An infrared human posture estimation method and system based on feature cross-domain migration

Abstract

This invention discloses an infrared human pose estimation method and system based on cross-domain feature transfer. A domain transfer module is inserted into each feature extraction layer of the human pose estimation model. This module is specifically designed to learn the transformation mapping relationship from the source domain feature space to the target domain feature space. By embedding the domain transfer module in each feature extraction layer of the network, domain transfer can be performed simultaneously at different abstraction levels of the network, achieving multi-scale and multi-level feature alignment. The domain transfer module is then trained using an infrared image dataset to learn the domain transformation from the source domain to the target domain. This invention solves the technical problems of traditional cross-domain pose estimation methods, such as the need for large amounts of labeled data, low training efficiency, and susceptibility to overfitting, thus improving the accuracy and precision of pose estimation after cross-domain adjustments.

Description

Technical Field

[0001] This invention belongs to the field of infrared human pose estimation technology, and relates to an infrared human pose estimation method and system based on feature cross-domain transfer. Background Technology

[0002] Human pose estimation, a core task in computer vision, aims to identify and locate key joints of the human body from images or videos to construct the human skeleton structure. With the rapid development of deep learning technology, human pose estimation methods based on convolutional neural networks have made breakthrough progress, significantly improving the accuracy and robustness of pose estimation. These methods have achieved near-human performance on large-scale public datasets such as COCO and MPII.

[0003] However, existing human pose estimation methods are primarily designed and optimized for visible light images. Visible light images rely on ambient lighting conditions, forming images by capturing the reflection of visible light from objects. Under sufficient lighting and favorable environmental conditions, visible light images can provide rich texture, color, and detail information, enabling deep learning-based pose estimation models to accurately identify and locate key human points. However, in practical applications, especially in critical scenarios such as military reconnaissance, emergency rescue, nighttime surveillance, and vital sign monitoring under adverse weather conditions, visible light imaging is often severely limited. Insufficient lighting, fog, rain, snow, and other adverse environmental conditions can lead to a sharp decline in visible light image quality, or even the inability to obtain effective image information, thus rendering visible light image-based human pose estimation methods ineffective.

[0004] Infrared imaging, as a passive imaging technology, forms images by detecting the thermal radiation emitted by an object itself, independent of external lighting conditions. Infrared images reflect the temperature distribution and thermal radiation characteristics of an object, enabling them to operate in all weather conditions, at all times, and in adverse weather conditions, providing crucial technical support for the aforementioned key applications. Especially in scenarios such as nighttime operations, disaster relief, and medical monitoring, infrared imaging technology possesses irreplaceable advantages.

[0005] However, infrared images and visible light images differ fundamentally in imaging principles, visual features, and data distribution; this difference is known as "domain difference." Therefore, directly applying models trained on visible light data to infrared images often fails to effectively extract key information, leading to a significant drop in pose estimation accuracy. Furthermore, acquiring high-quality infrared human pose estimation annotation data is extremely costly and challenging. In contrast, visible light image data is abundant, with public datasets like COCO containing hundreds of thousands of annotated images. This significant difference in data resources provides a natural advantage for cross-domain transfer learning. It allows leveraging the abundant annotated data and trained high-performance models in the source domain to transfer learned general features and structural knowledge to the target domain, thereby achieving high performance on limited target domain annotated data and addressing the problems of scarce target domain data and high annotation costs.

[0006] However, transfer learning from the visible light domain to the infrared domain faces numerous technical challenges. Traditional transfer learning methods mainly fall into two categories: domain adaptation and fine-tuning. Domain adaptation methods typically require simultaneous access to data from both the source and target domains, reducing domain differences through adversarial training and other methods. However, these methods are computationally expensive and have high requirements for datasets. Fine-tuning methods train the entire model using target domain data on top of a pre-trained model, i.e., directly using infrared data to fine-tune the entire pre-trained model. While this can adapt to the infrared domain to some extent, it requires updating all model parameters, consuming significant computational resources, resulting in long training times, and is prone to overfitting, especially when infrared labeled data is limited.

