A method for complex motion perception based on human visual cues

Abstract

The application discloses a complex motion perception method based on human visual inspiration, which comprises the following steps: firstly, capturing the image of an athlete in a motion process, identifying a key image area by evaluating the semantics and visual importance of the image area, and dynamically processing the area features of the key image area; secondly, processing the dynamically processed area features, using a low-rank active learning technology, and obtaining an optimized image area through an optimized objective function; and finally, classifying the optimized image area by using a support vector machine (SVM) to obtain a motion perception result. The application can ensure better calculation efficiency and the visual saliency of the extracted key area, and significantly improves the performance in a complex motion environment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence, computer vision and physical education, and particularly relates to a complex motion perception method based on human visual inspiration. BACKGROUND

[0002] Accurate scene classification is crucial for understanding and analyzing sports related to ball games, especially in educational environments. Just as real-time action statistics such as assists and dunks in NBA games provide valuable insights for spectators, the MobileEye method can enhance sports education by automatically detecting and analyzing key actions such as passing, defending, and shooting techniques. This human vision-inspired method can process real-time sports videos, providing instant feedback, making it a powerful tool for coaches and athletes to analyze performance and decision-making in real time. In addition, in the field of sports safety, it is crucial to accurately identify risks or unsafe actions during training or competition. The MobileEye method can detect dangerous behavior, such as sudden changes in body posture or potential collision risks, through its fast scene perception ability. This ensures real-time monitoring and feedback, helping to prevent injuries by highlighting unsafe actions, thereby promoting a safer training environment.

[0003] However, the prior art has the following significant deficiencies: lack of real-time performance, many motion scene analysis methods are difficult to meet real-time application requirements due to the high complexity of processing multi-modal inputs such as athlete positions, ball trajectories, and action dynamics; low action classification accuracy, current methods often have low accuracy due to inaccurate data set labeling or insufficient model capture ability for fast motion scenes; lack of human visual inspiration in model design, most methods fail to effectively simulate human visual attention allocation characteristics in dynamic scenes, and lack of attention to key action areas; poor adaptability, existing models are difficult to efficiently extract and utilize multi-modal features in different ball games or diverse scenes, resulting in significant decline in classification performance SUMMARY

[0004] To overcome the deficiencies of the prior art, the present application provides a complex motion perception method based on human visual inspiration to address the many deficiencies of traditional methods in the field of invention.

[0005] The complex motion perception method based on human visual inspiration mainly consists of four stages, and the specific process is as follows:

[0006] S1. Capture athlete images during the motion process, and identify key image regions by evaluating the semantics and visual importance of image regions.

[0007] S2. Dynamically process the region features of the key image regions.

[0008] S3. For the dynamic processed region features, a low-rank active learning (LAL) technique is used to obtain an optimized image region through an optimized objective function.

[0009] S4. The optimized image region is classified using a support vector machine (SVM) to obtain a motion-aware result.

[0010] Further, the BING objectness measure method is used in step S1 to extract the region features of the image, which is very effective in quickly extracting high-quality image blocks (referred to as "image regions") related to objects. In the present application, the concept of feature margin is used to quantify the correlation strength between a specific feature and a class label. The margin of feature V is defined as follows:

[0011]

[0012] where |E| represents the total number of samples in the data set. N(v (n) ) represents the nearest erroneous sample, i.e., the nearest sample that does not belong to the same class as v (n) . I(v (n) ) represents the nearest correct sample, i.e., the nearest sample that belongs to the same class as v (n) .

[0013] Further, in addition to evaluating the discrimination ability of a single feature, it is also crucial to explore the interaction between features. The relationship between features is quantified by calculating the information gain, which is as follows:

[0014] G(V|W)=EN(V)-EN(V|W)

[0015] where EN(V) is the entropy of feature V; EN(V|W) is the conditional entropy of feature V, which is given the observation of feature W.

[0016] Symmetric uncertainty can be used to normalize the information gain (IG) to provide a consistent measure of the dependence between features, which can be represented as follows:

[0017]

[0018] In order to evaluate the effectiveness of the discrimination ability of a certain feature relative to its processing time, the present application introduces an index called the discrimination ability to time cost ratio. For each transformed feature W′ i in the nonlinear domain, the balance between its discrimination ability and the required calculation is evaluated, as well as the correlation between features in terms of calculation requirements. For feature W′ i , the discrimination ability to time cost ratio TD(W′ i ), the time correlation ratio TC(V′ i ,W′ i ), is as follows:

[0019]

[0020] where V i and W i are original features, transformed into target features V′ i and W′ i respectively, T(W i ) is the total time of feature W i processing, and p∈(0,+∞) is its degree.

