Complex motion sensing method based on human visual inspiration

Abstract

The invention discloses a complex motion perception method based on human visual inspiration, and the method comprises the steps: firstly capturing an athlete image in a motion process, recognizing a key image region through evaluating the semantic and visual importance of an image region, and then dynamically processing the region features of the key image region. And then, for the dynamically processed region features, a low-rank active learning technology is adopted, an objective function is optimized, and an optimized image region is obtained. And finally, classifying the optimized image area by using a support vector machine (SVM) to obtain a motion sensing result. According to the method, the extracted key region can be ensured to have visual saliency while better calculation efficiency is ensured, and the performance in a complex motion environment is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the fields of artificial intelligence, computer vision and physical education, and in particular relates to a complex motion perception method based on human vision inspiration. Background Art

[0002] Accurate scene classification is essential for understanding and analyzing ball-related sports, especially in educational settings. Just as real-time action statistics such as assists and dunks in NBA games provide valuable insights to viewers, the MobileEye approach is able to enhance sports education by automatically detecting and analyzing key actions such as passing, defense, and shooting skills. This human vision-inspired approach is able to process real-time sports footage and provide instant feedback, making it a powerful tool for coaches and athletes to enable instant analysis of performance and decision-making. In addition, in the field of sports safety, it becomes crucial to accurately identify risky or unsafe actions during training or competition. The MobileEye approach is able to detect dangerous behaviors, such as sudden changes in body posture or potential collision risks, through its rapid scene perception capabilities. This ensures real-time monitoring and feedback, which helps prevent injuries by highlighting unsafe actions, thereby promoting a safer training environment.

[0003] However, the existing technologies have significant deficiencies in the following aspects: insufficient real-time performance. Many motion scene analysis methods are difficult to meet the needs of real-time applications due to the high complexity of processing multimodal inputs (such as player positions, ball trajectories, motion dynamics, etc.); low motion classification accuracy. Current methods often have low accuracy in motion classification due to inaccurate data set annotations or insufficient model capture of fast-motion scenes; lack of model design inspired by human vision. Most methods fail to effectively simulate the attention allocation characteristics of human vision in dynamic scenes and pay insufficient attention to key action areas; poor adaptability. In different ball sports or diversified scenarios, existing models find it difficult to efficiently extract and utilize multimodal features, resulting in a significant decrease in classification effect. Summary of the invention

[0004] In order to make up for the deficiencies of the prior art, the present invention provides a complex motion perception method based on human vision inspiration to solve many deficiencies of traditional methods in the field of invention.

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

[0006] S1. Capture images of athletes during motion and identify key image regions by evaluating their semantic and visual importance.

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

[0008] S3. For the regional features after dynamic processing, the low-rank active learning (LAL) technology is used to optimize the objective function to obtain the optimized image region.

[0009] S4. Use support vector machine (SVM) to classify the optimized image area and obtain motion perception results.

[0010] Furthermore, in step S1, the BING objectness metric method is used to extract the regional features of the image, which is very effective in quickly extracting high-quality image blocks related to objects (called "image regions"). In the present invention, the concept of feature margin is used to quantify the strength of association between specific features and category labels. 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 error sample, that is, the distance v (n) The closest sample but not belonging to the same category, I(v (n) ) represents the most recent correct sample, that is, (n) The closest sample of the same class.

[0013] Furthermore, in addition to evaluating the discriminative power of individual features, it is also crucial to explore the interactions between features. The relationship between features is quantified by calculating the information gain, as follows:

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

[0015] Among them, EN(V) is the entropy of feature V; EN(V|W) is the conditional entropy of feature V, expressed under the observation of given feature W.

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

[0017]

[0018] In order to evaluate the effectiveness of a feature's discriminative ability relative to its processing time, the present invention introduces an indicator called the ratio of discriminative ability to time cost. 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 follows:

[0019]

[0020] 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 characteristic W i The total processing time, where p∈(0,+∞) is the number of times.

[0021] 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.

[0022] Furthermore, the step S2 adopts the feature of dynamically selecting these regions. The invention uses a self-representation model to maintain the geometric structure of the key image region. 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 = [i 1 ,…,i N ] represents a matrix containing multiple visual features of N training blocks, and V is the matrix of extracted image region features and self-reconstruction coefficients. In order 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:

[0023]

[0024] where ||·|| v and ||·|| w represents the predefined matrix norm, where τ 1 and τ 2 are the non-negative weights of these norms respectively.

