Lightweight human body target detection method and system based on improved YOLOv11

Abstract

The invention discloses a lightweight human body target detection method and system based on improved YOLOv11, and belongs to the technical field of target detection. The method comprises the following steps: acquiring a background image and a human body target image, constructing a data set, and dividing the data set into a training set, a verification set and a test set; the method comprises the following steps: constructing a lightweight human body target detection model LightSeek-YOLO; training the constructed lightweight human body target detection model LightSeek-YOLO by using the training set, and evaluating a target detection result through an evaluation index; and testing the trained lightweight human body target detection model LightSeek-YOLO by using the test set. According to the method, the personnel target can be detected more accurately and quickly in a complex disaster scene, and powerful technical support is provided for disaster rescue.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a lightweight human target detection method and system based on an improved YOLOv11. Background Art

[0002] Globally, the frequency and impact of disasters are increasing, dealing a heavy blow to human society and the ecological environment. From raging forest fires to devastating earthquakes, from fierce floods to extreme weather disasters, these disasters not only seriously threaten people's lives and property, but also cause immeasurable damage to infrastructure and ecosystems. During post-disaster rescue operations, rapid and accurate detection of human targets within the affected area plays a crucial role in implementing effective rescue efforts and minimizing casualties.

[0003] After a disaster, the on-site environment becomes extremely complex. Collapsed buildings create vast rubble, twisted and deformed roads, and accumulation of debris hinders rescue efforts and greatly complicates human object detection. Fire scenes are filled with intense heat, smoke, and flames. The heat can damage detection equipment, while the smoke obscures vision, rendering vision-based detection methods ineffective. Floods can inundate large areas, trapping affected people on floating objects. The water flows rapidly, creating ever-changing conditions. Mudslides can cause massive accumulations of mud, sand, and rocks, burying victims and complicating detection and rescue efforts. These complex and ever-changing environmental factors place extremely high demands on the adaptability of human object detection technology.

[0004] The rapid development of deep learning technology has revolutionized human object detection. By building multi-layer neural networks, deep learning can automatically learn complex feature representations from large amounts of data, eliminating the need for manual feature design. In the field of human object detection, deep learning-based object detection algorithms include the Regional Convolutional Neural Network (R-CNN) series, the YOLO (YOLO) series, and the Single Shot Detector (SSD). Numerous researchers are also pursuing the application of deep learning-based object detection models to disaster personnel detection. Their achievements have brought new hope and possibilities for improving disaster rescue efficiency and protecting the lives of disaster victims. Some researchers are working to use deep learning to optimize human object detection models in complex disaster scenarios. Focusing on post-earthquake debris scenes, they collected a large amount of real-world post-disaster imagery and trained a modified R-CNN model. This effectively improved the accuracy of the detection model.

[0005] While these studies have made progress, they are computationally intensive and slow to detect. They cannot achieve real-time detection of disaster victims. Furthermore, in complex disaster environments, high temperatures, dense smoke, and flames significantly impact model detection accuracy. Summary of the Invention

[0006] The purpose of the embodiments of the present invention is to provide a lightweight human target detection method and system based on improved YOLOv11, which can detect human targets more accurately and quickly in complex disaster scenarios, providing strong technical support for disaster rescue, thereby solving at least one technical problem involved in the background technology.

[0007] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0008] The embodiment of the present invention provides a lightweight human target detection method based on an improved YOLOv11, comprising the following steps:

[0009] Step S1: obtain a background image and a human target image, use the U2Net model to extract a human foreground image from the human target image, and superimpose the human foreground image onto the background image to generate a composite image. The composite image is annotated in YOLO format to construct a dataset, and the dataset is divided into a training set, a validation set, and a test set.