[0021] According to the values of each region TD, the first G1 regions are selected to calculate their TC values, and according to the values of TC, the first G2 regions are sorted as key image regions.

[0022] Further, the step S2 employs dynamic selection of the features of these regions, and the invention utilizes a self-expression model to maintain the geometric structure of key image regions, which assumes that all samples (i.e. object regions) exist in the combination of multiple subspaces, so that each sample can be described by the linear combination of other samples, but not including itself. This relationship is mathematically expressed as I = IVwhere diag(V), I = [i1,…,i N ] represents a matrix containing multiple visual features of N training blocks, and V is the extracted image region feature which is also a matrix of self-reconstruction coefficients. In order to consider potential sample pollution, the model is modified as I = IV+L, where L is an error matrix. The refined self-expression model is obtained as follows:

[0023]

[0024] where ||·||τ1 v and ||·||τ2 w represent predefined matrix norms, and τ1 and τ2 are non-negative weights of these norms, respectively.

[0025] The matrix K is defined as |V|+|V T |, which acts as a weight matrix to quantify the relationship between all samples. If i i and i j belong to different categories, the proximity between Mi i and should be greater, where the element in the ith row and jth column represents the weight K ij ; the following objective function is obtained:

[0026]

[0027] where ⊙ represents the Hadamard product operation, n iThe label of the i-th sample in the respective class. M is the dimension reduction matrix, and matrix N contains the class labels of all training samples. A ij = a ||Bi i -B ij || 2 +(1-a)||n-n j || 2 is a comprehensive index to evaluate the similarity between sample i i and i j , which is based on their key image regions (Bi i , Bi j ) and their labels (n i , n j ), where a ∈ (0, 1) is a balance parameter between the two aspects.

[0028] In the present invention, the l1 norm is applied to ||V|| v and ||L|| w to promote the sparsity of the representation matrix and the error matrix. In order to ensure the column sparsity of L, the l 1,2 norm is adopted for feature selection. The objective function can be changed to the following:

[0029]

[0030] where τ3 is an additional non-negative weight factor, which deals with the key image regions through the improved objective function.

[0031] Further, step S3 adopts a low-rank active learning (LAL) technique to generate gaze shift paths (GSPs). The present invention adopts a novel active learning method to simulate how humans interact with different image regions. The present invention aims to minimize the overall reconstruction error to effectively identify and select the most meaningful object regions in the scene. The overall reconstruction error is minimized, and the formula is as follows:

[0032]

[0033] where λ > 0 is a tuning parameter that customizes the regularization term aiming to enhance the specificity of the model. The main goal of this function is to ensure high-quality reconstruction of the input scene patches, where T denotes the reconstruction coefficient matrix. C is the key image region processed by the objective function. Meanwhile, the additional term introduces a specific matrix norm for regularization, aiming to achieve high reconstruction accuracy, especially for the K most important object patches. This strategy significantly prioritizes these patches while reducing the focus on the N-K patches that are less important. For example, if all elements in a certain row of T are zeros, it indicates that the corresponding patch is considered non-essential and does not contribute to the reconstruction of other patches. Therefore, T is designed to exhibit row sparsity, with each row indicating the reconstruction importance of a specific patch.

[0034] To enhance the sparsity of matrix T, the norm ||T||1 m is adjusted to ||T||1 2,1 , considering that the target region lies within a subspace in the higher-dimensional feature space, T needs to be a low-rank matrix. The subsequent objective function optimization is:

[0035]

[0036] where η > 0 is the weight of the regularization term, and rank(·) is used to evaluate the rank of the matrix, aiming to minimize rank(T) to ensure that T remains low-rank, thus preserving the inherent set configuration of the input data.

[0037] Since the optimization objective is an NP-hard problem, using a convex surrogate method to approximate the minimization of rank, the optimization objective function can reduce the computational burden while maintaining the quality of the solution.

[0038] Further, the step S4 effectively classifies the image regions using support vector machines (SVM). For each of the L optimized image regions, a convolutional neural network (CNN) trained on ImageNet is used to extract 128-dimensional deep features from each region. These 128-dimensional deep features from the L regions are then combined into a single feature vector β of length 128L, creating a comprehensive representation of each sports scene. This approach captures the complexity of the game, including the movement of players, the trajectory of the ball, and the important actions on the field.