[0025] The matrix K is defined as |V|+|V T |, acts as a weight matrix, quantifying the relationship between all samples. i and i j Belong to different categories, then Mi i and The closer the pair is, the greater the proximity should be, where the element in the i-th row and j-th column represents the weight K ij ; The following objective function is obtained:

[0026]

[0027] where ⊙ represents the Hadamard product operation, ni Represents the label of the i-th sample in each category. M is the dimension reduction matrix, and matrix N contains the category labels of all training samples. ij =α||Bi i -B ij || 2 +(1-α)||nn j || 2 Is a comprehensive indicator used 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.

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

[0029]

[0030] Among them, τ 3 is an additional non-negative weight factor that processes the key image areas through the improved objective function.

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

[0032]

[0033] Among them, λ≥0 is a tuning parameter used to customize 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 blocks, where T represents the reconstruction coefficient matrix. C is the key image area after the objective function is processed. At the same time, the additional terms introduce a specific matrix norm for regularization, aiming to achieve high reconstruction accuracy, especially for the most important K salient object blocks. This strategy significantly prioritizes these blocks while reducing attention to the less important NK blocks. For example, if all elements in a row of T are zero, it means that the corresponding block is considered non-essential and does not contribute to the reconstruction of other blocks. Therefore, T is designed to exhibit row sparsity, and each row indicates the reconstruction importance of a specific block.

[0034] In order to enhance the sparsity of the matrix T, the norm ||T|| m Adjust to ||T|| 2,1 , considering that the target region is located in a subspace in a higher-dimensional feature space, T needs to be a low-rank matrix. The subsequent objective function is optimized as:

[0035]

[0036] 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 collective configuration of the input data.

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

[0038] Furthermore, step S4 uses a support vector machine (SVM) to effectively classify image regions. For each L optimized image region, 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, thereby creating a comprehensive representation of each sports scene. This approach captures the complexity of the game, including player movement, ball trajectory, and important actions on the field.

[0039] For P optimized image regions, each scene shows different motion actions, and the task of each binary SVM classifier is to distinguish motion actions by using the combined feature vector so that the model can achieve accurate classification results. The optimization objective of the support vector machine classifier can be expressed as follows:

[0040]

[0041] Among them, δ j represents the jth optimized image region, and the feature vector represents a specific sports action. 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. The variable β represents the coefficient of the hyperplane used to distinguish different sports action categories in the feature space, M kl Refers to the number of optimized image regions associated with the action pair k and l. It ensures that the support vector machine (SVM) classifier can effectively separate each pair of sports actions, thereby optimizing the decision boundary for accurate classification.

[0042] Beneficial effects of the present invention:

[0043] The present invention provides a complex motion perception method based on human vision inspiration. First, the BING object measurement method is used to extract regional features of images, and two indicators of distinguishing ability and time cost are designed. This can ensure that the extracted key areas have visual significance while ensuring better computational efficiency. The self-representation model is optimized and low-rank active learning technology is used to enable the method to have stronger feature selection and distinguishing capabilities, which significantly improves the performance in complex motion environments. A support vector machine is used to classify the optimized areas, with high recognition accuracy, and excellent accuracy in identifying key motion actions and strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the working flow chart of the present invention. DETAILED DESCRIPTION

[0045] Research in visual cognition and psychology has shown that human eyes naturally tend to focus on areas of significant semantic or visual importance in architectural scenes. This observation suggests that our visual attention is intentionally focused on specific key elements rather than randomly dispersed. Based on this direction, the present invention develops a complex motion perception method inspired by human vision, combining the detection of object center patches with the application of low-rank active learning (LAL). The construction of LAL assumes that spatially adjacent scene patches are closely distributed in the feature space because of their similar visual appearance and semantics. In theory, the strategy of the present invention 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 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 striking elements, the present invention uses the BING objectness metric. This tool is very effective in quickly isolating high-quality object-related image regions (i.e., "landscape tiles"). The BING metric is particularly valued for its speed in quickly locating relevant tiles, its key role in creating gaze shift paths (GSPs) (by providing a wide range of object-specific tiles), and its strong generalization capabilities across multiple object categories. This flexibility significantly improves the versatility of our scene classification framework across diverse datasets. Once these BING tiles are isolated, they will be characterized in detail through high-quality feature extraction, which will be elaborated in subsequent sections.

