Scalable adaptive full-view head pose estimation and normalization selection method

Abstract

The present application belongs to the technical field of computer vision and image recognition, and specifically relates to an extensible adaptive full-view head posture estimation and normalization selection method. The method comprises the following steps: firstly, resource perception and candidate algorithm screening are performed, and a set of adaptive algorithms is dynamically selected in combination with task requirements and device computing power; subsequently, the candidate algorithms are executed in parallel, and evaluation results in multiple dimensions of geometric stability, visual quality and downstream task adaptability are obtained; the optimal algorithm is determined based on the evaluation score, scene features are extracted, a scene library is established through clustering, and cluster-level optimal algorithm records are maintained. The present application can output stable normalized results under full-view and complex shooting conditions, dynamically balance the requirements of accuracy, speed and computing power, eliminate the influence of scene mutations through boundary smoothing decision, form a self-learning closed-loop optimization system, and is suitable for various scenes requiring face standardization processing such as identity verification and criminal investigation.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and image recognition technology, specifically involving a scalable adaptive full-view head pose estimation and normalization selection method. Background Technology

[0002] Head pose estimation and alignment is a crucial foundational step in computer vision and head recognition. Its core task is to estimate the 3D spatial pose parameters of a head from an input image and align the head to a standardized pose using geometric or learning methods. This process provides stable and consistent input features for subsequent applications such as face recognition, expression analysis, identity verification, 3D reconstruction, immersive interaction, and virtual avatar generation, making it a key preliminary step in the entire human perception and understanding system. With the development of artificial intelligence and deep learning, head pose estimation and alignment techniques have evolved into various implementation approaches.

[0003] However, in practical applications, image acquisition environments are highly diverse and uncontrollable. Different shooting angles, lighting conditions, occlusion, and imaging devices can significantly affect the quality and distribution of head images. Furthermore, the hardware for running pose estimation and alignment algorithms varies considerably, with different computational power, storage, and real-time constraints. Existing research often uses single algorithm models, which tend to perform best only on specific datasets or under controlled conditions, making it difficult to balance accuracy, speed, and computational constraints. Therefore, at the application level, there is still a lack of a scalable, adaptive, full-view head pose estimation and normalization selection system that can dynamically select the optimal algorithm based on the operating environment and input characteristics, and maintain consistent and stable output across multiple scenarios.

[0004] In the field of head pose estimation and normalized alignment, the mainstream methods can be summarized into three categories: (1) Geometric transformation-based methods. These methods establish affine, perspective, or projection geometric models, perform coordinate mapping and spatial correction on the detected head region, and transform head images from different perspectives to a unified standard pose. Their implementation usually relies on the key point detection results to calculate the transformation matrix.

[0005] (2) Methods based on 3D face models. This type of method estimates the 3D shape parameters, pose parameters and lighting conditions from the input image, reconstructs the 3D structure of the human head through the 3D shape and texture model, and then renders it to the standard pose to achieve alignment and normalization.

[0006] (3) Generative or regression methods based on deep learning. These methods use CNN or GAN to directly learn the mapping relationship from non-frontal human head images to standard frontal views. Common practices include regressing three-dimensional pose angle parameters, predicting key point coordinates, or directly outputting frontalized images through generative networks.

[0007] In recent years, some studies have attempted hybrid approaches combining 3D modeling and deep learning, introducing 3D geometric constraints or explicit shape priors into deep networks to balance geometric consistency and generation quality. However, these methods still belong to a single-model framework and lack the ability to dynamically select and optimize under different computing power, environmental, and scene conditions. In the field of head pose estimation and normalized alignment at the multi-model level, some studies have explored dynamic selection mechanisms based on model sets. The Pose-Aware Models (PAMs) framework proposed by Masi et al. (CVPR 2016) divides the pose space into several discrete intervals, trains independent sub-models for different pose ranges, and selects or weights the corresponding model based on the pose information of the input image during the inference stage, thereby improving robustness and accuracy under large pose changes. Patent EP2672424A1 proposes a dynamic switching mechanism based on multi-morphological models, which improves the adaptability of registration by automatically selecting matching pre-trained models under different facial conditions. Related works all pre-train multiple sub-models that perform well under different poses or lighting conditions, and select the best-matching model for estimation based on the features of the input image during the inference stage, thus improving the system's environmental adaptability to a certain extent.