[0039] For each of the P optimized image regions, each scene exhibits a different sports action, and the task of each binary SVM classifier is to distinguish the sports action by utilizing the combined feature vector, enabling the model to achieve accurate classification results. The optimization objective of the support vector machine classifier can be represented as follows:

[0040]

[0041] where, δ j represents the jth optimized image region, and the feature vector represents a specific sports action. The label o j represents the presence or absence of the jth action type, where o j = 1 represents the presence of the action, o j = -1 represents the absence of the action, and h(·) is a function for calculating a hash value. The variable β represents a coefficient of a hyperplane for distinguishing different sports action categories in a feature space, and M kl refers to the number of optimized image regions associated with the action pair k and l. The condition ensures that a support vector machine (SVM) classifier effectively separates each pair of sports actions, thereby optimizing the decision boundary to achieve accurate classification.

[0042] Advantages of the present application:

[0043] The complex motion perception method based on human visual inspiration provided by the present application first extracts the region features of an image by using a BING objectivity measurement method, designs two indexes of distinguishing ability and time cost, and can ensure better calculation efficiency while ensuring that the extracted key regions have visual saliency. The self-representation model is optimized, the low-rank active learning technology is used, the feature selection and distinguishing ability of the method are stronger, the performance in a complex motion environment is significantly improved, the support vector machine is used for classifying the optimized regions, the recognition accuracy is high, and the method exhibits excellent accuracy in recognizing key motion actions and strategies. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The present application is a work flow diagram. DETAILED DESCRIPTION

[0045] Visual cognition and psychology research shows that human eyes naturally tend to focus on regions with significant semantics or visual importance in a building scene. This observation suggests that our visual attention is intentionally focused on specific key elements rather than randomly dispersed. Based on this direction, the present application develops a complex motion perception method based on human visual inspiration, which combines the detection of object center patches and the application of low-rank active learning (LAL). The construction of LAL assumes that adjacent scene patches in space are closely distributed in the feature space because of their similar visual appearance and semantics. In theory, the strategy of the present application selectively targets those patches that are visually striking and semantically meaningful, reflecting the inherent tendency of human visual perception.

[0046] In everyday environments, people instinctively focus their attention on salient elements or objects, such as vehicles or architectural features, that stand out due to their importance and spatial location in the scene. To identify these attention-grabbing elements, the present invention employs the BING Objectness metric. This tool is very effective at rapidly isolating high-quality object-related image regions (i.e., "landscape patches"). The BING metric is particularly valued for its speed in quickly localizing relevant patches, its key role in creating Gaze Shift Paths (GSPs) by providing a wide range of object-specific patches, and its strong generalization ability across multiple object categories. This flexibility significantly improves the versatility of our scene classification framework across diverse datasets. Once these BING patches are isolated, we will characterize them in detail through high-quality feature extraction, the details of which will be elaborated in subsequent sections.

[0047] The human vision-inspired complex motion perception method of the present invention is further described below in conjunction with the accompanying drawings.

[0048] (1) Identify key image regions by evaluating their semantic and visual importance

[0049] To improve the performance and efficiency of semantic recognition, the present invention introduces a new feature selection method called Time-Discriminative (TD) and Time- Correlation (TC) based feature selection. This technique is divided into two stages: the first stage removes features that lack sufficient discriminative power through the Time-Discriminative (TD) index; the second stage eliminates features that exhibit high redundancy through the Time-Correlation (TC) index. This two-stage approach ensures that only the most effective and relevant features are retained, thereby improving the speed and accuracy of the emotion recognition method. Specifically as follows.

[0050] Using the BING Objectness metric method, which is very effective at rapidly isolating high-quality object-related image patches (referred to as "image regions"). Once these BING features are extracted, they will be characterized in detail through high-quality feature extraction.

