A motion recognition method, system, device and medium based on skeleton coordinate points

Abstract

The present invention discloses a method, system, device, and medium for action recognition based on skeletal coordinate points. The method comprises the following steps: obtaining a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence; and performing action recognition based on the obtained video sequence of skeletal coordinate points using a pre-trained action recognition model to obtain an action recognition classification result. The technical solution provided by the present invention is specifically a small-sample skeletal action recognition method based on parallel spatiotemporal interactive alignment, which can effectively compare the similarity between the query sample and each type of support sample and accurately classify small-sample actions.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and in particular relates to a motion recognition method, system, device and medium based on skeletal coordinate points. Background Art

[0002] In small sample scenarios, skeleton-based action recognition is a task that uses a small amount of skeleton actions to train a model to recognize human actions. This task usually uses a small amount of skeleton action data representing human actions to train the model, and then tests the generalization of the model on a new action category dataset. This task has many practical application scenarios, such as medical action video analysis, high-speed motion scene recognition, etc.

[0003] Currently, existing skeletal action recognition methods with small sample sizes still have the following problems:

[0004] (1) Existing models do not consider the temporal frame misalignment of the skeleton sequence and the spatial position differences of the joint points. These spatial and temporal information are very important for comparing small samples, which makes it difficult to distinguish when matching similar query and support actions;

[0005] (2) The human skeleton has a topological structure with strong semantics, and each joint has different physical meanings in the action; however, existing methods do not make reasonable use of topological information, and obtain non-discriminative semantic features, resulting in low small sample classification accuracy. Summary of the Invention

[0006] The present invention aims to provide a method, system, device, and medium for motion recognition based on skeletal coordinate points to address one or more of the above-mentioned technical problems. The technical solution provided by the present invention is a small-sample skeletal motion recognition method based on parallel spatiotemporal interactive alignment, which can effectively compare the similarity between a query sample and each type of supporting sample and accurately classify small-sample motions.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A first aspect of the present invention provides a method for motion recognition based on skeletal coordinate points, comprising the following steps:

[0009] Obtaining a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence;

[0010] Based on the obtained video sequence of skeletal coordinate points, performing action recognition using a pre-trained action recognition model to obtain an action recognition classification result;

[0011] The action recognition model includes:

[0012] The feature extraction module is used to input the video sequence of skeleton coordinate points for feature extraction and obtain joint point level features;

[0013] The topological encoding module is used to input joint-level features for feature extraction, obtaining body local features and semantic association features;

[0014] An attention interaction network module includes a parallel spatial interaction-based alignment network branch and a temporal interaction-based alignment network branch; wherein the spatial interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a spatial category distance score from the query sequence to the support sequence; the temporal interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a temporal category distance score from the query sequence to the support sequence; the attention interaction network module is used to obtain a weighted sum category distance score based on the spatial category distance score and the temporal category distance score;

[0015] The nearest neighbor classifier is used to input weighted and category distance scores and perform classification, and output action recognition classification results.

[0016] A further improvement of the method of the present invention is that, in the feature extraction module, the step of inputting a video sequence of skeletal coordinate points for feature extraction and obtaining joint point level features includes:

[0017] Convert the bone coordinate points into three sets of information: joint point position, bone length vector, and bone velocity vector;

[0018] A graph convolutional encoder based on front fusion is used to encode three sets of information, namely joint position, bone length vector and bone velocity vector, in parallel to generate query joint level features and support joint level features.

[0019] A further improvement of the method of the present invention is that, in the topological coding module, the step of inputting joint-level features for feature extraction and obtaining body local features and semantic association features includes:

[0020] According to the human skeletal structure, the human body structure is divided into five body parts: left arm, right arm, left leg, right leg and torso; the part-node adjacency matrix is ​​constructed based on prior knowledge V represents the number of joints, and P represents the number of locals; each column of the local-node adjacency matrix is ​​composed of one-hot vector encodings represented by 0 and 1, which are used to represent that the node belongs to a specific local of the human skeleton topology;

[0021] Use mutual attention mechanism to learn joint-level features F and local node features The semantic interaction between them makes the joints contain rich body local and overall movement patterns, which can be expressed as:

[0022]

[0023] Where, is the scaling factor, is the adjacency matrix normalized using the inverse degree matrix, W is the linear mapping weight, F part is the local feature of the body, Softmax() represents the normalized exponential function;

[0024] In spatial interaction, semantic information is given to each joint point to obtain semantic association features; a set of trainable parameters is assigned to each joint point pair as a bias item to form a semantic bias adjacency matrix As a semantic association feature, it is used to represent the semantic association between V×V pairs of nodes.