[0007] Therefore, an efficient domain transfer method is needed. This method should be able to make full use of the general features and structural knowledge in the pre-trained model of the source domain, and at the same time effectively reduce the domain difference between the source domain and the target domain through a targeted feature domain transfer mechanism, avoid the computational burden caused by full parameter fine-tuning, and achieve excellent performance while maintaining parameter efficiency. Summary of the Invention

[0008] The purpose of this invention is to address the problem that existing technologies suffer from domain differences between infrared and visible light images. Directly using infrared data to fine-tune the entire source domain pre-trained model consumes a lot of computational resources, has a long training time, and is prone to overfitting, resulting in low accuracy of pose estimation. This invention provides an infrared human pose estimation method and system based on feature cross-domain transfer.

[0009] To achieve the above objectives, the present invention employs the following technical solution: An infrared human pose estimation method based on feature cross-domain transfer includes the following steps: Obtain a trained human pose estimation model; A domain transfer module is inserted into each feature extraction layer of the human pose estimation model to obtain an infrared image dataset. The domain transfer module is trained based on the infrared image dataset to obtain the infrared image human pose estimation model after cross-domain transfer. The infrared image of the human body to be identified is acquired, and the pose of the human body to be identified is estimated based on the infrared image human pose estimation model, so as to obtain human body key point detection and pose estimation.

[0010] A further improvement of the present invention is that: The domain transfer module is used to generate scaling and offset parameters corresponding to feature transfer through convolution operations based on the input feature map, and to perform element-wise scaling and offset transformation on the input feature map.

[0011] The domain transfer module includes two parallel convolutional branches, one of which is used to learn feature scaling parameters and generate a scaling factor. Another convolutional branch is used to learn the feature offset parameters and generate the offset. .

[0012] The domain transfer module is trained based on the infrared image dataset to obtain a cross-domain transferred infrared image human pose estimation model, including: During training, all parameters of the source domain human pose estimation model are frozen, and only the parameters of the domain transfer module are trained to learn the domain transformation from the source domain to the target domain.

[0013] The process of obtaining the trained human pose estimation model includes: Obtain a visible light image dataset and train a human pose estimation network using the visible light image dataset.

[0014] During training, the output of the domain transfer module is iteratively optimized using the back gradient of the loss function.

[0015] An infrared human pose estimation system based on feature cross-domain transfer includes: The source domain pre-trained pose estimation model acquisition module is used to acquire a trained human pose estimation model. The infrared pose estimation domain transfer training module is used to insert the domain transfer module into each feature extraction layer of the human pose estimation model, obtain the infrared image dataset, train the domain transfer module based on the infrared image dataset, and obtain the infrared image human pose estimation model after cross-domain transfer. The infrared image human pose estimation model after cross-domain migration is used to acquire the infrared image of the human body to be identified. Based on the infrared image human pose estimation model, the pose of the infrared image of the human body to be identified is estimated, and the human body key point detection and pose estimation are obtained.

[0016] A computer program product includes a computer program that, when executed by a processor, implements any of the methods described above.

[0017] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of any of the methods described above.

[0018] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the methods described herein.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses an infrared human pose estimation method based on cross-domain feature transfer. A domain transfer module is inserted into each feature extraction layer of the human pose estimation model. This module is specifically designed to learn the transformation mapping relationship from the source domain feature space to the target domain feature space. By embedding the domain transfer module in each feature extraction layer of the network, domain transfer can be performed simultaneously at different abstraction levels of the network, achieving multi-scale and multi-level feature alignment. The domain transfer module is then trained using an infrared image dataset to learn the domain transformation from the source domain to the target domain, improving training efficiency, reducing computational resource consumption, mitigating the risk of overfitting, and enhancing the model's accuracy and the precision of human pose estimation. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is an overall flowchart disclosed in an embodiment of the present invention; Figure 2 This is a block diagram of the feature cross-domain transfer model of the embedded domain transfer module disclosed in an embodiment of the present invention; Figure 3 This is a comparison diagram of dual-path reasoning for the same model disclosed in an embodiment of the present invention; Figure 4 This is a structural diagram of the domain migration module SFT disclosed in an embodiment of the present invention; Figure 5 The HRNet network disclosed in this embodiment of the invention has a domain migration module embedded in Layer 1 and above. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0027] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0028] The present invention will now be described in further detail with reference to the accompanying drawings: See Figures 1 to 5This paper discloses an infrared human pose estimation method based on feature cross-domain transfer, which includes the following steps: Step 1: Construct and pre-train a source domain pre-trained model for human pose estimation based on visible light image dataset.