[0051] The present invention comprehensively evaluates each image region through three main performance indicators: processing speed, discriminative power, and interaction between different features. In the S1 stage, the processing time is divided into two parts: the time required for feature extraction and the time required for feature classification. The total processing time of any selected feature, denoted as V, covers the following two stages

[0052] T(V) = T f (V) + T d (V)

[0053] where T f represents the time required for feature extraction, Td time needed for classification

[0054] The ability of a feature to distinguish different types of image regions at the image patch level is referred to as its discriminative effectiveness, which is evaluated by its correlation with class labels. This concept is important in image classification, object detection, and other tasks, especially in feature selection. In the present invention, the concept of feature margin is adopted to quantify the strength of the association between a particular feature and class labels. It can be quantified by evaluating the discriminativeness that a feature provides for different class labels in a classification task. Simply put, the margin can reflect the role that the feature plays in distinguishing classes. The larger the feature margin, the greater the contribution of the feature to class distinction. The margin of feature V is defined as follows:

[0055]

[0056] where |E| represents the total number of samples in the dataset. N(v (n) ) represents the nearest erroneous sample, i.e., the sample closest to v (n) but not belonging to the same class, and I(v (n) ) represents the nearest correct sample, i.e., the nearest sample belonging to the same class as v (n) . The discriminative ability of a feature is evaluated by calculating the distance between the nearest erroneous sample and the nearest correct sample. A larger difference between these distances indicates that the feature has a stronger ability to distinguish between classes.

[0057] In addition to evaluating the discriminative ability of individual features, it is also crucial to explore the interactions between features. Interactions between features can significantly affect the predictive ability of a model, especially in complex nonlinear tasks. By identifying and modeling these interactions, it is possible to improve the performance of the model. When two highly discriminative features are highly correlated, it may indicate the presence of redundancy, providing an opportunity to improve the efficiency of the model by removing one of the redundant features. This strategy not only simplifies the feature extraction and classification process but also improves the overall performance of the method. The correlation between features is usually evaluated by information gain, which is determined based on the entropy of the relevant features. Information gain evaluates the importance of a feature by calculating the reduction in uncertainty (reduction in entropy) when the feature is used to partition the dataset. The larger the information gain, the greater the role of the feature in distinguishing between classes. The entropy of feature V is defined as follows:

[0058]

[0059] where P(v (n) ) represents the probability of v (n)

[0060] The conditional entropy of feature V, given the observation of feature W, can be represented as follows: ​

[0061]

[0062] The relationship between features can be quantified by computing the information gain, which assesses the reduction in uncertainty about one feature by knowing another, specifically:

[0063] G(V|W) = EN(V) - EN(V|W)

[0064] Symmetric uncertainty can be used to normalize the information gain (IG) to provide a consistent measure of dependence between features, which can be expressed as:

[0065]

[0066] To evaluate the effectiveness of a feature's discriminative ability relative to its processing time, the present invention introduces an index called the discriminative ability to time cost ratio. For a feature V, this ratio is calculated as follows:

[0067]

[0068] where p e (0, +∞) allows for adjustment according to the processing time required for each feature to adapt to more situations. A higher p value indicates greater concern for the time efficiency of the feature, and adjusting p values for different needs allows for the ability and efficiency to meet specific needs.

[0069] In addition, the time-dependent ratio is determined by the following equation:

[0070]

[0071] where Q represents the selected feature set, and R represents a certain feature currently being evaluated. This index is used to evaluate the impact of excluding R from the feature set, which helps to optimize the feature space.

[0072] Visual features typically exhibit a nonlinear distribution within their domain. In the field of computer vision and image processing, the nonlinear distribution of features means that their changes are not simply linear, but complex and highly dependent on the distribution characteristics of the input data. Therefore, in order to better understand their interactions, the present invention projects the input features into a nonlinear domain. In this transformed space, each feature combines multiple original attributes, thereby enhancing the depth of analysis. This approach is particularly effective in fields such as semantic recognition, as it employs a variety of nonlinear techniques.

[0073] For each transformed feature W' in the nonlinear domain i , the present invention evaluates the balance between its discriminative ability and the required computation, as well as the correlation between features in terms of computational demand, as follows:

[0074]

[0075] Where V i and W i are the original features, which are converted into target features V′ i and W′ i .

[0076] Sort by the TD value of each region, select the first G1 region to calculate its TC value, and sort according to the TC value, and use the first G2 regions as the key image region.

