method for detecting the falling of the old indoor based on the improvement YOLOv is characterized by comprising the following steps: S1, acquiring UMFD falling data sets, respectively marking data of target objects contained in multi-target video images in UMFD falling data sets, and dividing UMFD falling data sets into training data sets and test data sets; S2, constructing an indoor old people falling detection model based on YOLOv n, wherein the indoor old people falling detection model is based on a YOLOv n network, wherein BottleNeck in a C2f module in a backbone network of the YOLOv n network is replaced by BottleNeck-ODConv modules and named as C2f-ODConv modules, the FASTERNET modules are used for fusing the C2f module in the neck network, and a LSKA attention mechanism is fused in the SPPF and named as an SPPF-LSKA module; S3, training and optimizing the indoor old people falling detection model by using the training set, and testing by using the testing set to obtain a final indoor old people falling detection model; S4, acquiring a video image to be detected, and inputting the video image to be detected into an indoor old people falling detection model for detection.
  2. 2. The fall detection method for the indoor old based on the improvement YOLOv, according to claim 1, is characterized in that the C2f-ODConv module comprises two paths, one path is a convolution block and a segmentation module which are sequentially connected and serve as shallow features to the fusion module, the other path is a convolution block and a segmentation module which are sequentially connected, n series-connected Bottleneck-ODConv modules are connected and serve as depth features to the fusion module, and finally shallow features and deep features of the two paths are fused in the fusion module to obtain output.
  3. 3. The improved YOLOv-based fall detection method for elderly people indoors of claim 2, wherein the ODConv module is defined as: ; Wherein, the And Respectively representing input features and output features; The representation is composed of First of filter components A convolution kernel; The representation is composed of First of filter components A convolution kernel; , Represent the first The first convolution kernel A filter; representing multiplication operations performed in kernel space; Is shown in Different attention weights assigned to the convolution parameters in the spatial locations; Representing each convolution filter The individual channels are assigned different attention weights; Represented as The convolution filters assign different attention weights; representation for weighting Weights of (2); representation for weighting Is a convolution kernel weight of (2); representing different attention weights assigned to the convolution parameters in the spatial locations; representing that the input channels of each convolution filter are assigned different attention weights; the different attention weights are assigned to the output channel convolution filters; The channel dimension realizes dynamic coordinates; Coordinates representing the dynamics of the kernel dimension; And Is two orthogonal but cooperative dimensions that together give the sample adaptation capability prior to the convolution operation.
  4. 4. The improved YOLOv-based fall detection method for elderly people indoors of claim 1, wherein the output formula of LSKA in the SPPF-LSKA module is: ; ; ; ; Wherein, the A convolution operation is represented and is performed, Represents the Hadamard product operation, d represents the expansion rate, Indicating that the convolution kernel is of size Is a function of the output of the depth-expanded convolution of (c), Indicating that the convolution kernel is of size Is a function of the output of the depth convolution of (a), Representing an attention-seeking diagram, Representation of Is a convolution kernel of (2); The output of the representation LSKA is provided, Representing the height of the input feature map, Representing the width of the input feature map, is the spatial dimension of the feature map, Representing a deep convolution of the data, Representing the dilation rate of the dilation convolution, Representing the input profile.
  5. 5. The improved YOLOv-based fall detection method for elderly people indoors of claim 1, wherein the FASTERNET module includes a partial convolution, a point-by-point convolution, a batch normalization, a ReLU activation function, and a point-by-point convolution connected in sequence, and the residuals add the processed features to the module input connections.
  6. 6. The improved YOLOv-based fall detection method for elderly people indoors of claim 5, wherein the point-wise convolution and partial convolution calculation formula of the flow is as follows: ; ; the delay calculation formula from the point-wise and partially convolved FLOPS is as follows: ; The calculation formulas of the point-by-point convolution and the partial convolution are as follows: ; ; In the formula, , Representing the height and width of the input feature map, The number of total input channels is indicated, Representing the number of part of the input channels, Representing the size of the convolution kernel.
  7. 7. The method for detecting fall of elderly people in room based on improvement YOLOv according to claim 1, wherein the working flow of the SEAM attention mechanism is that firstly, a feature map is input to connect CNN, then the feature map is split into two branches, namely, one is directly connected with CAM and two paths are further derived, one is connected with PCM and the other is directly generated to be basic heat map, meanwhile, the other branch of the feature map split is directly connected with PCM and then generates a deeper heat map, and then the heat map is output through a prediction layer The loss function L1 is combined with isomorphism regularization to jointly optimize model performance.