[0029] Based on a large-scale visible light image dataset, a visible light human pose estimation model is pre-trained to obtain a visible light human pose estimation pre-trained model. Specifically, in this embodiment, the model is trained based on the COCO dataset.

[0030] The pre-trained model disclosed in this embodiment is trained on visible light data, learns rich general feature representations and spatial structure knowledge, and is directly loaded into the weights of the source domain pre-trained model. This makes full use of the existing knowledge accumulation in the source domain and avoids the difficulty of training from scratch when the target domain data is scarce. It serves as the source domain pre-trained model for subsequent cross-domain transfer learning.

[0031] Step 2: Insert a domain transfer module into the source domain pre-trained model to learn infrared domain feature transformations.

[0032] Lightweight domain transfer modules are embedded in each feature extraction layer of the pre-trained network. These modules are specifically designed to learn the transformation mapping relationship from the source domain feature space to the target domain feature space.

[0033] This step embeds domain transfer modules into each feature extraction layer of the network, enabling simultaneous domain transfer at different abstraction levels of the network, thus achieving multi-scale and multi-level feature alignment.

[0034] Step 3: Supervised training of the domain transfer module is performed using the infrared image dataset, where all parameters of the source domain pre-trained model are frozen, meaning that none of the parameters of the source domain pre-trained model are updated during the training process. The model is trained under supervision using an infrared image dataset and its corresponding human body key point annotation information.

[0035] Furthermore, during training, all parameters of the source domain pre-trained model are frozen, keeping them unchanged throughout the training process. This fully preserves the general knowledge learned by the source domain pre-trained model. Simultaneously, only the parameters of the embedded domain transfer modules are trained; these parameters are specifically designed to learn the domain transformation from the source domain to the target domain. This significantly improves training efficiency, reduces computational resource consumption, and lowers the risk of overfitting. See [link to relevant documentation]. Figure 2 .

[0036] Furthermore, during training, a human pose estimation network suitable for infrared images is obtained through iterative optimization using backpropagation gradients. During training, the model is forward-propagated using an infrared image dataset to calculate the loss between the predicted results and the true labels. Through iterative optimization using backward gradient descent, the loss gradient is backpropagated from the output layer to the domain transfer module, updating only the parameters of the domain transfer module while keeping all parameters of the source domain pre-trained model unchanged. Through multiple iterations, the domain transfer module gradually learns the transformation mapping from the source domain feature space to the target domain feature space, enabling the model's performance on the target domain data to continuously approach or even reach the performance of a model trained directly on the target domain data. This ultimately yields a human pose estimation network suitable for infrared images, achieving high performance in infrared human pose estimation tasks.

[0037] Furthermore, this embodiment also discloses the same model, direct prediction in the original domain and cross-domain prediction after introducing the domain adaptation module, see [link to relevant documentation]. Figure 3 A comparison of the overall inference process: The top shows the visible light pre-trained model. Inputting a visible light image, the pre-trained visible light pose estimation model sequentially generates features for each stage and directly outputs the pose result. The bottom shows the domain transfer process. Inputting an infrared image, it uses the same pre-trained visible light model, inserting a domain adaptation module at each stage to adapt features across domains, obtaining adapted features, and then outputting the pose result. The two processes are presented side-by-side, highlighting that "the input domains are different, the main model is the same, the domain adaptation module is responsible for cross-domain adaptation, and the output is the same pose estimate."