[0010] Step S2, constructing a lightweight human target detection model LightSeek-YOLO, whose network architecture includes a backbone network HGnetV2, a neck network Neck, and a head network Head. The backbone network HGnetV2 includes a feature extraction module HGBlock composed of a LightConV lightweight convolution module and a CBS convolution module; the neck network Neck includes a SeekDS downsampling module combined with a C3K2 module and placed before the C3K2 module; the head network Head is a Seek-CD detection head;

[0011] Step S3, using the training set to train the constructed lightweight human target detection model LightSeek-YOLO, and evaluating the target detection results through evaluation indicators;

[0012] Step S4: Use the test set to test the trained lightweight human target detection model LightSeek-YOLO.

[0013] Optionally, step S1 specifically includes:

[0014] Step S11, obtaining a background image and a human target image;

[0015] Step S12, using the U2Net model to perform background segmentation on the human target image, accurately extract the human foreground image, and remove redundant background;

[0016] Step S13, cropping and denoising the segmented human foreground image to retain only the human body part;

[0017] Step S14, randomly scaling the cropped and denoised human foreground image to a suitable size, and then superimposing it on the background image to generate a composite image;

[0018] Step S15: annotate the synthesized image according to the Yolo format to construct a data set.

[0019] Optionally, in step S11, a background image is obtained from the AIDER dataset, and a human target image is obtained from the LSP / MPII-MPHB dataset.

[0020] Optionally, in step S1, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0021] Optionally, in step S2, the backbone network HGnetV2 also includes an HGStem module, a DWConV module, an SPPF module and a C2PSA module, wherein the HGStem module, the HGBlock module and the DWConV module mainly constitute the functions of the backbone network HGnetV2; the SPPF module is a spatial pyramid pooling layer, which is used to aggregate features at multiple scales; the C2PSA module is an advanced feature extraction module that combines the CSP structure and the PSA attention mechanism, which can enhance the feature extraction capability.

[0022] Optionally, in step S3, training parameters need to be set before training, specifically including:

[0023] Set the training round epoch = 150, batch size batch_size = 32, set the network initial learning rate to 0.01, use the stochastic gradient descent optimizer, and uniformly scale the input image size to 640×640.

[0024] The present invention also provides a lightweight human target detection system based on an improved YOLOv11 for executing the method, comprising:

[0025] The dataset construction module is used to obtain background images and human target images, extract the human foreground image from the human target image using the U2Net model, and superimpose the human foreground image on the background image to generate a composite image. The composite image is annotated in the YOLO format to construct a dataset, and the dataset is divided into a training set, a validation set, and a test set.

[0026] A model construction module is used to build a lightweight human target detection model LightSeek-YOLO. Its network architecture includes a backbone network HGnetV2, a neck network Neck, and a head network Head. The backbone network HGnetV2 includes a feature extraction module HGBlock composed of a LightConV lightweight convolution module and a CBS convolution module; the neck network Neck includes a SeekDS downsampling module combined with a C3K2 module and placed before the C3K2 module; the head network Head is a Seek-CD detection head;

[0027] The training module is used to train the constructed lightweight human target detection model LightSeek-YOLO using the training set;

[0028] The testing module is used to test the trained lightweight human target detection model LightSeek-YOLO using the test set and evaluate the target detection results.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] 1. The lightweight design provided by the present invention can be better deployed on edge devices.

[0031] 2. The dataset provided by the present invention is suitable for disaster scenarios, and the module is designed specifically for this scenario, which can provide strong support for the search for disaster victims.

[0032] 3. The dual-path feature stream of the Seek-DS module provided by the present invention expands the effective receptive field to more than twice that of traditional methods. The adaptive feature fusion mechanism significantly improves the coupling degree of multi-scale features and is optimized for the characteristics of disaster scenarios, enabling it to maintain stable detection performance under harsh conditions such as smoke and lighting changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0034] Figure 1 A flowchart of a lightweight human target detection method based on improved YOLOv11 provided by the present invention;

[0035] Figure 2 A schematic diagram of the network structure of the lightweight human target detection model LightSeek-YOLO provided by the present invention;

[0036] Figure 3 A schematic diagram of the structure of the HGblock module provided by the present invention;

[0037] Figure 4 A schematic structural diagram of the HGstem module provided by the present invention;

