Multi-person head posture estimation method, device and medium based on geodesic loss
By using geodesic loss method in multi-person head pose estimation, matching and updating the parameters of the target detection network, the problems of complex detection, slow speed and high resource consumption in the prior art are solved. It realizes fast and accurate multi-person head posture estimation and efficient resource use.
Patent Information
- Application Number
- CN202411352336.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-26
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-09-26
AI Technical Summary
In the prior art, the multi-person head posture estimation method has complex detection, slow speed, high resource consumption, and insufficient detection accuracy, making it difficult to accurately detect the multi-person head posture.
Using the multi-person head pose estimation method based on geodesic loss, the detection information of multiple head targets is detected through the target detection network, the predicted rotation matrix and the real rotation matrix are matched, the geodesic loss is constructed, and the backpropagation loss is calculated. , guide network parameter updates and train multi-person head pose estimation model.
Fast and accurate multi-person head posture estimation is achieved, reducing detection complexity, improving detection accuracy, and reducing resource consumption.
Smart Images

Figure CN118865453B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of head posture detection, and in particular to a method, device and medium for estimating the head posture of multiple people based on geodesic loss. Background Art
[0002] Head posture estimation is based on computer vision technology to accurately identify and calculate the three-dimensional direction and position of the human head from images or videos. Its core is to analyze visual data through algorithms to accurately capture and understand the human head posture. Head posture estimation has a wide range of applications in human-computer interaction, AR and VR fields, driver fatigue monitoring systems, etc.
[0003] Regarding head posture estimation, the existing technologies usually focus on single-person head posture estimation, and there are relatively few solutions for multiple-person head posture estimation. Moreover, in the existing technologies, for multiple-person head posture estimation, face detection is usually performed first, and then the faces are sliced into multiple faces before single-person head posture estimation is performed separately. Since such methods need to perform face detection, segmentation, and single-person head posture estimation, there are problems such as complex detection implementation, slow detection speed, and high resource consumption. Some practitioners have proposed to add a head posture estimation prediction detection head on the basis of the target detection network. Although the detection speed can be improved to a certain extent, the direct addition of the prediction detection head will lead to the problem of insufficient detection accuracy, making it difficult to accurately detect the head postures of multiple people at the same time. Summary of the invention
[0004] The technical problem to be solved by the present invention is as follows: In view of the above-mentioned problems in the prior art, a method, device and medium for estimating multi-person head posture based on geodesic loss are provided, which have simple implementation method, low cost, fast detection speed, high detection accuracy and low resource consumption.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0006] A method for estimating multi-person head posture based on geodesic loss, comprising the following steps:
[0007] Inputting the training image into the target detection network, the target detection network detects detection information of multiple head targets, wherein the detection information includes a predicted rotation matrix;
[0008] Each predicted rotation matrix is matched with the real rotation matrix corresponding to the real head with the label. After matching, the positive samples are divided from the detection results obtained by the target detection network to obtain the predicted multi-person head posture positive sample matrix, and a label is matched to each positive sample to obtain the real label matrix corresponding to the positive sample;
[0009] The predicted multi-person head posture positive sample moments and the true label matrix corresponding to the positive samples are used according to the formula The geodesic loss is constructed to characterize the distance between the predicted rotation matrix and the true rotation matrix, where represents a function whose value is limited to (-1, 1). ) represents the positive sample matrix for predicting multiple head poses , the true label matrix corresponding to the positive sample function, and calculate the back propagation loss according to the geodesic loss;
[0010] In the process of training the target detection network using training pictures, a loss function including the back-propagation loss is used to guide the updating of network parameters until the target detection network converges on the multi-person head pose estimation task, and a multi-person head pose estimation model is obtained through training to realize multi-person head pose estimation.
[0011] Furthermore, the expression of the predicted rotation matrix is:
[0012]
[0013]
[0014]
[0015]
[0016]
[0017]
[0018] in, represents the predicted rotation matrix, are the 6 values of the 6D rotation matrix, with superscript T Indicates transpose.
[0019] Furthermore, the expression of the true rotation matrix is:
[0020]
[0021] in, represents the true rotation matrix, , , They are the real multi-person head posture labels corresponding to , , Axis coordinate values.
[0022] Furthermore, the expression of the geodesic loss is:
[0023]
[0024] in, represents the geodesic loss, is a hyperparameter and <3, in the clamp function, if it is greater than 1, it takes 1, and if it is less than -1, it takes -1.
[0025] Furthermore, the expression of the back propagation loss is:
[0026]
[0027] in, represents the back-propagation loss, represents the geodesic loss, represents the weight matrix of the positive sample, represents the weight of the i-th positive sample, Represents the number of positive samples.