[0025] A further improvement of the method of the present invention is that, in the attention interaction network module, the step of inputting joint-level features, body-part features, and semantic association features into the alignment network branch based on spatial interaction and outputting the spatial category distance score between the query sequence and the support sequence includes:

[0026] Bidirectional spatial mutual attention is used to adjust the frame-by-frame node responses between the query joint-level features and the support joint-level features to obtain query and support skeleton-level features. The obtained query and support skeleton-level features are divided into a set of all subsequence features in time series. The combined query and support time series features are obtained by traversing all subsequence combinations in time series. The similarity between the query and support time series features is measured using the one-way average symmetric surface criterion, and the cumulative sum of all similarity scores of the query time series features in time series is used as the spatial category distance score from the query sequence to the support sequence.

[0027] in,

[0028] The expression for querying bone-level features is:

[0029]

[0030]

[0031] Where, F q←s represents the aligned query skeleton-level features; GAP represents the global average pooling in the spatial dimension; FFN represents the feedforward neural network; A q←s represents the weight matrix of the support pair query; represents the linear mapping weight; To query joint-level features, T represents the time length, V represents the number of joints, and C represents the number of channel dimensions; F is the support joint level feature; part It is a local feature of the body; is a semantic association feature;

[0032] The expression supporting the bone-level features is,

[0033]

[0034]

[0035] Where, represents the linear mapping weight, A s←q represents the weight matrix of the query pair support, Represents the supporting bone-level features after alignment;

[0036] The expression of the spatial category distance score from the query sequence to the support sequence is,

[0037]

[0038] Where D(q,c) is the spatial category distance score from the query sequence to the support sequence, c is the action category, and are the combined query and support time series features, i and j are the indexes of the query and support features in the time series dimension, and T′ is the length of the time series subsequence.

[0039] A further improvement of the method of the present invention is that, in the attention interaction network module, the step of inputting joint-level features, body-part features, and semantic association features into the alignment network branch based on temporal interaction and outputting a temporal category distance score from the query sequence to the support sequence comprises:

[0040] Joint semantic association features and body local features are used to adjust the joint response frame by frame within the query and support joint-level features to obtain query and support skeleton-level features. The query and support skeleton-level features are first interacted in the time dimension and then in the channel dimension to obtain the interacted skeleton-level features. The interacted skeleton-level features are divided into a set of all subsequence features in time series to obtain the temporal feature sequence of the query and support. The temporal mutual attention network is used to learn the correlation between the query and support feature sequences and calculate the similarity to obtain the temporal category distance score from the query sequence to the support sequence.

[0041] in,

[0042] The expressions for querying and supporting bone-level features are,

[0043] G=GAP(FFN(A g V g));

[0044]

[0045]

[0046] Where G is the query and support skeleton level features; GAP represents the global average pooling in the spatial dimension; FFN represents the feedforward neural network, which consists of three layers of point-by-point convolutional networks; A sm is the semantic association feature; C represents the number of channel dimensions, X part is the local feature of the body; F is the query joint level feature or support joint level features and is the learnable weight matrix;

[0047] The expression of the bone-level feature after interaction is:

[0048] Z=U T +σ(LN(U T )W C );

[0049] U=G T +σ(LN(G T )W T );

[0050] Where U is the feature after time interaction, σ represents the GELU nonlinear activation function, LN represents the layer normalization method, and W T and W C A hybrid encoder consisting of two multi-layer perceptrons;

[0051]

[0052]

[0053] Where, and are the learnable matrices, represents the temporal feature sequence of the query, represents the correlation matrix of query to supporting time series features, represents the supporting temporal features of the corresponding action category c, LN is the standard layer normalization method, represents the weighted supporting temporal features relative to a specific query, Represents the weighted query time series features;

[0054] The expression of the time category distance score from the query sequence to the support sequence is,

[0055] Where T(q,c) is the temporal category distance score from the query sequence to the support sequence, t is the index of the query and support features in the temporal dimension, and T′ is the length of the temporal subsequence.