[0008] In head pose estimation and normalization alignment techniques, although existing methods can achieve pose correction and standardized output to varying degrees, there are still obvious limitations in various technical approaches.

[0009] Methods based on geometric transformations are simple in structure and computationally efficient, but they are prone to geometric distortion and keypoint misalignment under large-angle rotations, complex lighting, or partial occlusion. Methods based on 3D head models can maintain structural consistency at the geometric level and achieve high accuracy within a small to medium angle range, but they are highly dependent on model fitting quality and lighting estimation, resulting in high computational cost and insufficient real-time performance. Generative or regression methods based on deep learning can adapt to complex pose and lighting changes, but their fixed model structure, strong dependence on training data, and weak generalization ability are significant challenges.

[0010] Furthermore, a single algorithm struggles to achieve a balance between accuracy, speed, and computational constraints. Even state-of-the-art (SOTA) models often have advantages limited to specific experimental environments or controlled datasets, making it difficult to maintain consistent performance across various real-world scenarios. Current multi-model-level improvements alleviate the adaptability limitations of single algorithms to some extent, but their model switching typically relies on static rules, failing to comprehensively consider the dynamic trade-offs between algorithm performance, scene characteristics, and hardware resources. Simultaneously, they lack real-time feedback-based policy update mechanisms, hindering continuous optimization during long-term operation or the emergence of new scenarios, and making it difficult to stably output a unified, standardized representation in ever-changing real-world environments.

[0011] To address the shortcomings of existing methods, this application proposes to construct a scalable adaptive full-view head pose estimation and normalization selection system, introducing resource awareness, multi-candidate algorithm evaluation, and scene memory mechanisms, focusing on solving the following problems: (1) Multi-scenario adaptability problem: How to maintain good estimation and alignment results in complex real-world scenarios.

[0012] (2) Computing power and real-time constraints: How to automatically adjust the algorithm strategy according to the computing power, latency or task priority of the running equipment, so as to balance accuracy and speed.

[0013] (3) Algorithm selection and performance evaluation mechanism issues: How to make dynamic decisions based on actual performance and scenario characteristics.

[0014] (4) Continuous optimization and self-learning problem: How to automatically update the optimal strategy based on historical performance or environmental changes, and realize scene memory and strategy iteration mechanism. Summary of the Invention

[0015] To address the aforementioned technical problems, this invention provides a scalable adaptive full-view head pose estimation and normalization selection method, comprising the following steps: Resource awareness and candidate algorithm screening: After receiving the input image, the system extracts the basic meta-information of the task and generates a task description vector. It detects the computing power status of the device in real time to form a resource status table. It calculates the comprehensive resource index based on the task description vector and the resource status table. If there is a stable and optimal algorithm, it runs directly. Otherwise, it selects a set of candidate algorithms that meet the resource constraints and scenario adaptation requirements from the built-in algorithm registration library. The algorithm performs parallel estimation and alignment result generation, schedules and executes each algorithm in the candidate algorithm set, collects the head pose parameters, key point coordinate set and normalized image results output by each algorithm, evaluates the geometric consistency and stability of the key point set, scores the visual quality and geometric matching degree of the normalized image, and evaluates the downstream task adaptability if it is in task chain mode. Results quality assessment and optimal selection: After normalizing the above evaluation indicators, calculate the comprehensive score, select the algorithm with the highest comprehensive score as the optimal algorithm in the current scenario, and record its score and running information. Scene feature extraction and memory update: Extract the scene feature vector of the current input, perform cluster analysis on the continuous variable subspace to build a scene library, maintain an algorithm performance statistics table for each scene cluster, update the relevant information of the optimal algorithm, and adopt a boundary smoothing decision mechanism for continuous scene features; Self-learning and policy solidification: When the input sample belongs to an existing scene cluster, the preferred algorithm solidified in that cluster is called first. The performance of the solidified algorithm is continuously monitored. If a performance degradation or a new algorithm is registered, the solidification state is deactivated and the candidate evaluation and policy update process is re-executed.