[0047] The complex motion perception method based on human vision inspiration of the present invention is further described below with reference to the accompanying drawings.

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

[0049] In order to improve the performance and efficiency of semantic recognition, the present invention introduces a new feature selection method, called feature selection based on temporal discrimination (TD) and temporal correlation (TC). The technology is divided into two stages: the first stage removes features that lack sufficient distinguishing ability through the temporal discrimination (TD) indicator; the second stage removes features that show high redundancy through the temporal correlation (TC) indicator. 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. The details are as follows.

[0050] The BING objectness metric is used, which is very effective at quickly isolating high-quality image patches (called "image regions") associated with objects. Once these BING features are extracted, they are characterized in detail through high-quality feature extraction.

[0051] The present invention comprehensively evaluates each image region through three main performance indicators: processing speed, discrimination ability, and the 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 for any selected feature, denoted as V, covers the following two stages

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

[0053] Where T f Indicates the time required for feature extraction, Td Indicates the time required for classification

[0054] The ability of a feature to distinguish different types of image regions at the image block level is called its discriminative validity, which is evaluated by its correlation with the category label. This concept is very important in tasks such as image classification and object detection, especially in feature selection. In the present invention, the concept of feature margin is used to quantify the strength of the association between a specific feature and a category label. It can be quantified by evaluating the discriminability provided by the feature for different category labels in the classification task. Simply put, the margin can reflect the role of the feature in distinguishing categories. The larger the feature margin, the greater the contribution of the feature to distinguishing categories. The margin of feature V is defined as follows:

[0055]

[0056] Where |E| represents the total number of samples in the data set. N(v (n) ) represents the nearest error sample, that is, the distance v (n) The closest sample but not belonging to the same category, I(v (n) ) represents the most recent correct sample, that is, (n) The closest example of the same class. The discriminative power of a feature is evaluated by calculating the distance between the closest incorrect example and the closest correct example. 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 power of individual features, it is also crucial to explore the interactions between features. The interactions between features may significantly affect the predictive power of the model, especially in complex nonlinear tasks. By identifying and modeling these interactions, we can help improve the performance of the model. When two highly discriminative features are highly correlated, this may indicate 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 often evaluated by information gain, which is determined based on the entropy of the correlated features. Information gain evaluates the importance of a feature by calculating the uncertainty it reduces (the reduction in entropy) when partitioning a dataset. The larger the information gain, the greater the role of the feature in distinguishing categories. The entropy of feature V is defined as follows:

[0058]

[0059] Where P(v (n) ) indicates v (n) Possibility

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

[0061]

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

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

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

[0065]

[0066] In order to evaluate the effectiveness of a feature's discriminative power relative to its processing time, this paper introduces a metric called the ratio of discriminative power to time cost. For feature V, the ratio is calculated as follows:

[0067]

[0068] Where p∈(0,+∞), allows for adjustments based on the processing time required for each feature to accommodate more situations. A higher p value indicates a greater focus on the time efficiency of the feature, and the p value is adjusted to achieve the capabilities and efficiency of specific needs for different requirements.

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

[0070]

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

[0072] Visual features usually exhibit nonlinear distribution in their domain. In the field of computer vision and image processing, the nonlinear distribution of features means that their changes are not just simple linear relationships, but complex distribution characteristics that are highly dependent on the input data. Therefore, in order to better understand the interactions between them, 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 the analysis. This method is particularly effective in fields such as semantic recognition because it uses 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 the features in terms of computational requirements, as follows:

[0074]

[0075] Where V i and W i are the original features, which are transformed 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=[i 1 ,…,i N ] represents a matrix containing multiple visual features of N training blocks, and V is a matrix of self-reconstruction coefficients. In order 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, τ 1 and τ 2 are non-negative weights of the norm respectively.

