A low-resolution line-of-sight estimation method fusing global and local features

Abstract

The application discloses a low-resolution gaze estimation method fusing global features and local features, comprising: acquiring a frame image through a camera device; detecting and positioning a face region, a left eye region and a right eye region in the frame image, and cutting the frame image of the face to obtain a face image, a left eye image and a right eye image; extracting global features of the face in the face image by using a face feature extraction module based on super-resolution and convolution transformer; extracting left eye local features and right eye local features in the left eye image and the right eye image by using an eye attention reinforced feature extraction module; fusing the global features of the face image, the left eye local features and the right eye local features by using a gaze estimation module fusing global and local features to obtain a binocular gaze direction; the effect of low-resolution gaze estimation is improved, more robust features are obtained, reliable gaze estimation is performed, and the gaze estimation performance of a low-resolution scene is improved.

Description

Technical Field

[0001] This invention relates to the field of pattern recognition, and more particularly to a low-resolution gaze estimation method that integrates global and local features. Background Technology

[0002] The main task of gaze estimation is to predict the direction of human gaze or the location of the gaze point. It has broad application prospects in virtual reality, assisted driving, and human-computer interaction, attracting increasing attention from researchers. However, existing research often assumes that the input image is an ideal high-resolution image, ignoring the prevalence of low-resolution images in real-world applications. Gaze estimation at low resolution faces challenges such as difficulty in feature extraction and loss of gaze-related features, which collectively lead to a significant performance degradation. Therefore, research on gaze estimation in low-resolution scenes is of great significance, not only expanding the applicability of gaze estimation techniques but also improving their application value in real-world scenarios. Traditional gaze estimation methods are poor at extracting features from low-resolution images, and traditional convolutional networks suffer from information loss during feature extraction, further contributing to poor performance. Summary of the Invention

[0003] Purpose of the invention: To overcome the shortcomings of existing technologies, this invention provides a low-resolution gaze estimation method that integrates global and local features. It extracts global facial features and local features of the left and right eyes, and then fuses these features using a gaze estimation module that integrates global and local features to obtain the gaze angle direction. This improves the performance of low-resolution gaze estimation, obtains more robust features, enables reliable gaze estimation, and enhances gaze estimation performance in low-resolution scenes.

[0004] Technical Solution: To achieve the above objectives, the present invention provides a low-resolution gaze estimation method that integrates global and local features, comprising the following steps:

[0005] Step 1: Acquire frame images using a camera device;

[0006] Step 2: Detect and locate the face region, left eye region, and right eye region in the frame image, and crop the face frame image to obtain the face image, left eye image, and right eye image;

[0007] Step 3: Use a face feature extraction module based on super-resolution and convolutional transformer to extract global face features from the face image;

[0008] Step 4: Use the human eye attention enhancement feature extraction module to extract local features of the left eye and the right eye from the left eye image and the right eye image;

[0009] Step 5: The gaze estimation module, which integrates global and local features, fuses the global features of the face image, the local features of the left eye, and the local features of the right eye to obtain the gaze direction of both eyes.

[0010] Furthermore, in step three, the shallow feature extraction module of the SwinIR image super-resolution model is used to extract the face image. The shallow features will The low-resolution image is converted into an enhanced dimensional feature space characterized by C channels. The calculation process is as follows:

[0011] ;

[0012] In the formula, These are shallow features. This represents the transformation operation of the convolutional layer;

[0013] The deep feature extraction module of the image super-resolution model SwinIR includes K residual Swin Transformer blocks coupled to a CNN and a convolutional layer. The Swin Transformer blocks sequentially calculate the transition features. The calculation process is as follows:

[0014] ;

[0015] In the formula, Represented as the i-th RSTB block;

[0016] Deep features are extracted using convolutional layers. The calculation process is as follows:

[0017] ;

[0018] In the formula, This represents the convolution operation used to extract features from a convolutional layer.

[0019] Furthermore, the high-quality image reconstruction module of the SwinIR image super-resolution model combines shallow and deep features to reconstruct a high-quality image and output a high-resolution image; the calculation process is as follows:

[0020] ;

[0021] In the formula, This represents the reconstruction operation of the high-quality image reconstruction module.