[0016] Preferably, the task description vector includes parameters related to input resolution, frame rate, task priority, and latency constraints.

[0017] Preferably, the resource status table records CPU / GPU utilization, available video memory space, power consumption limits, and system latency parameters, and is periodically refreshed to reflect the real-time operating environment.

[0018] Preferably, the built-in algorithm registration library covers three representative methods: keypoint alignment algorithm based on geometric transformation, shape regression algorithm based on 3D model fitting, and frontalization algorithm based on deep generative networks.

[0019] Preferably, the candidate algorithm screening process is as follows: first, models that do not meet the requirements are eliminated based on resource consumption constraints; then, scene adaptation screening is performed based on task operation mode and image scene labels; finally, the top-3 models are selected according to the comprehensive score to form a candidate set. The comprehensive score takes into account both accuracy and latency performance, and accuracy takes precedence when scores are the same.

[0020] Preferably, the geometric consistency and stability assessment is achieved by measuring the degree of positional dispersion between key points and the deviation of the rotation matrix from the reference rotation matrix. The visual quality and geometric matching score covers image sharpness, alignment symmetry, and facial structure integrity indicators.

[0021] Preferably, the scene feature vector includes pose angle features, illumination-related parameters, occlusion ratio, and resource status level.

[0022] Preferably, the clustering analysis adopts an incremental online update algorithm, which calculates Mahalanobis distance to determine the similarity between the input sample and each cluster, determines the sample's affiliation based on a preset threshold, and triggers overall re-clustering when the number of newly added clusters exceeds the preset threshold.

[0023] Preferably, during the self-learning process, when an algorithm has more than a preset threshold in the number of samples in the corresponding scene cluster and has the highest average score, it is marked as the preferred algorithm for that cluster. The criterion for judging the performance degradation is that the overall score of the algorithm decreases by more than a preset threshold.

[0024] Preferably, the boundary smoothing decision mechanism is as follows: when the input feature is located near the boundary of the scene cluster interval, the final algorithm selection decision is generated by referring to the historical best algorithm and comprehensive score of the adjacent clusters and weighting the feature similarity or confidence.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Unified and stable output under all-view and complex environments: Through parallel evaluation of multiple candidate algorithms and scene clustering memory mechanism, the optimal algorithm can be dynamically selected according to the pose, lighting and occlusion features of the input image, and stable pose estimation and normalization results can be output under 360° all-view and various real shooting conditions.

[0026] 2. Dynamic adaptive balance between accuracy, speed and computing power: The introduction of a resource awareness module enables real-time detection of device computing power, latency constraints and task priorities, allowing the system to automatically balance efficiency and real-time performance.

[0027] 3. Continuous Scene Generalization and Boundary Smoothing Decision Mechanism: For the continuous changes of factors such as pose angle, illumination, and occlusion ratio in the real environment, the boundary similarity weighted decision method eliminates the output instability caused by scene abrupt changes.

[0028] 4. Self-learning driven closed-loop optimization system: During long-term system operation, a score fluctuation monitoring, performance degradation detection, and strategy solidification mechanism are introduced to form a self-learning closed loop from algorithm selection, evaluation feedback to strategy update, ensuring that the system maintains high accuracy, consistency and robustness in multi-perspective, multi-computing power and dynamic environments. Attached Figure Description

[0029] Figure 1 This is a diagram of the overall architecture of the present invention; Figure 2 This is a schematic diagram of the self-learning process of the present invention. Detailed Implementation

[0030] Please refer to Figures 1 to 2 The scalable adaptive full-view head pose estimation and normalization selection method provided by this invention is implemented through the following technical solution: Step 1: Resource Awareness and Candidate Algorithm Selection 1.1 Input Reception and Task Initialization: After receiving the input image, the system extracts basic meta-information related to the task, including input resolution, frame rate, task priority, and whether it is a real-time processing task, and generates a corresponding task description vector. This task description vector is the system's internal representation used for task evaluation and contains a set of parameters:

[0031] Where r represents the input resolution, f represents the frame rate, p represents the task priority, and d represents the latency constraint.