[0038] The present invention also discloses a specific embodiment: Step 1: Design the domain migration module: The domain transfer module is the core component of this invention, used to implement feature transformation from the visible light domain to the infrared domain. This embodiment uses the SFT (Spatial Feature Transform) module. The design concept of the SFT module is to learn scaling factors and offsets through convolution operations to perform spatial transformation on the features, enabling the feature representation in the visible light domain to adapt to the feature distribution in the infrared domain. For detailed design of the SFT module, please refer to [link to SFT module details]. Figure 4 .

[0039] The SFT module employs a lightweight design with few parameters and low computational overhead. This module consists of two parallel convolutional branches: One branch learns the feature scaling parameter (scale), and the other branch learns the feature shift parameter (shift). Specifically, for the input feature map x, the SFT module first generates the scaling factor through two parallel convolutional layers. and offset Then, the features are transformed by element-wise multiplication and addition:

[0040] Here, ⊙ represents element-wise multiplication. In this way, the SFT module can learn spatially relevant transformation parameters and adaptively modulate features at different spatial locations, thereby adapting to the feature differences between the visible and infrared domains.

[0041] In this embodiment, the SFT module has the following design features: (1) Fewer parameters: By using 1×1 convolution kernels and fewer channels, the number of parameters in the SFT module is much smaller than the number of parameters in the entire model; (2) High computational efficiency: The lightweight 1×1 convolution operation is used, which has low computational overhead and will not significantly increase the inference time of the model; (3) Good flexibility: The SFT module can be embedded into any feature extraction layer of the network to adapt to the feature transformation needs of different levels; (4) High versatility: The design of the SFT module does not depend on a specific network architecture and can be applied to various deep neural networks, such as HRNet, U-Net, etc.

[0042] Step 2: Embed a domain transfer module into the visible light human pose estimation network HRNet is a typical visible light human pose estimation network. It features a high-resolution network and achieves high-precision pose estimation by maintaining high-resolution feature representations throughout the network. This embodiment uses HRNet as an example, embedding the domain transfer module into each layer up to Layer 1 to achieve progressive feature transformation layer by layer.

[0043] HRNet first uses a "stem" to perform preliminary feature extraction on the input, followed by four consecutive stages (Layer 1, Stage 2-Stage 4). Layer 1 is the first high-resolution residual stage after the stem, consisting of four cascaded Bottleneck units that maintain a single resolution. Stages 2-4 progressively introduce multi-resolution parallel branches and continuously exchange information. Therefore, "Layer 1" refers to the first stage of HRNet after the stem, and "before Layer 1" refers to the stem; the subsequent Stages 2 / 3 / 4 correspond to the later stages as the network deepens.

[0044] Furthermore, for HRNet networks embedding domain migration modules in Layer 1 and earlier, see [link to relevant documentation]. Figure 5Layer 1 Embedding: In HRNet's Layer 1 stage (the first high-resolution residual stage), the domain transfer module is embedded after the convolutional layers of the last Bottleneck module. Specifically, for each convolutional layer of the last Bottleneck module, a domain transfer module is added after it to transform the features of the convolutional output.

[0045] Layer-by-layer progressive feature transformation: By embedding a domain transfer module in each layer, a layer-by-layer progressive feature transformation is achieved. Each layer's domain transfer module receives the features of that layer as input, generates scaling factors and offsets through convolution operations, and transforms the features. As the network depth increases, the features gradually adapt from the visible light domain to the infrared domain, ultimately outputting a feature representation suitable for the infrared domain.

[0046] Step 3: Training Process for Cross-Domain Transfer Learning This embodiment details the complete training process for cross-domain transfer learning, as shown below: Step 3.1: Load visible light pre-trained weights The weights of the HRNet model, pre-trained on the COCO visible light dataset, are directly loaded.

[0047] Step 3.2: Parameter Freeze Iterate through all the parameters of the model and freeze all the parameters of HRNet.

[0048] Step 3.3: Model Training The model was trained using an infrared human pose estimation dataset.

[0049] Training configuration: Optimizer: Use Adam optimizer, learning rate 5e-4, train for 210 epochs, update only the domain transfer module parameters.