[0038] Figure 5 A schematic diagram of the structure of the SeekDS downsampling module provided by the present invention;

[0039] Figure 6 A schematic diagram of the structure of the Seek-CD detection head provided by the present invention;

[0040] Figure 7 This is a structural block diagram of the lightweight human target detection system based on the improved YOLOv11 provided by the present invention;

[0041] Figure 8 This is one of the structural diagrams of the electronic device provided by the present invention;

[0042] Figure 9 This is the second structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0044] The terms "first," "second," and the like in the specification and claims of the present invention are used to distinguish similar objects, and are not used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments of the present invention can be implemented in an order other than that illustrated or described herein, and that the objects distinguished by "first," "second," and the like are generally of the same type, and do not limit the number of objects. For example, the first object can be one or more. In addition, the term "and / or" in the specification and claims refers to at least one of the connected objects, and the character " / " generally indicates that the objects connected are in an "or" relationship.

[0045] See Figure 1 As shown, the embodiment of the present invention provides a lightweight human target detection method based on improved YOLOv11, including the following steps:

[0046] Step S1: obtain a background image and a human target image, use the U2Net model to extract a human foreground image from the human target image, and superimpose the human foreground image onto the background image to generate a composite image. The composite image is annotated in YOLO format to construct a dataset, and the dataset is divided into a training set, a validation set, and a test set.

[0047] Step S2: Construct a lightweight human target detection model LightSeek-YOLO (see Figure 2 As shown), its network architecture includes a backbone network HGnetV2, a neck network Neck, and a head network Head. The backbone network HGnetV2 includes a feature extraction module HGBlock composed of a LightConV lightweight convolution module and a CBS convolution module (see Figure 3 The neck network Neck includes a SeekDS downsampling module combined with a C3K2 module and placed before the C3K2 module (see Figure 5 The head network Head is a Seek-CD detection head (see Figure 6 shown);

[0048] Step S3, using the training set to train the constructed lightweight human target detection model LightSeek-YOLO, and evaluating the target detection results through evaluation indicators;

[0049] Step S4: Use the test set to test the trained lightweight human target detection model LightSeek-YOLO.

[0050] In step S1, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. Step S1 specifically includes:

[0051] Step S11, obtaining a background image and a human target image;

[0052] Step S12, using the U2Net model to perform background segmentation on the human target image, accurately extract the human foreground image, and remove redundant background;

[0053] Step S13, cropping and denoising the segmented human foreground image to retain only the human body part;

[0054] Step S14, randomly scaling the cropped and denoised human foreground image to a suitable size, and then superimposing it on the background image to generate a composite image;

[0055] Step S15: annotate the synthesized image according to the Yolo format to construct a data set.

[0056] In step S11, a background image is obtained from the AIDER dataset, and a human target image is obtained from the LSP / MPII-MPHB dataset.

[0057] In step S14, the position and size of the human foreground image in the background image are randomly selected to simulate the diversity and complexity of human targets in real disaster scenes.

[0058] In step S2, the backbone network HGnetV2 also includes the HGStem module (see Figure 4 As shown in the figure), DWConV module, SPPF module and C2PSA module, among which the HGStem module, HGBlock module and DWConV module constitute a hybrid gated convolution, which mainly realizes the function of the backbone network HGnetV2. The hybrid gated convolution can perform initial feature extraction and improve the ability to retain high-frequency details.

[0059] The SPPF module is a spatial pyramid pooling layer used to aggregate features at multiple scales. Specifically, the SPPF module uses three-level serial maximum pooling to capture multi-scale contextual information and uses feature splicing to retain the original resolution information to aggregate multi-scale contextual information and improve the model's detection ability for targets of different sizes.

[0060] The C2PSA module is an advanced feature extraction module that combines the CSP structure and the PSA attention mechanism, which can enhance feature extraction capabilities. Specifically, the C2PSA module can fuse channel attention (CA) and spatial attention (SA), enhance the role of key features, retain original information and perform attention calculations.