[0032] 1.2 Equipment Resource Status Monitoring: The system monitors the computing power status of the current devices in real time by calling the resource monitoring module, and generates a resource status table. This table records information such as CPU / GPU utilization, available video memory space, power consumption limits, and system latency parameters, and is refreshed periodically to reflect the real-time operating environment.

[0033] 1.3 Resource Awareness Analysis and Execution Mode Selection: The system calculates the comprehensive resource index C=f(T,R) based on the task description vector T and the resource status table R. If there is a stable optimal algorithm under the current scenario and hardware resource conditions, the optimal algorithm is run directly, and the process jumps to step five.

[0034] 1.4 Candidate Algorithm Screening and Combination Generation: The system's built-in algorithm registration library covers three representative methods: Keypoint alignment algorithms based on geometric transformations (DlibAlignment, SDM, etc.); Shape regression algorithms based on 3D model fitting (3DDFA, PRNet, etc.); Positive transformation algorithms based on deep generative networks (TP-GAN, CAPG-GAN, etc.).

[0035] The aforementioned algorithm library is constructed based on the principle of "complementary capabilities": geometric transformation algorithms focus on preserving the authenticity of the original texture with low computational power in near-frontal view; while 3D fitting and generation algorithms work together to solve the problems of spatial structure correction and occluded information recovery in large pose view.

[0036] When registering a model, five head images covering common scenarios need to be run to record the model size, average inference time, peak memory usage, average accuracy, and scene labels.

[0037] When filtering the candidate set, models are first eliminated based on resource consumption constraints, retaining only those whose peak GPU memory does not exceed the currently available memory and whose average inference time does not exceed the latency budget. Then, scene adaptation filtering is performed based on task operation mode and image scene labels, prioritizing models that match the current scene. The models obtained from the initial filtering are ranked by their comprehensive score (S = α⋅accuracy + β⋅latency score), selecting the top-3. If scores are the same, accuracy is prioritized. The comprehensive score uses default weights and can be customized according to project needs. Finally, the candidate set is... Proceed to the next step to perform the evaluation.

[0038] Step 2: Parallel estimation and alignment result generation using multiple algorithms: 2.1 Execution of Candidate Algorithms and Result Collection: The system receives the candidate algorithm set A={a1,a2,…,a...} output from step one. nThe system executes each algorithm under a controlled parallel or sequential scheduling mechanism. Each algorithm aᵢ outputs the corresponding head pose parameters (including yawᵢ, pitchᵢ, rollᵢ), keypoint coordinate set Kᵢ, and normalized image result Nᵢ for the input image I. The system stores these results in an intermediate result buffer M={m1,m2,…,m n This serves as input for subsequent quality assessments.

[0039] 2.2 Geometric Consistency and Key Point Stability Assessment: The system performs geometric consistency analysis on the keypoint set Kᵢ output by each algorithm, and measures attitude stability by calculating the variance of the Euclidean distance between keypoints and the deviation of the deformation matrix. Stability metrics are defined as follows:

[0040] in Indicates the variance of key point locations. The rotation matrix output by the algorithm. As the reference rotation matrix, , These are weighting coefficients. A lower stability index indicates a more stable algorithm result.

[0041] 2.3 Visual quality and geometric matching score: For each normalized image Nᵢ, the system uses an image quality assessment model to calculate scores for sharpness, alignment symmetry, and facial structure integrity. The comprehensive index is defined as follows:

[0042] in For image sharpness index, Scoring for geometric consistency The facial symmetry score is calculated using α, β, and γ, which are weighted values ​​set by the system.