[0022] Furthermore, a ResNet18 network is used to extract high-resolution image features from the high-resolution image, and these high-resolution image features are flattened into vector features of two-dimensional tensors. The calculation process is as follows:

[0023] ;

[0024] In the formula, Feature extraction operations for the ResNet18 network This is a vector flattening operation;

[0025] Category labels generated for each input face image sample With vector features The above positions are concatenated to generate position codes, which are then processed by a Transformer encoder. Perform splicing to obtain features :

[0026] ;

[0027] In the formula, This is represented as a feature concatenation operation;

[0028] Features Input the vector into the Transformer to extract global face features, and select the first vector as the global face feature. The calculation process is as follows:

[0029] ;

[0030] In the formula, This is represented as the Transformer feature extraction operation.

[0031] Furthermore, in step four, the human eye attention enhancement feature extraction module uses a residual network ResNet18 to extract local features from the left and right eye images. The ResNet18 network includes an initial layer, convolutional layers, batch normalization layers, ReLU activation function, and max pooling layers; the network inputs features from the images of both eyes. The height and width of the input features are then subjected to adaptive average pooling, and then average pooling and max pooling are performed on the height and width of the input features respectively to obtain pooled features. The calculation process is as follows:

[0032] ;

[0033] ;

[0034] In the formula, Represented as adaptive average pooling, subscript This is represented as an operation on height and width. This is represented as an average pooling operation. This is represented as a max pooling operation;

[0035] The pooling features are concatenated and merged along the height and width dimensions respectively to obtain the height and width features:

[0036] ;

[0037] In the formula, This is represented as a vector concatenation operation;

[0038] The obtained height and width features are then combined and processed to obtain a comprehensive spatial feature. :

[0039] ;

[0040] In the formula, This is represented as a vector concatenation operation.

[0041] Furthermore, regarding the comprehensive spatial characteristics Perform one-dimensional convolution and batch normalization operations, then normalize the convolutional features using the Sigmoid activation function to generate a spatial attention map. The calculation process is as follows:

[0042] ;

[0043] In the formula, Represented as an activation function, This is represented as a convolution operation. This is represented as a batch normalization operation;

[0044] Spatial attention map By performing a weighted multiplication with the original vector, a feature map of the spatial location is obtained, which in turn yields the local features of the left eye. and local features of the right eye :

[0045] ;

[0046] In the formula, This represents element-wise multiplication.

[0047] Furthermore, in step five, the gaze estimation module, which fuses global and local features, obtains the global facial features. Local features of the left eye and local features of the right eye The features are then stitched together; after stitching together the global facial features and the local features of the eyes, a multilayer perceptron is used to obtain the two-dimensional features of the gaze. , to indicate the angle of the estimated line of sight;

[0048] ;

[0049] In the formula, Represented as a multilayer perceptron, This is represented as a feature splicing operation.

[0050] Furthermore, L1Loss is used as the loss function for gaze estimation. The calculation is as follows:

[0051] ;

[0052] In the formula, This represents the true value of the estimated line of sight. This is represented as the predicted value from the line-of-sight estimation model.

[0053] Beneficial Effects: This invention presents a low-resolution gaze estimation method that integrates global and local features. It proposes a face feature extraction module based on super-resolution and convolutional Transformer. This module uses SwinIR to recover low-resolution face features, and then captures global face features from the feature map using a convolutional hybrid Transformer to recover low-resolution face features. This improves gaze estimation performance in low-resolution scenes. It also proposes an eye attention enhancement feature extraction module, primarily used to enhance gaze-related regions, thereby improving the effect of low-resolution gaze estimation. Furthermore, it proposes a gaze estimation module that integrates global and local features, fusing global features from the face and local features from the eyes to obtain more robust features for reliable gaze estimation. Finally, it effectively reduces information loss when extracting eye features from low-resolution images, enhancing the model's low-resolution gaze estimation performance. Attached Figure Description

[0054] Figure 1 A flowchart of a low-resolution gaze estimation method that integrates global and local features;

[0055] Figure 2 A diagram illustrating the overall network framework for a low-resolution gaze estimation method that integrates global and local features.

[0056] Figure 3 This is a structural diagram of the GRSA module. Detailed Implementation

[0057] The invention will now be further described with reference to the accompanying drawings.