[0056] A further improvement of the method of the present invention is that, in the attention interaction network module, in the step of obtaining the weighted sum category distance score based on the spatial category distance score and the temporal category distance score,

[0057] The final weighted sum category distance score obtained is T(q,c)+λD(q,c); where T(q,c) is the temporal category distance score from the query sequence to the support sequence, λ is the weighting coefficient, and D(q,c) is the spatial category distance score from the query sequence to the support sequence.

[0058] A further improvement of the method of the present invention is that the pre-trained action recognition model uses a standard cross entropy loss when training, which is expressed as:

[0059] L=L cs (-D(q,·),y)+λL ct (-T(q,·),y);

[0060] Where λ is the balanced spatial interaction alignment loss L cs and temporal interaction alignment loss L ct The commonly used weights are , and y is the true category label of the query sequence.

[0061] A second aspect of the present invention provides a motion recognition system based on skeletal coordinate points, comprising:

[0062] A data acquisition module is used to acquire a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence;

[0063] A result acquisition module is used to perform action recognition based on the acquired video sequence of skeletal coordinate points using a pre-trained action recognition model to obtain an action recognition classification result;

[0064] The action recognition model includes:

[0065] The feature extraction module is used to input the video sequence of skeleton coordinate points for feature extraction and obtain joint point level features;

[0066] The topological encoding module is used to input joint-level features for feature extraction, obtaining body local features and semantic association features;

[0067] An attention interaction network module includes a parallel spatial interaction-based alignment network branch and a temporal interaction-based alignment network branch; wherein the spatial interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a spatial category distance score from the query sequence to the support sequence; the temporal interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a temporal category distance score from the query sequence to the support sequence; the attention interaction network module is used to obtain a weighted sum category distance score based on the spatial category distance score and the temporal category distance score;

[0068] The nearest neighbor classifier is used to input weighted and category distance scores and perform classification, and output action recognition classification results.

[0069] A third aspect of the present invention provides an electronic device, comprising:

[0070] at least one processor; and,

[0071] a memory communicatively connected to the at least one processor; wherein,

[0072] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the action recognition method based on skeletal coordinate points as described in any one of the first aspects of the present invention.

[0073] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the motion recognition method based on skeletal coordinate points as described in any one of the first aspects of the present invention.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] In the action recognition method disclosed by the present invention, spatial and temporal alignment is performed through two parallel spatiotemporal interaction alignment branches, which improves the feature space learning when matching query set and support set samples from different angles; at the same time, by proposing spatial attention associated with local body topology and joint semantics, the discriminability of spatial semantic features is improved; the accuracy of the final small sample classification can be improved. In the present invention, the spatiotemporal graph convolutional network is first used to extract the basic skeletal joint feature sequence, and then the joint features are enriched through local division and semantic association in the spatial dimension. On this basis, a parallel spatiotemporal interaction alignment network is used to align the query and support features from the perspectives of spatial interaction and temporal interaction respectively; finally, query and support matching are performed respectively in the above two feature spaces that focus on temporal and spatial consistency. The two parallel networks complement each other and can improve the accuracy of the final small sample classification.

[0076] The method proposed in this invention addresses the difficulty of existing methods in comparing the similarity of skeletal motion sequences in spatial and temporal dimensions under small sample sizes. For the first time, it improves the similarity measurement scheme of skeletal motion sequences by aggregating the spatiotemporal context content of features within the sequence and associating the spatiotemporal states between sequences. It also focuses on feature learning of skeletal spatial semantics and utilizes the inherent local division and semantic association of the skeleton to enhance the specific correlation between joint points. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art; obviously, the drawings described below are some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0078] Figure 1 1 is a flow chart of a method for motion recognition based on skeletal coordinate points provided by an embodiment of the present invention;

[0079] Figure 2 1 is a schematic diagram of the training process of the model in an embodiment of the present invention;

[0080] Figure 3 is a schematic diagram of the network structure of the model in an embodiment of the present invention;

[0081] Figure 4 is a schematic diagram of the structure of spatial and temporal mutual attention in an embodiment of the present invention; wherein, Figure 4 (a) is a schematic diagram of the spatial mutual attention between query and support samples; Figure 4 (b) is a schematic diagram of the temporal mutual attention between the query sample and the supporting samples of category c;

[0082] Figure 5 Schematic diagram of the local-joint mutual attention sub-network structure in an example of the present invention;