[0028] Furthermore, the loss function including the back propagation loss is:
[0029]
[0030]
[0031]
[0032] in, represents the overall loss, 、 、 are the weights of back propagation loss, classification loss, and detection loss, respectively. represents the back-propagation loss, represents the classification loss, represents the detection loss, represents the classification prediction value, represents the true value of the classification, represents the BCE (Binary Cross Entropy) loss function, Represents the detection box prediction value, Represents the true value of the detection box, represents the DFL (Deep Focal Loss) loss function, Represents the CIOU (Complete IoU, intersection over union loss) loss function.
[0033] Furthermore, the head pose estimation model is used to predict the head pose according to the following formula:
[0034]
[0035]
[0036]
[0037]
[0038] in, Respectively represent the predicted head posture , , Axis coordinate values, represents the predicted rotation matrix, express The elements in .
[0039] Furthermore, when the training images are input into the target detection network, adaptive frame processing is also performed on the training images to uniformly adjust each training image to an image of a specified size.
[0040] An electronic device comprises a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.
[0041] A computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.
[0042] Compared with the prior art, the advantages of the present invention are as follows: in the process of using the target detection network to train the model, the present invention obtains a predicted multi-person head posture positive sample matrix by matching the predicted rotation matrix detected by the network with the real rotation matrix, and combines the multi-person head posture positive sample matrix and The function constructs the geodesic loss and then calculates the back propagation loss. The loss function including the back propagation loss is used to guide the update of network parameters during network training. It can make full use of the positive samples obtained by matching the predicted rotation matrix with the true rotation matrix, and make some rough prediction results not participate in the back propagation, so that the model prediction is more accurate and can quickly and accurately realize multi-person head posture estimation, which can not only reduce the detection complexity and improve the detection speed, but also improve the detection accuracy and reduce the required resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 4 is a schematic diagram of the implementation flow of the multi-person head posture estimation method based on geodesic loss in this embodiment.
[0044] Figure 2 Schematic diagram of the principle of using the target detection network for detection in this embodiment. DETAILED DESCRIPTION
[0045] like Figure 1 As shown, the steps of the multi-person head posture estimation method based on geodesic loss in this embodiment include:
[0046] Step S01. Input the training image into the target detection network, and the target detection network detects the detection information of multiple head targets, the detection information includes the predicted rotation matrix, the detection box, the classification confidence, etc.;
[0047] Step S02. Match each predicted rotation matrix with the real rotation matrix corresponding to the real head with the label. After matching, divide the positive samples from the detection results obtained by the target detection network to obtain the predicted multi-person head posture positive sample matrix, and match a label for each positive sample to obtain the real label matrix corresponding to the positive sample;
[0048] Step S03: Use the predicted multi-person head posture positive sample matrix and the true label matrix corresponding to the positive samples according to the formula The geodesic loss is constructed to characterize the distance between the predicted rotation matrix and the true rotation matrix, where Indicates that the value is limited to (-1, 1). ) represents the positive sample matrix for predicting multiple head poses , the true label matrix corresponding to the positive sample function, and calculate the back propagation loss according to the geodesic loss;
[0049] Step S04. In the process of training the target detection network using the training images, a loss function including back propagation loss is used to guide the update of network parameters until the target detection network converges on the multi-person head pose estimation task, and a multi-person head pose estimation model is obtained through training to realize multi-person head pose estimation.
[0050] This embodiment is based on a target detection network, and uses the target detection network to detect detection frames, classification confidences, and predicted rotation matrices of multiple head targets. In the process of training the target detection network, the predicted rotation matrix is matched with the true rotation matrix to obtain a predicted multi-person head posture positive sample matrix, and the multi-person head posture positive sample matrix is used to construct a geodesic loss to characterize the distance between the predicted rotation matrix and the true rotation matrix, and then the back propagation loss is calculated. The loss function including the back propagation loss is used to guide the update of network parameters during network training. Since the geodesic loss makes full use of the positive samples obtained by matching the predicted rotation matrix with the true rotation matrix, and through The function prevents some rough prediction results from participating in the back propagation, making the model prediction more accurate, so that a model that can accurately estimate the head postures of multiple people can be trained. This not only reduces the detection complexity and improves the detection speed, but also improves the detection accuracy and reduces the required resource consumption.
[0051] In this embodiment, after the training images are input into the target detection network and before training, the training images are also subjected to adaptive frame processing to uniformly adjust the training images to images of a specified size, so as to enhance the feature extraction capability of the target detection network. For example, the input images can be uniformly adjusted to an image size of 1280×1280 pixels to adaptively preprocess the training images into images of a size of 1280×1280, so as to facilitate feature extraction by the target detection network.