[0043] 2.4 Downstream task suitability assessment: If the system is in task chain mode, the impact of Nᵢ on the performance of subsequent tasks is further evaluated. The system can calculate the recognition confidence or facial expression consistency as the fitness score Aᵢ in the downstream task model, reflecting the task usability of the algorithm results.

[0044] Step 3: Result Quality Assessment and Optimal Selection The system calculates the overall score after normalizing the above indicators:

[0045] in, , , These are preset weighting coefficients used to balance the influence of different evaluation indicators. The weighting coefficients can be set to fixed values ​​based on system design or typical application scenarios, or they can be manually adjusted via a configuration file. Without special configuration, the system uses the default weights to perform scoring calculations.

[0046] The algorithm with the highest overall score is determined to be the optimal algorithm for the current scenario. The algorithm's score and running information are recorded in the scene memory module for use in subsequent strategy updates.

[0047] Step 4: Scene Feature Extraction and Memory Update 4.1 Scene Feature Extraction and Normalization: The system extracts scene feature vectors from the current input. This includes attitude angle features. Light intensity / color temperature, etc. Occlusion ratio Resource status level .

[0048] 4.2 Continuous Variable Clustering and Scene Library Establishment: The system in a continuous variable subspace Cluster analysis is performed to distinguish different scenario types. An incremental K-Means online update algorithm is used. For each cluster... The system maintains the centroid vector of this type of sample. Covariance matrix and sample size To determine the input sample The system uses Mahalanobis distance as the matching metric to measure the similarity between clusters:

[0049] To determine the affiliation, the system calculates the 95th percentile as the radius threshold based on the historical sample distance distribution of each cluster. And set the judgment rules:

[0050] When an input sample meets the criteria for a certain clustering, the system assigns it to the corresponding cluster. And update the statistics (including) If none of the clusters meet this condition, the system treats the input as a new scene and uses its feature vector. Initialize new cluster centers; when the number of new clusters exceeds a preset threshold, trigger an overall re-clustering process.

[0051] 4.3 Registration and Update of Cluster-Level Optimal Algorithm: The system establishes an algorithm performance statistics table S for each scene cluster k. k This table records the historical performance of each candidate algorithm in this scenario. The statistics table includes the algorithm identifier 'a', the average overall score, and so on. Score variance and sample size (a) When a new input sample is determined to belong to a scene cluster Then, if the optimal algorithm for the current frame is The system performs a sliding update on the mean and variance of the corresponding entries for the algorithm in the table, and increments the sample count by one. When the number of samples for a certain algorithm in this scene cluster exceeds a preset threshold... When an algorithm achieves the highest average score, the system marks it as the preferred algorithm for that cluster. This is for quick access later.

[0052] 4.4 Continuous Scene Generalization and Boundary Smoothing Decision Making: To address the smooth transition characteristics of continuous variables such as pose and illumination in real-world scenarios, the system models continuous features as intervals. When the input feature is near the interval boundary, the system simultaneously references the historical best algorithms of neighboring clusters and the comprehensive score, generating the final decision based on feature similarity or confidence.

[0053] Step 5: Self-learning and strategy consolidation: When the system recognizes that the input sample belongs to a scene cluster In this case, the resource screening and multi-candidate evaluation process in steps two through four will be skipped directly, and the fixed preferred algorithm will be called first. Perform inference and normalize the output.

[0054] During system operation, the overall score changes of the fixed algorithm are continuously monitored. When a performance degradation exceeding a threshold is detected or a new algorithm version is registered in the system algorithm library, the system will de-fix the state and return to step two to perform candidate evaluation and strategy update processes to redetermine the new preferred algorithm.

[0055] To further illustrate the technical solution of the present invention, the implementation process of the system of the present invention is illustrated below with examples of different application scenarios.