[0050] Loss function: KeypointMSELoss loss.

[0051] Through the above three steps, the cross-domain transfer learning training process based on the HRNet network and the domain transfer module is completed. After training, the model can transfer the knowledge of the HRNet network pre-trained in the visible light domain to the infrared domain, achieving high performance on the infrared human pose estimation task.

[0052] Furthermore, this embodiment also discloses experimental results and analysis. The method of this invention demonstrates excellent performance in infrared human pose estimation tasks. Taking an input resolution of 225x192 as an example, after embedding a domain transfer module (Layer1_before) before Layer 1 of the HRNet network, the average precision (AP) of keypoint detection is improved from 0.599 at the baseline (Low) to 0.770, an improvement of 28.5% compared to the baseline (Low). Further embedding a domain transfer module (Layer1&prior) at or before Layer 1 further improves the AP to 0.830, an improvement of 38.6% compared to the baseline (Low), reaching 89.1% of the baseline (High). Among them, AP50 achieved near-baseline (High) performance, while AP75 improved by 37.9% and 46.1% respectively under Layer1_before and Layer1&prior configurations, reaching 90.4% and 95.7% of the baseline (High) respectively; AR improved by 25.8% and 34.3% respectively, with both AR50 and AR75 reaching near-baseline (High) performance. While achieving the above-mentioned excellent performance, the number of parameters in Layer1_before and Layer1&prior only increased by 0.2% and 0.4% respectively, while the computational load increased by 3.5% and 5.4% respectively, achieving a balance between performance improvement and parameter efficiency.

[0053] This invention achieves efficient cross-domain transfer with minimal additional computational overhead. Under Layer1_before and Layer1&prior configurations, this method requires only 0.2% and 0.4% new parameters, respectively, to complete the cross-domain transfer, effectively reducing computational resource consumption and the risk of overfitting. With minimal parameter additions, this method effectively transfers knowledge from a visible-domain pre-trained model to the infrared domain, achieving near-baseline (High) performance and demonstrating efficient parameter transfer learning. See Tables 1 and 2 for details.

[0054] Table 1 Evaluation Indicators

[0055] Table 2 Computational and Parameter Quantities

[0056] The infrared image human pose estimation model disclosed in this implementation has the following advantages in training: (1) Fewer parameters: The domain transfer module is designed to be lightweight and has fewer parameters, usually accounting for only a small portion of the total number of parameters in the model. Compared with the full model fine-tuning method, the number of parameters is greatly reduced, significantly reducing the model training complexity and storage requirements.

[0057] (2) High training efficiency: Since only the parameters of the domain transfer module are trained, all parameters of the pre-trained model are frozen. The gradients of these parameters do not need to be calculated during training, which reduces the amount of computation and improves training efficiency.

[0058] (3) Excellent performance: Through feature transformation of the domain transfer module, the knowledge of the pre-trained model in the source domain can be effectively transferred to the target domain. Experimental results show that, on the target domain task, the performance of the method of the present invention can approach that of the model trained directly on the target domain data with a significantly reduced number of parameters.

[0059] (4) Simple implementation: The method of the present invention is simple to implement. It only requires embedding a domain transfer module in the pre-trained model, and then freezing the parameters and training. It does not require complex domain adaptation algorithms, nor does it require simultaneous access to data in the source domain and the target domain, resulting in low implementation cost.

[0060] (5) High versatility: The domain transfer module is designed to be universal and can be embedded into any human pose estimation network based on deep learning structure to achieve cross-domain transfer learning.

[0061] (6) Less data requirements: Since the knowledge of the source domain pre-trained model is fully utilized, the method of the present invention has less requirements for the target domain labeled data, and can achieve high performance on limited labeled data, reducing the cost of data acquisition and labeling.