[0058] As attached Figure 1 As shown, a low-resolution gaze estimation method that integrates global and local features includes the following steps:

[0059] Step 1: Acquire frame images of the person's face using the camera device on the equipment;

[0060] Step 2: Detect and locate the face region, left eye region, and right eye region in the frame image of the person, and crop the frame image of the face to obtain the face image, left eye image, and right eye image;

[0061] Step 3: Use a face feature extraction module based on super-resolution and convolutional transformer to extract global face features from the face image;

[0062] Step 4: Use the human eye attention enhancement feature extraction module to extract the local features of the left eye and the right eye from the left eye image and the right eye image;

[0063] Step 5: The gaze estimation module, which integrates global and local features, fuses the global features of the face image, the local features of the left eye, and the local features of the right eye to obtain the gaze direction of both eyes.

[0064] As attached Figure 1 As shown, in step two, a face detection model based on a convolutional neural network (CNN) is used to detect and locate the face region, left eye region, and right eye region of the frame image. Based on the detected and located face region, left eye region, and right eye region, the frame image is cropped to obtain the face image, left eye image, and right eye image. Specifically, the CNN-based face detection model uses the "dlib.cnn_face_detection_model_v1" model. This model includes the weights and structure of a CNN, which are trained on a large amount of image data. After loading the "dlib.cnn_face_detection_model_v1" model, faces can be detected in the image. Compared to traditional face detection models, the CNN-based face detection model has higher accuracy, can handle various complex situations, and has stronger robustness, real-time performance, and adaptability.

[0065] As attached Figure 2As shown, in step three, the face feature extraction module based on super-resolution and convolutional transformers uses the image super-resolution model SwinIR-2× to improve image quality and recover the details required for gaze estimation. The SwinIR image super-resolution model is an image super-resolution model designed based on the Swin Transformer. Compared with other models, it can effectively capture long-range dependencies in images and effectively recover face features. Furthermore, the SwinIR model combines features at different scales, enabling it to better reconstruct high-resolution images. The SwinIR image super-resolution model includes a shallow feature extraction module, a deep feature extraction module, and a high-quality image reconstruction module, corresponding to the shallow feature extraction stage, the deep feature extraction stage, and the high-quality image reconstruction stage, respectively.

[0066] The primary purpose of the shallow feature extraction module is to extract preliminary features from low-resolution images. This stage is a preparatory stage, employing the shallow feature extraction module of the SwinIR image super-resolution model to extract face images. The shallow features will The low-resolution image is converted into an enhanced dimensional feature space characterized by C channels. Specifically, it is represented as... The face image is transformed using a convolutional layer with a 3×3 kernel. The calculation process is as follows:

[0067] ;

[0068] In the formula, These are shallow features. This represents the transformation operation of the convolutional layer;

[0069] Adding a convolutional layer at the beginning of the Transformer helps it stabilize and converge faster during training. Following the shallow feature extraction module is the deep feature extraction module, which is the core of SwinIR. The deep feature extraction module of the SwinIR image super-resolution model contains K residual SwinTransformer blocks coupled to the CNN, i.e., Residual Swin Transformer Blocks (RSTBs), and a 3×3 convolutional layer. The Swin Transformer blocks are used to extract complex and abstract features, and they sequentially calculate transition features. The calculation process is as follows:

[0070] ;

[0071] In the formula, This is represented as the i-th RSTB block; the specific formula for the above calculation process is shown below, ultimately yielding the transition features. :

[0072] ,

[0073] ,

[0074] ,

[0075] ,

[0076] ;

[0077] Using convolutional layers to extract deep features The calculation process is as follows:

[0078] ;

[0079] In the formula, This represents the convolution operation for feature extraction by a convolutional layer. By placing a convolutional layer at the end of the feature extraction stage, the inductive bias of the convolution operation can be introduced into the Transformer-based network, thus laying a solid foundation for shallow and deep feature extraction.

[0080] The high-quality image reconstruction module of the SwinIR image super-resolution model combines shallow and deep features to reconstruct a high-quality image and output a high-resolution image; the calculation process is as follows:

[0081] ;

[0082] In the formula, This represents the reconstruction operation of a high-quality image reconstruction module; shallow features are mainly used to encapsulate low-frequency details, while deep features encapsulate high-frequency details. The complexity of reconstructing the latter requires a... arrive The long-range skip connections enable deep feature extraction to focus on recovering high-frequency details.