[0077] (2) Dynamic processing of image region features

[0078] The present invention uses machine learning to dynamically process the features of image regions. The present invention uses a self-representation model to maintain the geometric structure of key image regions. The self-representation model is a machine learning model commonly used for subspace clustering. Its basic principle is to use the correlation between objects in the data set to construct a low-dimensional representation space, so that each object can be approximately represented by a linear combination of other objects. At the same time, high-quality features are selected for each object patch. The model assumes that all samples (i.e., object regions) exist in a combination of multiple subspaces, so that each sample can be described by a linear combination of other samples, but not itself. This relationship is mathematically expressed as I=IV where diag(V), I=[i1,…,i N ] represents a matrix containing multiple visual features of N training blocks, and V is the matrix of self-reconstruction coefficients. To consider potential sample contamination, the model is modified to I = IV + L, where L is the error matrix. The refined self-representation model is as follows:

[0079]

[0080] where ||·|| v and ||·|| w represents the predefined matrix norm, and τ1 and τ2 are the non-negative weights of the norm.

[0081] Introduce a matrix M that projects the original gaze path transfer matrix (GSP) features into a low-dimensional space. In order to achieve effective feature selection (FS), the matrix M needs to be sparse on the columns. If two samples i i and i j From the same category, their projection features Mi in low-dimensional space i and Mi j should show similarity, thus giving a higher weight Mi to the connection between them j In addition, the matrix K is defined as |V|+|VT |, which acts as a weight matrix quantifying the relationship between all samples. If i i and i j belong to different classes, then Mi i and Mi j should be greater, resulting in a lower weight K ij . These relationships and assumptions are summarized in the following objective function:

[0082]

[0083] where ⊙ denotes the Hadamard product operation, n i represents the label of the i-th sample in the respective class; M is the dimensionality reduction matrix, and matrix N contains the class labels of all training samples; A ij = α||Bi i -Bi j || 2 +(1-α)||n i -n j || 2 is a comprehensive index that evaluates the similarity between sample i i and i j , based on their key image regions (Bi i , Bi j ) and their labels (n i , n j ), where α∈(0,1) is a balance parameter between the two aspects.

[0084] By integrating these concepts, the objective function for feature selection can be reconstructed to simultaneously include the self-representation model that preserves geometric structure and the label consistency model, allowing the model to better understand and maintain label consistency while preserving the geometric structure information of the data. This approach is particularly useful when dealing with visual data, especially in cases of limited data or incomplete labels. This optimization method ensures that the selected features not only preserve the inherent geometric structure of the data but also accurately correspond to the label information of the samples.

[0085]

[0086] In the present invention, the l1 norm is applied to ||V|| v and ||L|| w to promote the sparsity of the representation matrix and the error matrix. To ensure the column sparsity of L, which is crucial for feature selection, the l 1,2 norm is adopted. A loss function is minimized by multiple regularization terms, combining sparsity, low-rank constraint, and smoothness. The objective function can be changed to the following:

[0087]

[0088] s, t. I = IV, diag(U) = 0.

[0089] where τ3 is an additional non-negative weight factor that handles the critical image regions by the improved objective function.

[0090] (3) The low-rank active learning (LAL) technique is used to determine the image regions that can attract human attention, and gaze shift paths (GSPs) are generated

[0091] The low-rank active learning (LAL) technique is used to determine the image regions that can attract human attention, and gaze shift paths (Gaze Shift Paths, GSPs) are generated. By analyzing the shift of gaze points, the visual search patterns, cognitive strategies and attention focusing areas of the observer in a specific situation can be revealed. The present invention uses a novel active learning method to simulate how humans interact with different image regions. Generally, many regions, especially those depicting background elements, lack clear semantic meaning and usually do not attract human attention. To solve this problem, a technique called low-rank active learning (LAL) is used, which is specifically designed to identify regions that are semantically rich in different scenes, thereby enhancing the model's ability to focus on regions that are important to human perception.

[0092] The method of the present invention aims to minimize the reconstruction error to effectively identify and select the most meaningful image regions in the scene. This objective can be expressed by the following formula:

[0093]

[0094] where, represents the reconstruction error, i.e. the difference between the original data C and the data CT processed by the transformation matrix T. The goal is to convert the original image C into a more meaningful form through the transformation matrix T and minimize the error in the conversion process, through ||T||2. mA regularization constraint is imposed on the transformation matrix T to control its sparsity. This regularization term encourages the elements of matrix T to be sparse, such that the resulting transformation matrix T only retains those regions or features that are most significant, thereby identifying important regions in the image. λ ≥ 0 is a tuning parameter that customizes the regularization term, aiming to enhance the specificity of the model. The main goal of this function is to ensure high-quality reconstruction of the input scene patches, where T represents the reconstruction coefficient matrix. C is the input sample matrix. Meanwhile, the additional term introduces a specific matrix norm for regularization, aiming to achieve high reconstruction accuracy, especially for the K most significant object patches. This strategy significantly prioritizes these patches while reducing the focus on the N-K patches that are less important. For example, if all elements in a certain row of T are zero, it indicates that the corresponding patch is considered non-essential and does not contribute to the reconstruction of other patches. Therefore, T is designed to exhibit row sparsity, with each row indicating the reconstruction importance of a specific patch.