[0062] This embodiment also discloses an infrared human pose estimation system based on feature cross-domain transfer, including: The source domain pre-trained pose estimation model acquisition module is used to acquire a trained human pose estimation model. The infrared pose estimation domain transfer training module is used to insert the domain transfer module into each feature extraction layer of the human pose estimation model, obtain the infrared image dataset, train the domain transfer module based on the infrared image dataset, and obtain the infrared image human pose estimation model after cross-domain transfer. The infrared human pose estimation model after cross-domain migration is used to acquire the infrared image of the human body to be identified. Based on the infrared image human pose estimation model, the pose of the infrared image of the human body to be identified is estimated, and the human body key point detection and pose estimation are obtained.

[0063] This embodiment discloses a globally frozen source domain backbone combined with locally trainable domain transfer: directly loading a high-precision backbone network pre-trained on large-scale visible light data, keeping all original parameters and structures unchanged, and inserting a lightweight domain transfer module only into the key feature layer of the backbone. Cross-domain transfer can be completed with a small number of trainable parameters without accessing the source domain data or performing full fine-tuning.

[0064] Secondly, the spatial feature-modulated domain transfer structure: the domain transfer module reshapes the features of each layer by learning position-sensitive scaling and offset, and achieves the mapping of visible light and infrared distribution with very few parameters, which significantly reduces training / computing costs while maintaining the accuracy of human keypoint estimation.

[0065] A schematic diagram of a terminal device according to an embodiment of the present invention. The terminal device of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.

[0066] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.

[0067] The terminal device can be a desktop computer, laptop computer, cloud server, or other device with strong computing power. The terminal device may include, but is not limited to, a processor and memory.

[0068] The optimal choice for the processor is a multi-core high-speed central processing unit (CPU).

[0069] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.

[0070] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0071] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An infrared human pose estimation method based on feature cross-domain transfer, characterized in that, Includes the following steps: Obtain a trained human pose estimation model; A domain transfer module is inserted into each feature extraction layer of the human pose estimation model to obtain an infrared image dataset. The domain transfer module is trained based on the infrared image dataset to obtain the infrared image human pose estimation model after cross-domain transfer. The infrared image of the human body to be identified is acquired, and the pose of the human body to be identified is estimated based on the infrared image human pose estimation model, so as to obtain human body key point detection and pose estimation.

2. The infrared human pose estimation method based on feature cross-domain transfer as described in claim 1, characterized in that, The domain transfer module is used to generate scaling and offset parameters corresponding to feature transfer through convolution operations based on the input feature map, and to perform element-wise scaling and offset transformation on the input feature map.

3. The infrared human pose estimation method based on feature cross-domain transfer as described in claim 2, characterized in that, The domain transfer module includes two parallel convolutional branches, one of which is used to learn feature scaling parameters and generate a scaling factor. Another convolutional branch is used to learn the feature offset parameters and generate the offset. .

4. The infrared human pose estimation method based on feature cross-domain transfer according to claim 1, characterized in that, The domain transfer module is trained based on the infrared image dataset to obtain a cross-domain transferred infrared image human pose estimation model, including: During training, all parameters of the source domain human pose estimation model are frozen, and only the parameters of the domain transfer module are trained to learn the domain transformation from the source domain to the target domain.

5. The infrared human pose estimation method based on feature cross-domain transfer according to claim 1, characterized in that, The process of obtaining the trained human pose estimation model includes: Obtain a visible light image dataset and train a human pose estimation network using the visible light image dataset.

6. The infrared human pose estimation method based on feature cross-domain transfer according to claim 1, characterized in that, During training, the output of the domain transfer module is iteratively optimized using the back gradient of the loss function.

7. An infrared human pose estimation system based on feature cross-domain transfer, characterized in that, include: The source domain pre-trained pose estimation model acquisition module is used to acquire a trained human pose estimation model. The infrared pose estimation domain transfer training module is used to insert the domain transfer module into each feature extraction layer of the human pose estimation model, obtain the infrared image dataset, train the domain transfer module based on the infrared image dataset, and obtain the infrared image human pose estimation model after cross-domain transfer. The infrared image human pose estimation model after cross-domain migration is used to acquire the infrared image of the human body to be identified. Based on the infrared image human pose estimation model, the pose of the infrared image of the human body to be identified is estimated, and the human body key point detection and pose estimation are obtained.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-6.

Images