[0061] The neck network Neck also includes a Concat splicing module and an Upsample module, thus forming an efficient aggregation network that can pay more attention to image details.

[0062] Among them, the Seekds feature downsampling module adjusts the number of channels, retains the most significant local features, and is computationally efficient. It also extracts effective features and then fuses the features. The C3K2 module and the Concat splicing module achieve multi-scale feature map fusion, further improving the model's feature extraction capabilities for maps of different scales. Specifically, the Seekds feature downsampling module consists of a Maxpooling path and a CBS path in parallel. One path first undergoes the Maxpooling operation, then undergoes a 1×1 convolution to adjust the number of channels, thereby retaining the most significant local features and achieving computational efficiency. The other convolution path first adjusts the number of channels, then undergoes a 1×1 convolution, with W2 as the weight and X as the weight. conv1 is the output feature, b2 is the bias formula as follows:

[0063]

[0064] Where X is the input feature; B is the batch size; C out is the number of channels, H is the height of the feature map size; W is the width of the feature map size; is a four-dimensional tensor with the shape

[0065] Through 3×3 convolution operations, this path captures more complex spatial patterns with learnable parameters, achieving feature downsampling, concatenating by channel number, and fusing features at different scales, further focusing on the image details after feature fusion. The resulting image then passes through the C3k2 module, performing this operation at different scales to produce a multi-scale fused feature map.

[0066] It can be seen that the feature maps of multiple scales obtained by the neck network Neck are used as input. Through the lightweight shared convolutional detection head SeekCD, the amount of computation is significantly reduced, the bounding box regression is converted into discrete probability distribution prediction, and the accuracy of small target detection is improved.

[0067] The backbone network HGnetV2 adopts a five-stage hierarchical processing flow. Each stage realizes progressive feature extraction and transformation through a specific module combination. It is divided into input preprocessing stage, primary feature abstraction stage, intermediate feature fusion stage, high-level semantic extraction stage, and multi-order feature enhancement stage.

[0068] The input preprocessing stage uses the HGstem module, replacing the traditional Stem layer, and employing hybrid gated convolution for initial feature extraction, improving the ability to retain high-frequency details. Similarly, a parallel two-branch convolution is sampled. One branch samples a 3×3 depthwise separable convolution to extract spatial features. Stem2b, stem2a, stem3, and stem4 can be understood as ConV operations at different stages to reduce computational complexity. The other branch uses MaxPool to preserve low-level features. The specific formula is as follows:

[0069] X out =stem4(stem3(concat(Maxpool(X1),stem2b(stem2a(X1)))));

[0070] Where, X out is the output feature; concat(·) is the concatenation operation; Maxpool(·) is the maximum pooling operation.

[0071] The other branch samples standard 3×3 convolution to enhance local texture information, effectively preserving the characteristic information of human targets in disaster scenes; at the same time, the CA channel attention mechanism is used to perform channel weighting on the dual-branch output to suppress noise characteristics. After adding the dual-branch features, the ReLU6 activation function is used to output a 48-channel feature map to achieve feature fusion.

[0072] The primary feature abstraction stage receives the output of the input preprocessing stage, processes it using the HGBlock module, constructs the bottom layer of the feature pyramid, and adopts a lightweight residual structure to balance the amount of computation and feature expression capabilities. Specifically, it is achieved through cross-stage partial connections: the input features are divided into two parts, and only half of the channels are convolutionally calculated to reduce redundancy; the shortcut method is enabled based on the complexity of the input features to achieve dynamic residual connections; and a bottleneck structure is used to process features, using 1×1 convolution to compress the number of channels, 3×3 convolution to extract spatial features, and then expanding back to the original number of channels to compress the amount of computation, making the backbone network HGnetV2 lighter, and obtaining a feature map with an output channel number of 128. The feature map undergoes separated spatial convolution and channel fusion convolution of the DWConV module, and the amount of computation is significantly reduced compared to traditional convolution. The specific principle is as follows: the computational amount of DW and PW is added together:

[0073]

[0074] Where, is the square of the convolution kernel size; M is the number of input channels; is the square of the size of the feature map, N is the number of output channels; D k *D F *M is the size of the input feature map; D k *D F *N is the size of the output feature map compared to conventional convolution. The computational complexity of conventional convolution is:

[0075]

[0076] Therefore, the computational reduction ratio of depth-wise separable convolution is:

[0077]

[0078] By reducing the redundant calculation of convolution kernels, lightweight is further achieved.