[0083] Then, the face feature module based on super-resolution and convolutional transformer uses a hybrid convolutional neural network and Transformer model to extract local features and capture global contextual relationships, thereby improving the global feature extraction capability. Specifically, a ResNet18 network is used to extract high-resolution image features from high-resolution images, and these high-resolution image features are flattened into vector features of two-dimensional tensors. The calculation process is as follows:

[0084] ;

[0085] In the formula, Feature extraction operations for the ResNet18 network This is a vector flattening operation;

[0086] In the face feature extraction module based on super-resolution and convolutional Transformer, the powerful global context modeling capability of VisionTransformer is utilized to extract global face features from the feature map; and class labels are generated for each input face image sample. With vector features The above positions are concatenated to generate position codes, which are then processed by a Transformer encoder. Perform splicing to obtain features The calculation process is as follows:

[0087] ;

[0088] In the formula, This is represented as a feature concatenation operation;

[0089] Features Input the vector into the Transformer to extract global face features, and select the first vector as the global face feature. The calculation process is as follows:

[0090] ;

[0091] In the formula, This is represented as the Transformer feature extraction operation.

[0092] As attached Figure 2 As shown, in step four, the human eye attention enhancement feature extraction module uses a residual network ResNet18 to extract local features from the left and right eye images. The ResNet18 network, through residual connections, effectively transmits local features in the image, enabling it to perform well in tasks requiring detail, such as gaze estimation. The ResNet18 network includes an initial layer, convolutional layers, batch normalization layers, ReLU activation function, max pooling layer, residual stages Res1, Res2, Res3, and Res4; each of the four residual stages has a similar structure. In the human eye attention enhancement feature extraction module, a GRSA module is embedded in the first two stages of the ResNet18 network, as shown in the attached diagram. Figure 3 As shown, the main function of the GRSA module is to enhance the line-of-sight related areas and complete the extraction of key features.

[0093] When the input image resolution is low, useful information is lost, and traditional ResNet feature extraction networks perform poorly when extracting features from low-resolution images. By focusing on features in gaze-related regions, the accuracy of gaze estimation can be improved; specifically, this involves focusing on the input features of the binocular images. The height and width of the input feature are adaptively averaged to reduce its spatial dimensionality. Then, the height and width of the input feature are averaged and max-pooled respectively to obtain the pooled feature. The calculation process is as follows:

[0094] ;

[0095] ;

[0096] In the formula, Represented as adaptive average pooling, subscript This is represented as an operation on height and width. This is represented as an average pooling operation. This is represented as a max pooling operation;

[0097] As attached Figure 3 As shown, the pooling features are concatenated and merged along both the height and width dimensions to obtain the height and width features:

[0098] ;

[0099] In the formula, This is represented as a vector concatenation operation;

[0100] As attached Figure 3 As shown, the obtained height and width features are then combined and processed to obtain the comprehensive spatial features. :

[0101] ;

[0102] In the formula, This is represented as a vector concatenation operation.

[0103] As attached Figure 3 As shown, for comprehensive spatial features Perform one-dimensional convolution and batch normalization operations, then normalize the convolutional features using the Sigmoid activation function to generate a spatial attention map. The calculation process is as follows:

[0104] ;

[0105] In the formula, Represented as an activation function, This is represented as a convolution operation. This is represented as a batch normalization operation;

[0106] Spatial attention map By performing a weighted multiplication with the original vector, a feature map of the spatial location is obtained, which in turn yields the local features of the left eye. and local features of the right eye :

[0107] ;

[0108] In the formula, This represents element-wise multiplication.

[0109] Compared to ordinary spatial attention, the human eye attention-enhanced feature extraction module can fuse channel information from feature maps, processing features in both height and width dimensions separately. This allows for a more comprehensive understanding of the spatial distribution of input features. Furthermore, it employs adaptive average pooling to capture contextual information from the feature map, enabling the model to accurately capture spatial structure information even in low-resolution images lacking rich detail. This enhances gaze-related regions, further improving the performance of convolutional neural networks on low-resolution gaze estimation tasks. By embedding gaze-related attention after the first two stages of the ResNet18 network, the module's ability to extract low-resolution eye features is enhanced.