  8. 8. The improved YOLOv-based indoor elderly fall detection method of claim 7, wherein the isomorphism regularization formula is: ; Wherein, the Representing a network prediction function Represents any spatial affine transformation that is performed, Representing the spatially affine transformed image, Representing a corresponding class activation graph; The evaluation formula of the PCM is: ; Wherein, the Representing input feature vectors, where And Respectively representing different pixels or feature points, The transpose of the matrix is represented, Representing the embedded function, in order to perform the similarity calculation, Represents a ReLU activation function, for non-linear transformation of the similarity score, Representing the pixel value corresponding to the original CAM, Representing the final CAM.
  9. 9. An electronic device comprising a memory for storing computer instructions and a processor for invoking the computer instructions from the memory to perform the detection method of any of claims 1-8.
  10. 10. A computer-readable storage medium, characterized in that a computer program is stored thereon, which, when being executed by a processor, implements the detection method according to any of claims 1-8.

Description

Indoor old people falling detection method based on improvement YOLOv8 Technical Field The invention relates to the technical field of video analysis and identification, in particular to an indoor old people falling detection method based on an improvement YOLOv. Background Along with the aging of population, falling becomes one of the main factors threatening the health and life safety of the elderly, and not only does falling cause serious bone injury, but also secondary injury and even death can be caused by untimely rescue. However, the existing manual monitoring and wearable sensor methods have the problems of poor compliance, limited coverage range and the like, and are difficult to meet the requirements of large-scale and real-time monitoring. The existing fall detection method comprises an algorithm based on a sensor and traditional computer vision, and a computer vision fall detection method. Algorithms based on sensors and traditional computer vision rely on hardware devices such as acceleration sensors, pressure sensors, etc. Sensor fall detection relies on wearable devices or indoor fixed mounted sensors to monitor the physical condition and movement status of the elderly. Although the algorithm based on the sensor and the traditional computer vision has certain real-time performance and precision, the algorithm is limited by equipment dependence, wearing burden of a user, deployment and maintenance cost and privacy risk, and is difficult to realize large-scale application. In contrast, the computer vision fall detection technology mainly utilizes a deep learning model to analyze the action gesture of the human body in real time, has the advantages of non-contact and no wearing burden, and can realize continuous monitoring under the condition of not interfering with the daily activities of users. A fall detection algorithm based on YOLOv n is proposed in the prior art. The SC module is designed and embedded into the C2f module, the edge characteristic of falling behaviors is extracted by utilizing an edge detection algorithm, the detection precision of the model is improved, and the problem of large model volume is still difficult to solve. A learner introduces a multidimensional cooperative attention Mechanism (MCA) on the basis of YOLOv n model, and adds the MCA after the up-sampling stage at Neck end and each C2f module so as to enhance the local feature interaction capability. And Alpha-SIoU loss function is adopted to provide higher-precision positioning information for the boundary box, so that model convergence is quickened, detection efficiency is improved, and the parameter quantity is also greatly improved. The scholars also use SPD-Conv to reserve complete channel information based on YOLOv n, so that the detection performance of low resolution and small targets is improved, a position information attention mechanism is introduced to enhance the positioning capability of human targets, LSKNet is added in feature extraction to dynamically adjust receptive fields, and the model improves the perception precision of complex scenes, but the quantity of parameters and the calculated quantity are increased. The algorithm based on YOLOv improvement is excellent in many public data sets. However, most of the fall detection algorithms are researched at present to improve the detection precision of targets, and the situations that false detection occurs under illumination interference, light weight and accuracy are difficult to balance still exist, and privacy of users is ignored. Therefore, it is needed to provide a detection method capable of solving the problems of high false detection rate, difficult light weight and difficult accuracy under the condition of shielding and illumination interference. Disclosure of Invention The invention provides an indoor old people falling detection method based on an improvement YOLOv, and aims to solve the problems that the false detection rate is high under shielding and illumination interference, and the weight reduction and the accuracy are difficult to achieve. The technical scheme is as follows: the invention provides an indoor old people falling detection method based on an improvement YOLOv, which comprises the following steps: S1, acquiring UMFD falling data sets, respectively marking data of target objects contained in multi-target video images in UMFD falling data sets, and dividing UMFD falling data sets into training data sets and test data sets; S2, constructing an indoor old people falling detection model based on YOLOv n, wherein the indoor old people falling detection model is based on a YOLOv n network, wherein BottleNeck in a C2f module in a backbone network of the YOLOv n network is replaced by BottleNeck-ODConv modules and named as C2f-ODConv modules, the FASTERNET modules are used for fusing the C2f module in the neck network, and a LSKA attention mechanism is fused in the SPPF and named as an SPPF-LSKA module; S3, training and optimiz