[0079] In the intermediate feature fusion stage, the feature map is downsampled by the 512-channel DWConv module and then input into three sets of improved HGBlock modules. The input features are compressed by 1×1 convolution channels and then spatial features are extracted by 5×5 convolution. The detailed mathematical principle formula is as follows:

[0080]

[0081] Where, F large To represent the large receptive field features extracted by 5×5 convolution; Conv 5×5 (·) is a 5×5 convolution operation; i and j are the relative coordinate offsets of the convolution kernel; W i,jis the weight parameter of the convolution kernel at position (i, j), which is the parameter that the model needs to learn; X x+i,y+j It is the value of the input feature map XX at position (x+i,y+j), that is, the input pixel or feature value in the local area covered by the convolution kernel.

[0082] In order to achieve enhanced receptive field and improve the detection ability of small targets, a gating mechanism is introduced again to control the activation of the residual path. The gating mechanism formula is as follows:

[0083] i=-2j=-2;

[0084] Dynamically control the activation strength of the residual path to avoid gradient vanishing. Through the above operation, the input channel is changed from 192 to 1024, and the cross-stage feature reorganization is enhanced to enhance feature diversity. The formula is:

[0085] Y=α·X+(1―α)·Conv(X),α=Sigmoid(FC(GAP(X)));

[0086] Where Y is the output feature map; α is the dynamic weight coefficient; Sigmoid(·) is the activation function, which compresses the input value to the interval [0, 1] to ensure the rationality of α; FC(·) is the fully connected layer; GAP(·) is the global average pooling;

[0087] The high-level semantic extraction stage still uses the HGblock module, adopting dynamic deep convolution to adaptively adjust the convolution kernel weights. The feature reconstruction layer enhances cross-channel interaction, and uses 1×1 convolution for channel interaction, replacing the traditional BN layer with parameterized linear transformation to achieve feature reconstruction. Obviously, the multi-order feature enhancement stage is where the spatial pyramid pooling and dual attention mechanism C2PSA module play a role. Through the SPPF module, three-level serial maximum pooling captures multi-scale contextual information, and feature splicing is used to retain the original resolution information to aggregate multi-scale contextual information and improve the model's detection ability for targets of different sizes. The SPEEF multi-scale pooling formula is as follows;

[0088]

[0089] Where, F pool is the output feature map after multi-scale pooling; MaxPool is the maximum pooling.

[0090] The C2PSA module can leverage the fusion of channel attention (CA) and spatial attention (SA) to enhance key features. Input features are split into two paths (path A retains the original information, and path B performs attention calculations). Path B uses multi-head self-attention (MHSA) to calculate spatial correlations, and then uses FFN to enhance nonlinearity. After concatenating the two feature paths, a 1×1 convolution is used to output the final features. C2PSA attention feature analysis is implemented using the Split operation, multi-head attention is processed using the Softmax function, and the output fusion calculation formula is as follows:

[0091] A,B=Split(Conv 1×1 (X));

[0092]

[0093] Y = Conv 1×1 (Concat[A,FFN(Attention(B))]);

[0094] Where Split(·) is to split the input features equally along the channel dimension; Conv 1×1 (·) is a 1×1 convolution; Attention(·) is a multi-head self-attention; Softmax(·) is a normalization function; FFN(·) is a feedforward neural network; Q is the query vector; K is the key vector; V is the value vector; d k is the scaling factor.