[0081] The matrix M is introduced, which 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 in columns. If two samples i i and i j From the same category, their projection features Mi in low-dimensional space i and Mi jshould show similarity, thus giving a higher weight Mi to the connection between them j In addition, the matrix K is defined as |V|+|V T |, acts as a weight matrix, quantifying the relationship between all samples. i and i j Belong to different categories, then Mi i and Mi j The proximity between should be greater, resulting in a lower weight K ij These relationships and assumptions are summarized in the following objective function:

[0082]

[0083] where ⊙ represents the Hadamard product operation, n i represents the label of the i-th sample in each category; M is the dimension 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 that evaluates 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.

[0084] By integrating these concepts, the objective function of feature selection can be reconstructed to include both the self-representation model that maintains the geometric structure and the label consistency model, which can enable the model to better understand and maintain the consistency of the label while ensuring the geometric structure information of the data. This method is very useful when processing visual data, especially when there is less data or incomplete labels. This optimization method ensures that the selected features not only maintain the inherent geometric structure of the data, but also accurately correspond to the label information of the sample.

[0085]

[0086] In the present invention, ||V|| v and ||L|| w Application 1 norm to promote the sparsity of the representation matrix and the error matrix. In order to ensure the column sparsity of L, which is crucial for feature selection, we use l 1,2norm. Minimize a loss function through multiple regularization terms, combining sparsity, low-rank constraints and smoothness. The objective function can be changed to the following:

[0087]

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

[0089] Among them, τ 3 is an additional non-negative weight factor that processes the key image areas through the improved objective function.

[0090] (3) Using low-rank active learning (LAL) technology to determine image regions that can attract human attention and generate gaze transfer paths (GSPs)

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

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

[0093]

[0094] in, represents the reconstruction error, that is, the difference between the original data C and the data CT processed by the transformation matrix T. The goal is to transform the original image C into a more meaningful form through the transformation matrix T and minimize the error in the conversion process, through ||T|| m, imposes a regularization constraint on the transformation matrix T to control its sparsity. This regularization term encourages the elements of the matrix T to be sparse, so that the final transformation matrix T retains only the most meaningful regions or features, thereby identifying the important regions in the image. ,λ≥0 is a tuning parameter used to customize 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 blocks, where T represents the reconstruction coefficient matrix. C is the input sample matrix. At the same time, the additional term introduces a specific matrix norm for regularization, aiming to achieve high reconstruction accuracy, especially for the most important K salient object blocks. This strategy significantly prioritizes these blocks while reducing attention to the less important NK blocks. For example, if all elements in a row of T are zero, it means that the corresponding block is considered non-essential and does not contribute to the reconstruction of other blocks. Therefore, T is designed to exhibit row sparsity, and each row indicates the reconstruction importance of a specific block.

[0095] To enhance the sparsity of the matrix T, the norm ||T|| m Adjust to ||T|| 2,1 or ||T|| ∞ In the present invention, ||T|| is selected 2,1 The matrix T plays two roles in the framework of the present invention: (i) capturing the reconstruction parameters, where each column represents the combination of key object patches required to reconstruct the entire scene image; and (ii) serving as a self-representing matrix, where the matrix from the set T N Each column of s i As a feature vector . C is an implicit dictionary.

[0096] Considering that the object patch lies in a subspace in a higher-dimensional feature space, T needs to be a low-rank matrix. This requirement affects 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, which 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 of unsupervised learning algorithms. This reformulation enables efficient solution of the matrix S and helps in determining the K most representative frames to capture the essence of each object patch sequence.

[0101] The goal of this function is to accurately locate key object patches using low-rank representation. Typically, the reconstruction error is calculated using the input sample matrix C. However, samples (such as object regions) may be affected by noise or other distortions. In this case, directly reconstructing each sample may reduce the effectiveness of model training. To this end, this method uses 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 so 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 method is that even if the data itself is affected by noise or distortion, it can still be well represented by suitable dictionary elements, thereby reducing the impact of noise. This method combines subspace learning with low-rank constraints to enhance the robustness of the model to noise. Therefore, the strategy of this method not only aims to maintain the low-rank structure, but also optimize sample selection based on this structure. This leads to a robust objective function that is resilient to noise and improves the efficiency of sample selection, ultimately resulting in a refined objective function that excels in sample selection while maintaining a low-rank framework.