[0056] Example 1: Access Control and Authentication System Scenario Description: This is a smart access control system for a park or train station. Different entrance / exit cameras vary significantly in angle, lighting, and obstruction conditions, requiring the system to quickly complete identity verification.

[0057] Implementation steps: Step 1: Image Acquisition When a user approaches the access control device, the system captures the input image in real time and extracts basic information such as attitude angle, illumination features and occlusion ratio.

[0058] Step 2: Device Status Detection and Algorithm Invocation The system detects the hardware computing power and latency constraints of the current access control terminal and retrieves the preferred algorithm for the corresponding scenario from the scenario memory. If the scenario already has a fixed strategy, the preferred algorithm is directly called for attitude estimation and normalization.

[0059] Step 3: Real-time normalization output The selected algorithm quickly outputs standardized images with a uniform pose, and the results are passed to the backend identity feature extraction and comparison module for instant identity verification. Only a single optimal algorithm is run in this stage to ensure real-time response performance.

[0060] Step 4: Strategy Monitoring and Update Mechanism The system continuously records the algorithm's performance metrics in this scenario and periodically evaluates its long-term performance. When a performance degradation is detected or a new algorithm version is registered, the system automatically enters "strategy update mode," returns to step 2 to perform an offline multi-candidate evaluation, and redetermines the optimal algorithm for the scenario.

[0061] Step 5: Strategy Consolidation and Rapid Deployment Once the new optimal algorithm is running stably, the system will re-establish it as the preferred algorithm for that scenario category for subsequent real-time calls.

[0062] Implementation results: The system can achieve a uniform posture alignment result at different entrances and exits, improving the stability of the access control verification system under complex lighting and different angles.

[0063] Example 2: Criminal Investigation Evidence Collection and Criminal Identification System Scenario Description: Applied to police criminal investigation video evidence collection scenarios. The input video sources are complex, including street surveillance, ATM camera footage, and dashcam video, with significant differences in resolution, angle, and lighting conditions.

[0064] Implementation steps: Step 1: Resource Awareness and Candidate Algorithm Selection After receiving the image, the system first extracts features such as the input pose angle, illumination intensity, and sharpness, while simultaneously detecting the resource status of the current computing node. Based on scene features and resource constraints, it dynamically selects a set of executable head pose estimation and normalization algorithms from the system's registry.

[0065] Step 2: Parallel execution and result evaluation of multiple candidate algorithms The system runs the selected candidate algorithms in parallel, outputs multiple pose estimation and normalization results, calculates key point stability and other indicators for each algorithm, and generates a corresponding comprehensive score.

[0066] Step 3: Quality Assessment and Determination of Optimal Algorithm The system performs normalized weighted calculations on multiple evaluation metrics to obtain a comprehensive score. The algorithm with the highest comprehensive score is determined as the optimal algorithm for the current scene, and the system outputs its standardized image, pose parameters, and key point information. Simultaneously, the relevant metrics are written to a database for subsequent memorization and updates.

[0067] Step 4: Scene Feature Extraction and Memory Update The system extracts scene features. When the input belongs to an existing cluster, the system updates the statistical information and the optimal algorithm score of that cluster. If it is a new scene, a new cluster is created and the algorithm performance data is recorded.

[0068] Step 5: Continuous Scene Generalization and Boundary Smoothing Decision To address the continuous changes in pose and lighting in the video, the system employs a weighted smoothing strategy in the boundary region of adjacent scene clusters: when the input feature is located at the intersection of two clusters, the system comprehensively refers to the historical best algorithm scores of the two clusters and determines the final selection based on similarity weighting, thereby achieving a smooth transition of algorithm switching and reducing continuous frame jitter.

[0069] Step Six: Self-Learning and Strategy Consolidation The system statistically analyzes the historical performance of algorithms within the same scenario cluster. When an algorithm consistently achieves the highest overall score in a given scenario and its volatility is below a preset threshold, the system designates it as the preferred algorithm for that scenario. If performance degradation is detected or a new algorithm is registered, the system re-enters step two to perform multi-candidate comparison and policy update, enabling self-learning optimization and evolution of the algorithm library.