[0095] To enhance the sparsity of matrix T, the norm ||T||1 m can be adjusted to ||T||1 2,1 or ||T||1 ∞ . In this invention, ||T||1 2,1 is chosen. Matrix T plays two roles in the framework of this invention (i) to capture the reconstruction parameters, with each column representing the combination of key object patches needed to reconstruct the entire scene image; (ii) as a self-representative matrix, where each column s N from the set T i serves as a feature vector . C is an implicit dictionary.

[0096] Considering that the object patches lie within a subspace of a higher-dimensional feature space, T needs to be a low-rank matrix. This requirement influences the refined version of the objective function described later:

[0097]

[0098] Among them, η≥0 is used as the weight of the regularization term, rank(·) is used to evaluate the rank of the matrix, and its goal is to minimize rank(T) to ensure that T remains low-rank, thereby preserving the intrinsic set configuration of the input data. However, minimizing the rank of the matrix, as required by the initial goal, is an NP-hard problem. NP-hard problems refer to problems whose correctness cannot be verified, or the verification process requires super-polynomial time, which is difficult to solve. In order to deal with NP-hard problems, this method uses a convex alternative method to approximate rank minimization, such as the nuclear norm. The nuclear norm is a convex function, so convex optimization theory and algorithms can be applied to solve the problem, avoiding the computational complexity of NP-hard problems. Although the nuclear norm is not equal to the rank of the matrix, it provides a good approximation in many applications. By minimizing the nuclear norm, a low-rank matrix can usually be obtained, retaining most of the structural information of the matrix. Nuclear norm optimization is much more computationally efficient than direct rank minimization, especially in large-scale matrices and high-dimensional data. It can reduce the computational burden while maintaining the quality of the solution, as follows:

[0099]

[0100] where ||T|| * stands for the nuclear norm, which is used to transform the optimization problem into a convex form, making the solution process more manageable, as elaborated in the literature on unsupervised learning algorithms. This reformulation enables efficient solution of the matrix S and helps determine the K most representative frames that capture the essence of each object patch sequence.

[0101] The goal of this function is to precisely locate the key object patches using low-rank representation. Typically, the reconstruction error is computed using the input sample matrix C. However, samples (such as object regions) can be affected by noise or other distortions. In this case, directly reconstructing each sample can reduce the effectiveness of model training. To this end, the present method employs a dictionary-based approach to process training samples, where each new sample is represented by a series of reconstruction parameters. Dictionary learning is an unsupervised learning method that aims to learn a set of basis functions (dictionary elements) from training data, such that each sample can be represented as a sparse linear combination of dictionary elements. Each sample is represented by a set of sparse reconstruction coefficients, which correspond to the weights of each element in the dictionary. The core idea of this approach is that even if the data itself is affected by noise or distortion, it can still be well represented by appropriate dictionary elements, thereby reducing the impact of noise. This approach combines subspace learning with low-rank constraints, enhancing the model's robustness to noise. Therefore, the strategy of the present method aims not only to maintain a low-rank structure but also to optimize sample selection based on this structure. This results in a robust objective function that is resilient to noise and improves the efficiency of sample selection, ultimately forming a refined objective function that performs well in sample selection while maintaining a low-rank framework. As follows:

[0102]

[0103] where μ ≥ 0 is used to adjust the emphasis of sample reconstruction, where μ = 1 is set to ensure optimal performance. Matrix U is the reconstruction dictionary

[0104] (4) Efficiently classifying image regions using support vector machines (SVMs)

[0105] Efficiently classifying image regions using support vector machines (SVMs), in the development of MobileEye for complex action perception related to ball sports, the present method focuses on extracting key visual features from each scene to enable fast and accurate understanding of sports actions. For each sports scene containing L key regions, a convolutional neural network (CNN) trained on ImageNet is used to extract 128-dimensional deep features from each region. These 128-dimensional features from the L regions are then combined into a single feature vector β of length 128L, creating a comprehensive representation of each sports scene. This approach captures the complexity of the game, including the movement of players, the trajectory of the ball, and important actions on the field.