[0110] As attached Figure 2 As shown, in step four, the human eye attention enhancement feature extraction module needs to extract local features of the left eye and the right eye from the left eye image and the right eye image. The human eye attention enhancement feature extraction module does not extract local features from the left eye image and the right eye image at the same time. It needs to extract local features from the left eye image and the right eye image separately. The human eye attention enhancement feature extraction module extracts local features only for one image.

[0111] As attached Figure 2 As shown, in step five, the gaze estimation module, which fuses global and local features, obtains the global facial features. Local features of the left eye and local features of the right eye By stitching together the features to maintain dimensional consistency among the three types, this method fully leverages the complementarity of facial, left eye, and right eye features, fusing richer features. Furthermore, after stitching together the global facial features and the local features of both eyes, a multilayer perceptron is used to obtain the two-dimensional features of the gaze. , to represent the angle of gaze estimation, which is the model prediction result, representing the direction of the person's gaze in the captured frame image;

[0112] ;

[0113] In the formula, Represented as a multilayer perceptron, This is represented as a feature splicing operation.

[0114] L1Loss, also known as Mean Absolute Error, is used as the loss function for gaze estimation. It represents the absolute value of the error between the model's predicted value and the true value. Using the L1Loss loss function allows the model to converge continuously during training, resulting in a more accurate model. The calculation is as follows:

[0115] ;

[0116] In the formula, This represents the true value of the estimated line of sight. This is represented as the predicted value from the line-of-sight estimation model.

[0117] Example

[0118] This invention uses the DIV2K and Flickr2K datasets to pre-train SwinIR; the parameters of the ResNet18 network are obtained through pre-training on ImageNet; because Transformer is highly dependent on pre-training, the entire network is pre-trained on the ETH-XGaze dataset, a large gaze estimation dataset. The authors of this dataset provided normalized data, so normalized data was directly used for training. A gaze estimation model is formed by combining three modules: a face feature extraction module based on super-resolution and convolutional transformers, a human eye attention enhancement feature extraction module, and a gaze estimation module that fuses global and local features. This model is trained and tested using modified MPIIFaceGaze datasets at different resolutions, following the preprocessing methods for the MPIIFaceGaze dataset in mainstream research.

[0119] To simulate various low-resolution scenarios, this experiment adjusted the preprocessed resolution of the dataset from 224×224 to 128×128. Furthermore, using scaling factors of 2× and 4×, the images at this resolution were bicubic downsampled again, resulting in three different low-resolution MPIIFaceGaze datasets with resolutions of 128×128, 64×64, and 32×32. Experiments on these three different resolution datasets demonstrate the versatility of this invention in low-resolution scenes and increase the persuasiveness of the experiments. The gaze estimation model was trained for 30 iterations with a batch size of 256, a learning rate of 0.0005, and a weight decay of 0.5. The decay step size was set to 60.

[0120] To verify the effectiveness of the network proposed in this invention, experiments were conducted on the MPIIFaceGaze dataset to compare it with other advanced methods such as the MNIST model, GazeNet model, Full Face model, Gaze360 model, RT-Gaze ​​model, and CDBN model. For a fair comparison, the experimental settings for each model were based on the settings in the corresponding paper, including the model's framework and hyperparameters, to reproduce their network performance. The performance comparison results of the proposed method with other advanced methods on the MPIIFaceGaze dataset are shown in Table 1 below. In Table 1, the top row represents the resolution of the dataset, and each column represents the angular error of different models at different resolutions.

[0121] Table 1. Comparative experimental results of the method of this invention with other advanced methods on the MPIIFaceGaze dataset.

[0122]

[0123] As can be seen from the table, in common low-resolution scenarios such as 128×128 and 64×64 resolutions, the method of this invention exhibits smaller angle errors and demonstrates superior performance. Even at a lower resolution of 32×32, the angle error of this invention remains very small, ranking second and placing it among the leading methods. Therefore, the overall performance of the method of this invention is leading at these three resolutions. The experimental results fully verify that the method of this invention has a significant effect on improving the line-of-sight estimation performance under common low-resolution conditions.