[0070] Implementation results: The system can still output a uniform standard pose image under multi-source video frame conditions, which improves the consistency of subsequent identity comparison and feature retrieval.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A scalable adaptive full-view head pose estimation and normalization selection method, characterized in that, Includes the following steps: Resource awareness and candidate algorithm screening: After receiving the input image, the system extracts the basic meta-information of the task and generates a task description vector. It detects the computing power status of the device in real time to form a resource status table. It calculates the comprehensive resource index based on the task description vector and the resource status table. If there is a stable optimal algorithm, it will be run directly. Otherwise, it will select a set of candidate algorithms that meet the resource constraints and scenario adaptation requirements from the built-in algorithm registration library. The algorithm performs parallel estimation and alignment result generation, schedules and executes each algorithm in the candidate algorithm set, collects the head pose parameters, key point coordinate set and normalized image results output by each algorithm, evaluates the geometric consistency and stability of the key point set, scores the visual quality and geometric matching degree of the normalized image, and evaluates the downstream task adaptability if it is in task chain mode. Results quality assessment and optimal selection: After normalizing the above evaluation indicators, calculate the comprehensive score, select the algorithm with the highest comprehensive score as the optimal algorithm in the current scenario, and record its score and running information. Scene feature extraction and memory update: Extract the scene feature vector of the current input, perform cluster analysis on the continuous variable subspace to build a scene library, maintain an algorithm performance statistics table for each scene cluster, update the relevant information of the optimal algorithm, and adopt a boundary smoothing decision mechanism for continuous scene features; Self-learning and policy solidification: When the input sample belongs to an existing scene cluster, the preferred algorithm solidified in that cluster is called first. The performance of the solidified algorithm is continuously monitored. If a performance degradation or a new algorithm is registered, the solidification state is deactivated and the candidate evaluation and policy update process is re-executed.

2. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The task description vector includes parameters related to input resolution, frame rate, task priority, and latency constraints.

3. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The resource status table records CPU / GPU utilization, available video memory, power consumption limits, and system latency parameters, and is periodically refreshed to reflect the real-time operating environment.

4. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The built-in algorithm registry covers three representative methods: keypoint alignment algorithm based on geometric transformation, shape regression algorithm based on 3D model fitting, and frontalization algorithm based on deep generative networks.

5. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The candidate algorithm selection process is as follows: first, models that do not meet the requirements are eliminated based on resource consumption constraints; then, scene adaptation is performed based on task operation mode and image scene labels; finally, the top-3 models are selected to form a candidate set according to the comprehensive score. The comprehensive score takes into account both accuracy and latency performance, and accuracy takes precedence when scores are the same.

6. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The geometric consistency and stability assessment is achieved by measuring the degree of positional dispersion between key points and the deviation of the rotation matrix from the reference rotation matrix. The visual quality and geometric matching score covers image sharpness, alignment symmetry, and facial structure integrity indicators.

7. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The scene feature vector includes pose angle features, lighting-related parameters, occlusion ratio, and resource status level.

8. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The clustering analysis employs an incremental online update algorithm. It calculates Mahalanobis distance to determine the similarity between the input sample and each cluster, determines the sample's affiliation based on a preset threshold, and triggers overall re-clustering when the number of newly added clusters exceeds the preset threshold.

9. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, During the self-learning process, when an algorithm has more than a preset threshold in the number of samples in the corresponding scene cluster and has the highest average score, it is marked as the preferred algorithm for that cluster. The performance degradation is judged when the overall score of the algorithm decreases by more than a preset threshold.

10. The scalable adaptive full-view head pose estimation and normalization selection method according to claim 1, characterized in that, The boundary smoothing decision mechanism is as follows: when the input feature is located near the boundary of the scene cluster interval, the final algorithm selection decision is generated by referring to the historical best algorithm and comprehensive score of the adjacent clusters and weighting the feature similarity or confidence.

Images