[0052] In this embodiment, after the input image is detected using the YOLOv (You Only Look Once) 9 target detection network, the bounding box of each detected head and the classification confidence are output respectively, and a 6D (dimensional) detection branch composed of multiple convolution kernels is added to the YOLOv9 network to detect the 6D rotation matrix, such as Figure 2 As shown in the figure, the 6D rotation matrix contains detailed information about the head rotation, which enables more accurate capture of each person's head posture changes. By converting the predicted 6D rotation matrix into a rotation matrix and then using the geodesic loss function to measure the difference between the true rotation matrix and the predicted rotation matrix, the accuracy of the model prediction can be improved.
[0053] Specifically, the YOLOv9 target detection network predicts several detection boxes, and the 6D detection branch at each detection box predicts 6 values , These are the six values predicted by the 6D detection branch, namely the 6D rotation matrix representation. By transforming these six values, an orthogonal matrix can be obtained, which is the predicted rotation matrix. After the training image is input into the target detection network, in addition to the output classification confidence and detection frame information, a 6D rotation matrix representation is also output, that is, each detection frame will output classification confidence, detection frame and 6D rotation matrix representation. Since only one class, the human head, is predicted, 1+4+6=11 values can be predicted:
[0054] (1)
[0055] in, Represents the classification confidence. Since only one class, human head, is predicted, there is only one category confidence. Indicates the center point of the detection box. Respectively represent the width and height of the detection box, The 6 values representing the 6D rotation matrix.
[0056] As mentioned above, after the target detection network, the output is a A vector of is the total number of detection boxes.
[0057] The 6 values of the 6D branch prediction form two three-dimensional vectors:
[0058] (2)
[0059] (3)
[0060] Then the two three-dimensional vectors , Do the following conversion:
[0061] (4)
[0062] (5)
[0063] (6)
[0064] in, represents the modulus of the orientation quantity, Represents the cross product of vectors.
[0065] After the above transformation, we can get three mutually orthogonal unit vectors , , , and then based on these three vectors, the predicted rotation matrix can be constructed:
[0066] (7)
[0067] in, Denotes the predicted rotation matrix, the superscript T represents the transpose. The predicted rotation matrix is an orthogonal matrix, and is a .
[0068] For realistic multi-person head pose labels , where x, y, and z are the three angles (Euler angles) of head posture estimation. Both The matrix of is the actual number of heads, then the actual rotation matrix can be expressed as:
[0069] (8)
[0070] (9)
[0071] (10)
[0072] Further according to the internal rotation method, the ZYX rotation order obtains the true rotation matrix:
[0073] (11)
[0074] Then the specific expression of the true rotation matrix can be obtained as:
[0075] (12)
[0076] in, represents the true rotation matrix, , , They are the real multi-person head posture labels corresponding to , , Axis coordinate value. The real rotation matrix is also an orthogonal matrix and can be written as:
[0077] (13)
[0078] in, , , are three mutually orthogonal three-dimensional unit vectors, then the real multi-person head posture label is converted into a multi-person head posture rotation matrix label. is a The matrix of .
[0079] In this embodiment, the task alignment assigner (TaskAlignedAssigner) allocation strategy is specifically used to match the prediction results and the real targets. According to the matching results, each detection frame obtained by the target detection network is divided into positive and negative samples, and the positive samples are matched with corresponding labels. TaskAlignedAssigner is a positive and negative sample matching strategy used in target detection tasks, that is, a method for finding positive samples. It belongs to a dynamic allocation strategy. The core idea of this strategy is that as the model is optimized, the prediction frame will gradually get closer to the real target (Ground Truth, GT), so that the positive and negative samples can be reallocated according to the prediction results of the model. Based on the TaskAlignedAssigner matching strategy, this embodiment calculates the matching degree of all prediction frames with the real frame, and takes the prediction frame whose center point falls inside the real frame as the positive sample for preliminary screening. For each real frame, based on the calculated matching degree, the prediction frame with the highest matching degree is selected as the positive sample, and then the label of the positive sample is set in combination with the label of the real frame and the classification score of the prediction frame. The label of the positive sample is the corresponding category label, and the label of the negative sample is the background label.
[0080] This embodiment uses the TaskAlignedAssigner strategy to match the predicted rotation matrix with the real rotation matrix, and obtains the predicted positive samples of multiple head postures. :
[0081] (14)
[0082] Above yes part of which is a The matrix of , that is, in (Np, 3, 3) matches as positive samples, among which is the number of positive samples, whose size is less than .