[0106] These aggregated feature vectors are used to construct training data for a multi-class support vector machine (SVM), which is trained to classify different ball-related sports actions such as passing, dribbling, and shooting. For P training scenes, each exhibiting a different sports action, the present method constructs a binary classifier The task of each binary SVM classifier is to distinguish between the sports actions by utilizing the combined feature vector, enabling the model to achieve an accurate classification result. The optimization objective of the support vector machine classifier can be expressed as follows:

[0107]

[0108] where δ j represents the feature vector of the jth training scene, which represents a specific sports action. The label o j indicates the presence or absence of the jth action type, where o j = 1 indicates the presence of the action, and o j = -1 indicates the absence of the action, and h(·) is a function that calculates the hash value. The variable β represents the coefficients of the hyperplane used to distinguish between different sports action categories in the feature space, and M kl refers to the number of training scenes associated with the action pair j and l. The condition ensures that the support vector machine (SVM) classifier effectively separates each pair of sports actions, thereby optimizing the decision boundary to achieve accurate classification. This structure enables MobileEye to accurately identify and classify different sports actions related to the ball, such as passing, scoring, or tackling, providing real-time, actionable insights, and enhancing the educational significance of sports analysis.

[0109] Compared with the prior art,

[0110] In summary, the present invention has the following advantages: 1. Computational efficiency and visual saliency: By adopting a low-rank active learning paradigm, the present invention can quickly identify the most visually salient regions in each sports scene. This not only improves computational efficiency, but also enhances accuracy, especially in fast-moving sports scenes where quick and accurate identification is crucial. 2. Feature selection and discrimination ability: The feature selection algorithm designed by the present invention optimizes computational efficiency and the discrimination ability of the selected features. These algorithms adjust the weights of feature channels based on their importance, significantly improving performance in complex sports environments where accuracy and resource management are crucial. 3. High accuracy in recognition: The present invention has high accuracy in identifying key sports actions and strategies, demonstrating excellent accuracy in this area.

[0111] To validate the capabilities of the present method, a simple experiment was designed. The discriminative power of 128L deep gaze shift path (GSP) features was evaluated within the multi-class support vector machine (SVM) framework, following the structure outlined in previous studies. The evaluation focused on the ability of MobileEye to accurately classify sports actions related to the ball, such as passing, dribbling, and shooting, utilizing human vision dynamics-inspired visual features. Subsequently, the present invention was compared with seven state-of-the-art deep learning models, which have demonstrated excellence in handling sports-specific data and action recognition across different sports scenarios. The evaluation utilized a comprehensive dataset from a leading study in the field of sports action recognition.

[0112] The present invention was rigorously compared with leading models to benchmark its performance in the evolving field of sports scenario recognition. The results, as shown in Table 1, highlight the precise and reliable classification capabilities of the present method for sports actions related to the ball, enhancing its suitability for educational platforms and real-time sports analysis. To ensure the reliability of the results, each test was conducted 20 times, and the standard deviation was recorded to assess the consistency of the results. The evaluation consistently demonstrated that the present method outperformed the competing models in terms of classification accuracy and stability.

[0113] Table 1 Experimental Results

[0114] Dataset R-CNN Fast R-CNN Faster R-CNN The invention Scene-15 0.863±0.012 0.901±0.012 0.919±0.013 0.922±0.008 Scene-67 0.682±0.014 0.710±0.012 0.739±0.012 0.750±0.006 ZJU Aerial 0.782±0.013 0.792±0.013 0.804±0.011 0.835±0.007 ILSVRC-2010 0.384±0.012 0.408±0.013 0.409±0.010 0.443±0.008 SUN397 0.468±0.013 0.528±0.013 0.517±0.012 0.554±0.008 Places205 0.441±0.013 0.480±0.013 0.496±0.011 0.515±0.007 MSEI 0.497±0.011 0.627±0.014 0.642±0.013 0.731±0.005