[0095] Seek-CD, a lightweight shared convolutional detection head, is particularly well-suited for object detection tasks in computer vision. Through innovative network structure design and computational optimization strategies, this approach significantly reduces computational complexity while maintaining high detection accuracy. Seek-CD primarily consists of a multi-scale feature fusion module, a shared convolutional computation module, and a dual-branch detection head. It utilizes a unique weight sharing mechanism and distributed focus loss (DFL) technology.

[0096] Receive multi-scale feature inputs from the Feature Pyramid Network (FPN), including: high-resolution feature map P3 (for small target detection), medium-resolution feature map P4, and low-resolution feature map P5 (for large target detection). The feature map of each scale first passes through an independent 1×1 convolution layer for channel dimensionality reduction. Through the shared convolution calculation module, the module adopts a two-level structure to achieve computational sharing. In order to standardize the input features and solve the distribution offset problem caused by the resolution difference of different levels of features (P3 / P4 / P5) from FPN / PANet, each scale feature is normalized to a similar numerical range through channel-by-channel normalization to achieve multi-scale feature alignment; at the same time, a dynamic adaptation strategy is adopted, using learnable affine parameters (γ, β) to retain the feature expression ability based on normalization. For example, a larger γ value is given to P3 (small target feature) to enhance high-frequency details. The formula of its input feature normalization module is as follows:

[0097]

[0098] Where P i is the input feature map (P3 / P4 / P5)μ i ; σ i is the mean / standard deviation calculated channel by channel, γ, β are learnable affine parameters; ⊙ is the Hadamard product;

[0099] A 1×1 convolutional dimensionality reduction module utilizes feature channel compression to uniformly project 256 / 512 / 1024-dimensional input features into a HiDC (e.g., 256)-dimensional space, reducing subsequent computational effort. A shared 1×1 convolution kernel implicitly establishes associations between features of different resolutions. This promotes cross-scale feature fusion, allowing high-resolution features (P3) to receive greater gradient weight during backpropagation, balancing multi-scale learning rates and achieving gradient redistribution. The specific implementation formula is as follows:

[0100]

[0101] Where * is the convolution operation, To convolve the weight, For bias top;

[0102] By sharing the grouped convolution layer and 1×1 fusion layer of the convolutional computing module, the number of parameters can be reduced by 75%.

[0103] Maintaining the spatial correlation of features, the regression branch and the classification branch are used to decouple the positioning and classification tasks. Decentralized Flow (DFL) is used to improve regression accuracy. The DFL decoding module converts discrete distributions into continuous coordinates to achieve sub-pixel positioning accuracy. Dynamic mesh generation is used to adapt to different input resolutions and avoid the limitations of predefined anchor points. The specific functions of the dynamic mesh generation module are as follows:

[0104]

[0105] Where a x is the adjusted coordinate of the dynamic grid in the horizontal direction (x-axis); img_width is the actual width of the input image; w is the original coordinate of the current grid unit; s is the scaling factor of the dynamic grid.

[0106] In step S3, training parameters need to be set before training, including:

[0107] Set the training round epoch = 150, batch size batch_size = 32, set the network initial learning rate to 0.01, use the stochastic gradient descent optimizer, and uniformly scale the input image size to 640×640.

[0108] Furthermore, the evaluation indicators include precision, recall rate, mAP@0.5, Params and GFLOPs; the precision is used to evaluate the ratio of positive samples correctly identified as positive samples in the recognized images. The higher the precision, the better the detection effect of the detection model; the recall rate is used to evaluate the proportion of positive samples among all samples detected as positive; the mAP is used to evaluate the average value of the average accuracy AP of multiple categories. The larger the mAP value, the better the performance of the detection model. mAP@0.5 indicates the mAP value when the IOU threshold is 0.5. When the IOU between the predicted box and the annotated box is greater than 0.5, the object is considered to be correctly predicted. Under this premise, the mean AP (mAP) is calculated. The Params and GFLOPs are used to evaluate the lightweight characteristics of the improved model. GFLOPs can measure the amount of floating-point operations, and then compare the computational complexity of different models. The larger the GFLOPs, the more complex the model, and the more data it can "throughput", which can complete complex tasks, but has greater requirements for hardware computing resources. A smaller GFLOPs value indicates that the task or model has lower computational requirements and may be more efficient or lightweight. Since the target task of this model is human target detection in disaster scenarios, it needs to be deployed on hardware with high computing power, such as drones. The model is required to ensure lightweight and real-time performance without significantly affecting the recognition accuracy of the model.