[0083] When matching positive and negative samples, each positive sample will be matched with a label, that is, when the size is ( ,3,3) Select one as the label matching the positive sample, and the true label matrix corresponding to the positive sample can be expressed as :
[0084] (15)
[0085] Above of size (matrix size) and Consistent, the same The matrix of of In addition, using the TaskAlignedAssigner matching strategy during the matching process, you can also get a size of The weight matrix , which represents the weight of each positive sample.
[0086] In order to measure the distance between the predicted rotation matrix and the actual rotation matrix, this embodiment uses the geodesic loss shown below:
[0087] (16)
[0088] in, represents the trace of the transposed product of two rotation matrices, is a hyperparameter and <3, clamp means Intermediate value function, used to ensure that the intermediate value is If it is greater than 1, it takes 1, and if it is less than -1, it takes -1.
[0089] Since the multi-person head pose estimation is based on the target detection network, the predicted results and the real labels are in a many-to-one relationship, so some of the predicted results will be relatively rough, affecting the training effect. When , due to the restriction of the clamp function, some rough prediction results will not participate in the back propagation, and as As increases, the model will pay more attention to predicting more accurate prediction results, making the model prediction more accurate.
[0090] This embodiment adopts the above-mentioned geodesic loss. When it is too large, the model may be difficult to converge due to the small number of prediction results involved in back propagation. value, we can also get a more accurate geodesic loss. For example, when the geodesic loss When taking 1, we get:
[0091] (17)
[0092] Using the geodesic loss in the above formula can not only prevent the rough prediction results from participating in the back propagation, making the model prediction more accurate, but also avoid the problem of difficult model convergence.
[0093] It is understood that in addition to the above forms, ) Other forms can also be used according to actual needs. The key is to use the positive sample matrix To represent the distance between the predicted rotation matrix and the true rotation matrix.
[0094] Further substituting the value of the rotation matrix into the above geodesic loss, the final geodesic loss is:
[0095] (18)
[0096] in represents the inner product of two vectors, then is a The matrix of .
[0097] The above geodesic loss is further used to calculate the back propagation loss, and the back propagation loss is obtained as:
[0098] (19)
[0099] in, represents the back-propagation loss, represents the geodesic loss, represents the weight matrix of the positive sample, represents the weight of the i-th positive sample, Represents the number of positive samples.
[0100] For the classification branch and the detection box branch, this embodiment uses the following losses:
[0101] (20)
[0102] (twenty one)
[0103] in, represents the classification loss, represents the detection loss, represents the classification prediction value, represents the true value of the classification, represents the BCE loss function, Represents the detection box prediction value, Represents the true value of the detection box, represents the DFL loss function, represents the CIOU loss function.
[0104] The above BCE loss function uses binary cross entropy as the classification loss, which measures the distance between the probability output by the sigmoid function and the true label. The DFL loss function optimizes the probabilities of the two left and right positions closest to the label y in the form of cross entropy, so that the network can focus on the target position and the distribution of the neighboring area more quickly. IoU (intersection-union ratio) is an evaluation index for the accuracy of the bounding box positioning of the object detection model. It describes the overlap between the boxes by calculating the ratio of the intersection of the predicted area and the real area to their union. The DIoU loss is based on IoU by introducing the normalized distance between the center point of the predicted box and the target box as a penalty term to accelerate the convergence speed. The CIOU loss function adds a constraint on the aspect ratio on the basis of the DIoU (Distance-IoU Loss) loss, so that the anchor box is closer to the loss of the label value. It is understandable that the above loss functions can also adopt other types according to actual needs.
[0105] The loss function is a function that evaluates the difference between the model prediction output and the true label. In this embodiment, the overall loss is composed of back propagation loss, detection loss and classification loss, that is, the loss function is:
[0106] (twenty two)
[0107] in, represents the overall loss, 、 、 are the weights of back propagation loss, classification loss, and detection loss, respectively. represents the back-propagation loss, represents the classification loss, Represents the detection loss.
[0108] This embodiment optimizes the classification task by using the binary cross entropy loss BCEloss, and optimizes the detection box regression by combining the deep focus loss DFL and the intersection-over-union loss CIoU Loss. The comprehensive loss functions can quickly and accurately guide the update of network parameters, so that the model can quickly converge on the multi-person head pose estimation task, and finally obtain a more accurate multi-person head pose estimation model.