[0115] After experimental comparison, the final experimental results were summarized as follows: (1) the present method consistently provided competitive accuracy rates in each sports category, while maintaining significantly lower standard errors; (2) the minimum standard deviation associated with the present method emphasized its robustness, showing more stable performance compared to other models in repeated trials.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A complex motion perception method based on human vision inspiration, characterized by: The following steps are involved: S1. Capture images of athletes in motion and identify key image regions by evaluating their semantic and visual importance. The specific implementation process is as follows: The marginal value of the feature is used to quantify the strength of the association between the feature and the category label. The marginal value of the feature V is defined as follows: Where |E| is the total number of samples in the data set; N(v (n) ) is the nearest error sample, that is, the distance v (n) The nearest sample but not of the same type; I(v (n) ) is the nearest correct sample, that is, (n) The most recent sample of the same class; The relationship between features is quantified by calculating information gain, as follows: G(V|W)=EN(V)-EN(V|W) Where EN(V) is the entropy of feature V; EN(V|W) is the conditional entropy of feature V, expressed given the observation of feature W; Symmetric uncertainty is used to normalize the information gain, which is expressed as follows: For each transformed feature W′ in the nonlinear domain i , evaluate the balance between its discriminative ability and the required computation, as well as the correlation between features in terms of computational requirements, for feature W′ i , the ratio of capacity to time cost TD(W′ i ), time correlation ratio TC(V′ i , W′ i ), as shown below: Where V i and W i are the original features, which are converted into target features V′ i and W′ i , T(W i ) is the feature W i The total processing time, where p∈(0,+∞) is the number of times; Sort by the TD value of each region, select the first G1 region to calculate its TC value, and sort according to the TC value, and use the first G2 regions as the key image region; S2. Dynamically process the regional features of key image areas. The specific implementation process is as follows: The self-representation model is used to maintain the geometric structure of the key image area, which is expressed as I = IV where diag(V) = 0, I = [i1,…,i N ] represents a matrix of multiple visual features containing N training blocks, and V is the matrix of extracted image region features and self-reconstruction coefficients. The self-representation model is modified to I = IV + L, where L is the error matrix. The refined self-representation model is as follows: where ||·|| v and ||·|| w represents the predefined matrix norm, τ1 and τ2 are the non-negative weights of the norm; The matrix K is defined as |V|+|V T |, acts as a weight matrix, where the element in row i and column j represents the weight K ij ; The following objective function is obtained: Where ⊙ represents the Hadamard product operation, n i Represents the label of the i-th sample in each category; M is the dimensionality reduction matrix, and the matrix N contains the category labels of all training samples; A ij =α||Bi i -Bi j || 2 +(1-α)||n i -n j || 2 Is a comprehensive indicator to evaluate sample i i and i j The similarity between them is evaluated based on their key image regions (Bi i ,Bi j ) and its label (n i , n j ), where α∈(0,1) serves as a balance parameter between these two aspects; Based on the refined self-representation model, the objective function is improved and the objective function is obtained as follows: Where τ3 is an additional non-negative weight factor; the key image area is processed by the improved objective function; S3. Using low-rank active learning technology to optimize the objective function of the dynamically processed regional features, an optimized image region is obtained. S4. Use support vector machine (SVM) to classify the optimized image area and obtain motion perception results.

2. The complex motion perception method based on human vision inspiration according to claim 1, characterized in that: The specific implementation process of step 3 is as follows: Minimize the overall reconstruction error, the formula is as follows: Where λ ≥ 0 is a tuning parameter used to customize the regularization term; T represents the reconstruction coefficient matrix; C is the key image area after the objective function is processed; The subsequent objective function is optimized as: Among them, η ≥ 0 is used as the weight of the regularization term, rank(·) is used to evaluate the rank of the matrix, and its goal is to minimize rank(T); The objective function is optimized by approximating the rank minimization using a convex surrogate method.

3. The complex motion perception method based on human vision inspiration according to claim 2, characterized in that: The specific implementation process of step S4 is as follows: For the optimized image regions, a convolutional neural network trained on ImageNet is used to extract deep features from each region; the deep features from L regions are combined into a single feature vector β; For P optimized image regions, the task of each binary SVM classifier is to distinguish motion actions by using the combined feature vector. The optimization objective of the support vector machine classifier is expressed as follows: Among them, δ j Represents the jth optimized image region, label o j Indicates the existence or non-existence of the j-th action type, where o j =1 indicates the existence of action, o j =-1 indicates the absence of an action; h(·) is the function for calculating the hash value, M kl refers to the number of optimized image regions associated with the action pair k and l; the condition Ensure that the support vector machine (SVM) classifier can effectively separate each pair of sports actions.

Images