[0102]

[0103] Among them, μ ≥ 0 is used to adjust the emphasis of sample reconstruction, and μ = 1 is set to ensure the best performance. The matrix U is the reconstruction dictionary

[0104] (4) Use support vector machine (SVM) to effectively classify image regions

[0105] Using support vector machines (SVM) to efficiently classify image regions, in MobileEye's development of complex motion perception related to ball sports, this method focuses on extracting key visual features from each scene to achieve 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 showing a different sports action, this method constructs a binary classifier The task of each binary SVM classifier is to distinguish motion actions by using the combined feature vector so that the model can achieve accurate classification results. The optimization goal of the support vector machine classifier can be expressed as follows:

[0107]

[0108] Among them, δ j The feature vector representing the jth training scene represents a specific sports action. 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. The variable β represents the coefficient of the hyperplane used to distinguish different sports action categories in the feature space, M kl refers to the number of training scenes associated with action pair j and l. This ensures that the support vector machine (SVM) classifier effectively separates each pair of sports actions, thereby optimizing the decision boundary for accurate classification. This structure enables MobileEye to accurately identify and classify different ball-related sports actions, such as passing, scoring or tackling, providing real-time, actionable insights and enhancing the educational significance of sports analysis.

[0109] Compared with the existing technology,

[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 areas in each motion scene. This not only improves computational efficiency, but also improves accuracy, especially in fast-moving motion scenes, where fast and accurate recognition 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 selected features. These algorithms adjust the weights of feature channels according to their importance, significantly improving performance in complex motion environments where accuracy and resource management are critical. 3. Highly competitive recognition accuracy: The present invention has high accuracy and demonstrates excellent accuracy in identifying key motion actions and strategies.

[0111] To validate the capabilities of our approach, a simple experiment was designed. The discriminative power of 128L deep gaze transfer path (GSP) features was evaluated within a multi-class support vector machine (SVM) framework, following the structure outlined in previous research. The evaluation focused on the ability of MobileEye to accurately classify ball-related sports actions, such as passing, dribbling, and shooting, leveraging visual features inspired by human visual dynamics. The present invention was then compared with seven state-of-the-art deep learning models that excel in handling sports-specific data and action recognition across different sports scenarios. The evaluation used a comprehensive dataset from leading research 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 scene recognition. The results, shown in Table 1, highlight the ability of the present method to accurately and reliably classify ball sports actions, enhancing its applicability in educational platforms and real-time sports analysis. To ensure the reliability of the results, each test was performed 20 times and the standard deviation was recorded to assess the consistency of the results. The evaluation results consistently show that the present method outperforms 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 present 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 summarize the following two points: (1) the method consistently provides competitive accuracy in each sports category while maintaining a significantly lower standard error; (2) the minimum standard deviation associated with the method emphasizes its robustness, showing that it has more stable performance than other models in repeated experiments.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to 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 in that: The following steps are involved: S1. Capture images of athletes during sports and identify key image regions by evaluating the semantic and visual importance of image regions; S2. Dynamically process regional features of key image areas; S3. For the regional features after dynamic processing, low-rank active learning technology is used to optimize the objective function to obtain the optimized image region; 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 evaluating the semantic and visual importance of the image region is as follows: The marginal value of a feature is used to quantify the strength of the association between the feature and the class label. The marginal value of 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 closest 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 the information gain as follows: G(V|W)=EN(V)-EN(V|W) Among them, EN(V) is the entropy of feature V; EN(V|W) is the conditional entropy of feature V, expressed under the observation of given 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 follows: 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 characteristic 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.

3. The complex motion perception method based on human vision inspiration according to claim 2 is characterized in that: The specific implementation process of step S2 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 of 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, and the refined self-representation model is obtained 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 the i-th row and j-th column 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 dimension reduction matrix, and 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 that evaluates 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 as follows: Among them, τ3 is an additional non-negative weight factor; the key image area is processed by the improved objective function.

4. The complex motion perception method based on human vision inspiration according to claim 3 is 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.

5. The complex motion perception method based on human vision inspiration according to claim 4 is characterized in that: The specific implementation process of step S4 is as follows: For the optimized image regions, deep features are extracted from each region using a convolutional neural network trained on ImageNet; 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 Make sure that the support vector machine (SVM) classifier can effectively separate each pair of sports actions.

Images