[0109] Recombination Figure 7 As shown, the present invention also provides a lightweight human target detection system based on improved YOLOv11 for executing the method described, including a data set construction module 1, a model construction module 2, a training module 3 and a testing module 4.

[0110] The dataset construction module 1 is used to obtain a background image and a human target image, extract a human foreground image from the human target image using the U2Net model, and superimpose the human foreground image onto the background image to generate a synthetic image, annotate the synthetic image according to the Yolo format to construct a dataset, and divide the dataset into a training set, a validation set, and a test set.

[0111] The model construction module 2 is used to construct a lightweight human target detection model LightSeek-YOLO, whose network architecture includes a backbone network HGnetV2, a neck network Neck and a head network Head. The backbone network HGnetV2 includes a feature extraction module HGBlock composed of a LightConV lightweight convolution module and a CBS convolution module; the neck network Neck includes a SeekDS downsampling module combined with a C3K2 module and placed before the C3K2 module; the head network Head is a Seek-CD detection head.

[0112] The training module 3 is used to train the constructed lightweight human target detection model LightSeek-YOLO using the training set.

[0113] The testing module 4 is used to test the trained lightweight human target detection model LightSeek-YOLO using the test set and evaluate the target detection results.

[0114] See also Figure 8 As shown, an embodiment of the present invention further provides an electronic device 600, which includes a processor 601, a memory 602, and a program or instruction stored in the memory 602 and executable on the processor 601. When the program or instruction is executed by the processor 601, each process of the embodiment of the above-mentioned lightweight human target detection method based on the improved YOLOv11 is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0115] It should be noted that the electronic devices in the embodiments of the present invention include the mobile electronic devices and non-mobile electronic devices mentioned above.

[0116] Figure 9 The present invention is a hardware structure diagram of an electronic device.

[0117] The electronic device 700 includes but is not limited to components such as a radio frequency unit 701 , a network module 702 , an audio output unit 703 , an input unit 704 , a sensor 705 , a display unit 706 , a user input unit 707 , an interface unit 708 , a memory 709 , and a processor 710 .

[0118] Those skilled in the art will understand that the electronic device 700 may also include a power source (such as a battery) to power each component, and the power source may be logically connected to the processor 710 through a power management system, thereby implementing functions such as charging, discharging, and power consumption management through the power management system. Figure 9 The electronic device structure shown in the figure does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently, which will not be repeated here.

[0119] It should be understood that in embodiments of the present invention, the input unit 704 may include a graphics processing unit (GPU) 7041 and a microphone 7042. The graphics processor 7041 processes image data of still images or videos obtained by an image capture device (such as a camera) in video capture mode or image capture mode. The display unit 706 may include a display panel 7061, which may be configured in the form of a liquid crystal display, an organic light-emitting diode, or the like. The user input unit 707 includes a touch panel 7071 and other input devices 7072. The touch panel 7071 is also called a touch screen. The touch panel 7071 may include two parts: a touch detection device and a touch controller. Other input devices 7072 may include, but are not limited to, a physical keyboard, function keys (such as volume control keys, power keys, etc.), a trackball, a mouse, and a joystick, which will not be described in detail here. The memory 709 may be used to store software programs and various data, including, but not limited to, applications and operating systems. The processor 710 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and applications, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 710.

[0120] An embodiment of the present invention also provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the various processes of the above-mentioned lightweight human target detection method embodiment based on the improved YOLOv11 are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0121] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk.

[0122] An embodiment of the present invention further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned lightweight human target detection method embodiment based on the improved YOLOv11, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0123] It should be understood that the chip mentioned in the embodiment of the present invention can also be called a system-on-chip, a system-on-chip, a chip system, or a system-on-chip chip, etc.