[0109] The above-mentioned multi-person head posture estimation model can be used to estimate the head postures of multiple people in the test image. When using the multi-person head posture estimation module for prediction, the test image is first preprocessed into a size of 1280×1280 through an adaptive frame and then input into the network. Non-Maximum Suppression (NMS) is performed on the network inference output to obtain n prediction results. Each prediction result includes classification, prediction frame, confidence, and 6D rotation matrix representation, a total of 1+4+1+6=12 values. The 6D rotation matrix representation can be used to obtain the prediction rotation matrix according to the above formula (7):
[0110]
[0111] Further prediction, the rotation matrix can also be written as:
[0112] (twenty three)
[0113] The final predicted x and z values range from -90° to 90°, and the y value range is from -180° to 180°. Specifically, the head postures predicted by the multi-person head posture estimation model are: (twenty four)
[0114] (25)
[0115] (26)
[0116] (27)
[0117] in, Respectively represent the predicted head posture , , Axis coordinate values, represents the predicted rotation matrix, express The elements in .
[0118] Using the above formula, we can get the predicted head posture ,in Both The matrix of .
[0119] This embodiment further provides an electronic device, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.
[0120] It is understandable that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario and completed by multiple devices in cooperation with each other. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and multiple devices interact to complete the above method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., for executing related programs to implement the above method of this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device. The memory can store an operating system and other applications. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.
[0121] This embodiment further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above method is implemented.
[0122] Those skilled in the art should understand that the above-mentioned embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process Figure 1 A process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including an instruction device, which implements the functions specified in the process. Figure 1 A process or multiple processes and / or boxes Figure 1These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0123] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. A multi-person head pose estimation method based on geodesic loss, It is characterized in that the steps include: Inputting the training image into the target detection network, the target detection network detects detection information of multiple head targets, wherein the detection information includes a predicted rotation matrix; Each predicted rotation matrix is matched with the real rotation matrix corresponding to the real head with the label. After matching, the positive samples are divided from the detection results obtained by the target detection network to obtain the predicted multi-person head posture positive sample matrix, and a label is matched to each positive sample to obtain the real label matrix corresponding to the positive sample; The predicted multi-person head posture positive sample matrix and the real label matrix corresponding to the positive samples are used according to the formula The geodesic loss is constructed to characterize the distance between the predicted rotation matrix and the true rotation matrix, where represents a function whose value is limited to (-1, 1). ) represents the positive sample matrix for predicting multiple head poses , the true label matrix corresponding to the positive sample function, and calculate the back propagation loss according to the geodesic loss; In the process of training the target detection network using the training images, a loss function including the back-propagation loss is used to guide the updating of network parameters until the target detection network reaches convergence on the multi-person head pose estimation task, and a multi-person head pose estimation model is obtained through training to realize multi-person head pose estimation; The expression of the predicted rotation matrix is: in, represents the predicted rotation matrix, are the 6 values of the 6D rotation matrix, with superscript T represents transpose; The expression of the true rotation matrix is: in, represents the true rotation matrix, , , They are the real multi-person head posture labels corresponding to , , Axis coordinate values; The expression of the geodesic loss is: in, represents the geodesic loss, is a hyperparameter and <3, clamp If the function is greater than 1, it takes 1, and if it is less than -1, it takes -1.
2. The method for multi-person head pose estimation based on geodesic loss according to claim 1, It is characterized in that The expression of the back propagation loss is: in, represents the back-propagation loss, represents the geodesic loss, represents the weight matrix of the positive sample, represents the weight of the i-th positive sample, Represents the number of positive samples.
3. The method for estimating multiple head poses based on geodesic loss according to any one of claims 1 to 2, It is characterized in that The loss function including the back propagation loss is: in, represents the overall loss, represents the back-propagation loss, represents the classification loss, represents the detection loss, 、 、 are the weights of back propagation loss, classification loss, and detection loss, respectively. represents the classification prediction value, represents the true value of the classification, represents the BCE loss function, Represents the detection box prediction value, Represents the true value of the detection box, represents the DFL loss function, represents the CIOU loss function.
4. The method for estimating multiple head poses based on geodesic loss according to any one of claims 1 to 2, It is characterized in that Use the multi-person head pose estimation model to predict the head pose according to the following formula: in, Respectively represent the predicted head posture , , Axis coordinate values, represents the predicted rotation matrix, express The elements in .
5. The method for estimating multiple head poses based on geodesic loss according to any one of claims 1 to 2, It is characterized in that Inputting the training images into the target detection network also includes performing adaptive frame processing on the training images to uniformly adjust each training image to an image of a specified size.
6. An electronic device comprising a processor and a memory, wherein the memory is used to store a computer program. It is characterized in that The processor is configured to execute the computer program to perform the method according to any one of claims 1 to 5.
7. A computer-readable storage medium storing a computer program, It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.