[0124] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. A low-resolution gaze estimation method that integrates global and local features, characterized in that: Includes the following steps: Step 1: Acquire frame images using a camera device; Step 2: Detect and locate the face region, left eye region, and right eye region in the frame image, and crop the face frame image to obtain the face image, left eye image, and right eye image; Step 3: Use a face feature extraction module based on super-resolution and convolutional transformer to extract global face features from the face image; Step 4: Use the human eye attention enhancement feature extraction module to extract local features of the left eye and the right eye from the left eye image and the right eye image; Step 5: The gaze estimation module, which integrates global and local features, fuses the global features of the face image, the local features of the left eye, and the local features of the right eye to obtain the gaze direction of both eyes; In step three, the shallow feature extraction module of the SwinIR image super-resolution model is used to extract the face image. The shallow features will The low-resolution image is converted into an enhanced dimensional feature space characterized by C channels. The calculation process is as follows: ； In the formula, These are shallow features. This represents the transformation operation of the convolutional layer; The deep feature extraction module of the image super-resolution model SwinIR includes K residual SwinTransformer blocks coupled to a CNN and a convolutional layer. The SwinTransformer blocks sequentially calculate the transition features. The calculation process is as follows: ； In the formula, Represented as the i-th RSTB block; Using convolutional layers to extract deep features The calculation process is as follows: ； In the formula, This represents the convolution operation used to extract features from a convolutional layer. In step four, the human eye attention enhancement feature extraction module uses a ResNet18 residual network to extract local features from the left and right eye images. The ResNet18 network includes an initial layer, convolutional layers, batch normalization layers, ReLU activation function, and max pooling layers. The network inputs features from the images of both eyes. The height and width of the input features are then subjected to adaptive average pooling, and then average pooling and max pooling are performed on the height and width of the input features respectively to obtain pooled features. The calculation process is as follows: ； ； In the formula, Represented as adaptive average pooling, subscript This is represented as an operation on height and width. This is represented as an average pooling operation. This is represented as a max pooling operation; The pooling features are concatenated and merged along the height and width dimensions respectively to obtain the height and width features: ； In the formula, This is represented as a vector concatenation operation; The obtained height and width features are then combined and processed to obtain a comprehensive spatial feature. : ； In the formula, This is represented as a vector concatenation operation; Comprehensive spatial characteristics Perform one-dimensional convolution and batch normalization operations, then normalize the convolutional features using the Sigmoid activation function to generate a spatial attention map. The calculation process is as follows: ； In the formula, Represented as an activation function, This is represented as a convolution operation. This is represented as a batch normalization operation; Spatial attention map By performing a weighted multiplication with the original vector, a feature map of the spatial location is obtained, which in turn yields the local features of the left eye. and local features of the right eye : ； In the formula, This represents element-wise multiplication.

2. The low-resolution gaze estimation method fusing global and local features according to claim 1, characterized in that: The high-quality image reconstruction module of the SwinIR image super-resolution model combines shallow and deep features to reconstruct a high-quality image and output a high-resolution image; the calculation process is as follows: ； In the formula, This represents the reconstruction operation of the high-quality image reconstruction module.

3. The low-resolution gaze estimation method fusing global and local features according to claim 2, characterized in that: The ResNet18 network is used to extract high-resolution image features from high-resolution images, and these high-resolution image features are flattened into vector features of two-dimensional tensors. The calculation process is as follows: ； In the formula, Feature extraction operations for the ResNet18 network This is a vector flattening operation; Category labels generated for each input face image sample With vector features The above positions are concatenated to generate position codes, which are then processed by a Transformer encoder. Perform splicing to obtain features : ； In the formula, This is represented as a feature concatenation operation; Features Input the vector into the Transformer to extract global face features, and select the first vector as the global face feature. The calculation process is as follows: ； In the formula, This is represented as the Transformer feature extraction operation.

4. The low-resolution gaze estimation method fusing global and local features according to claim 1, characterized in that: In step five, the gaze estimation module, which fuses global and local features, obtains the global facial features. Local features of the left eye and local features of the right eye The features are then stitched together; after stitching together the global facial features and the local features of the eyes, a multilayer perceptron is used to obtain the two-dimensional features of the gaze. , to indicate the angle of the estimated line of sight; ； In the formula, Represented as a multilayer perceptron, This is represented as a feature splicing operation.

5. The low-resolution gaze estimation method fusing global and local features according to claim 1, characterized in that: L1Loss is used as the loss function for line-of-sight estimation. The calculation is as follows: ； In the formula, This represents the true value of the estimated line of sight. This is represented as the predicted value from the line-of-sight estimation model.

Images