[0124] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0125] Furthermore, it should be noted that the scope of the methods and systems of the present invention is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in reverse order, depending on the functions involved. For example, the methods described may be performed in an order different from that described, and various steps may be added, omitted, or combined. Furthermore, features described with reference to certain examples may be combined in other examples.

[0126] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are protected by the present invention.

Claims

1. A lightweight human target detection method based on improved YOLOv11, characterized in that: The steps include: Step S1: obtain a background image and a human target image, use the U2Net model to extract a human foreground image from the human target image, and superimpose the human foreground image onto the background image to generate a composite image. The composite image is annotated in YOLO format to construct a dataset, and the dataset is divided into a training set and a test set. Step S2, constructing a lightweight human target detection model LightSeek-YOLO, whose network architecture includes a backbone network HGnetV2, a neck network Neck, and a head network Head. The backbone network HGnetV2 includes a feature extraction module HGBlock composed of a LightConV lightweight convolution module and a CBS convolution module; the neck network Neck includes a SeekDS downsampling module combined with a C3K2 module and placed before the C3K2 module; the head network Head is a Seek-CD detection head; Step S3, using the training set to train the constructed lightweight human target detection model LightSeek-YOLO, and evaluating the target detection results through evaluation indicators; Step S4: Use the test set to test the trained lightweight human target detection model LightSeek-YOLO.

2. The method according to claim 1, characterized in that Step S1 specifically includes: Step S11, obtaining a background image and a human target image; Step S12, using the U2Net model to perform background segmentation on the human target image, accurately extract the human foreground image, and remove redundant background; Step S13, cropping and denoising the segmented human foreground image to retain only the human body part; Step S14, randomly scaling the cropped and denoised human foreground image to a suitable size, and then superimposing it on the background image to generate a composite image; Step S15: annotate the synthesized image according to the Yolo format to construct a data set.

3. The method according to claim 2, characterized in that In step S11, a background image is obtained from the AIDER dataset, and a human target image is obtained from the LSP / MPII-MPHB dataset.

4. The method according to claim 1, wherein In step S1, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:

1.

5. The method according to claim 1, characterized in that In step S2, the backbone network HGnetV2 also includes an HGStem module, a DWConV module, an SPPF module, and a C2PSA module. The HGStem module, the HGBlock module, and the DWConV module mainly constitute the functions of the backbone network HGnetV2; the SPPF module is a spatial pyramid pooling layer used to aggregate features at multiple scales; the C2PSA module is an advanced feature extraction module that combines the CSP structure and the PSA attention mechanism, which can enhance the feature extraction capability.

6. The method according to claim 5, characterized in that In step S3, training parameters need to be set before training, including: Set the training round epoch = 150, batch size batch_size = 32, set the network initial learning rate to 0.01, use the stochastic gradient descent optimizer, and uniformly scale the input image size to 640×640.

7. A lightweight human target detection system based on improved YOLOv11 for executing the method according to any one of claims 1 to 6, characterized in that: include: The dataset construction module is used to obtain background images and human target images, extract the human foreground image from the human target image using the U2Net model, and superimpose the human foreground image on the background image to generate a composite image. The composite image is annotated in the YOLO format to construct a dataset, and the dataset is divided into a training set, a validation set, and a test set. A model construction module is used to build a lightweight human target detection model LightSeek-YOLO. Its network architecture includes a backbone network HGnetV2, a neck network Neck, and a head network Head. The backbone network HGnetV2 includes a feature extraction module HGBlock composed of a LightConV lightweight convolution module and a CBS convolution module; the neck network Neck includes a SeekDS downsampling module combined with a C3K2 module and placed before the C3K2 module; the head network Head is a Seek-CD detection head; The training module is used to train the constructed lightweight human target detection model LightSeek-YOLO using the training set; The testing module is used to test the trained lightweight human target detection model LightSeek-YOLO using the test set and evaluate the target detection results.

Images