[0083] Figure 6 is the heat map result between the subsequences of query action features and supporting action features improved by temporal mutual attention in the example of the present invention; wherein, Figure 6 (a) is a schematic diagram of the query action "cutting nails" and the supporting action "combing hair"; Figure 6 (b) is a schematic diagram of the query action "cut nails" and the supporting action "cut nails"; Figure 6 (c) is a schematic diagram of the query action "cutting nails" and the supporting action "standing";

[0084] Figure 7: is a schematic diagram of the visualization results of the t-SNE clustering effect of the basic features and adaptive features of the graph convolution in the example of the present invention; wherein, Figure 7 (a) is the t-SNE visualization of the basic features under 5 categories; Figure 7 (b) is the t-SNE visualization of the basic features under 10 categories; Figure 7 (c) is the t-SNE visualization of the adaptability features under five categories; Figure 7 (d) is the t-SNE visualization of the adaptability features under 10 categories;

[0085] Figure 8 Schematic diagram of a skeletal motion recognition system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0086] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0087] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0088] The present invention is described in further detail below with reference to the accompanying drawings:

[0089] See also Figure 1 The embodiment of the present invention provides a method for motion recognition based on skeleton coordinate points, comprising the following steps:

[0090] Step 1: Obtain a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence;

[0091] Step 2: Based on the video sequence of the skeleton coordinate points obtained in step 1, a pre-trained action recognition model is used to perform recognition to obtain an action recognition classification result;

[0092] The action recognition model includes:

[0093] A feature extraction module is used to input a video sequence of skeletal coordinate points for feature extraction, and obtain joint-level features of the input video sequence of skeletal coordinate points; wherein the feature extraction module can be an encoder constructed by a skeletal-based graph convolution model;

[0094] A topological encoding module is used to input the joint-level features for feature extraction, and obtain body local features and semantic association features corresponding to the input joint-level features;

[0095] A parallel attention interaction network module includes an alignment network branch based on spatial interaction and an alignment network branch based on temporal interaction; the alignment network branch based on spatial interaction is used to input joint-level features, body-part features, and semantic association features, and output a spatial category distance score from a query sequence to a support sequence; the alignment network branch based on temporal interaction is used to input joint-level features, body-part features, and semantic association features, and output a temporal category distance score from a query sequence to a support sequence; the parallel attention interaction network module is used to obtain a weighted sum category distance score based on the spatial category distance score and the temporal category distance score;

[0096] The nearest neighbor classifier is used to input weighted and category distance scores and perform classification, and output action recognition classification results.

[0097] See also Figures 2 to 4 In an embodiment of the present invention, the training step of the pre-trained action recognition model includes:

[0098] Step 1: Obtain a training sample set; each sample includes a video sequence of skeletal coordinate points and an action category label;

[0099] Specifically, step 1 is to divide the small sample data set and perform sample category and length sampling, including:

[0100] Step 1.1) Divide the dataset into training set, validation set, and test set according to the action categories;

[0101] Step 1.2) Sequences of different lengths are uniformly sampled into sequences of 30 frames in length.

[0102] Step 2: Use the skeleton-based graph convolution model to build an embedding network to obtain a feature extraction module. The processing steps include:

[0103] Step 2.1) Preprocessing of bone coordinate points;

[0104] Step 2.2) Generate joint-level features based on the front-fused graph convolutional encoder.

[0105] Step 3: Extract body parts and semantic association features, including:

[0106] Step 3.1) extracting body part related features;

[0107] Step 3.2) Extract semantic association features of joint points.

[0108] Step 4: Use attention to build a parallel spatiotemporal interaction alignment network, including:

[0109] Step 4.1) Construct query and support alignment network branches based on spatial interactions, including:

[0110] Step 4.1.1) Use bidirectional spatial mutual attention to adjust the frame-by-frame node responses between query and support features;

[0111] Step 4.1.2) Divide the skeleton-level representation into the set of all subsequence representations;

[0112] In step 4.1.3), the similarity between the query and support features is measured using the one-way average symmetric surface criterion (ASSM).

[0113] Step 4.2) Construct the query and support alignment network branches based on temporal interactions, including:

[0114] Step 4.2.1) Using the above joint semantic association features and body part association features, adjust the joint response frame by frame within the query and support;

[0115] Step 4.2.2) First, perform interaction on the skeleton-level features in the time dimension, and then in the channel dimension;

[0116] Step 4.2.3) Divide the skeleton-level representation into a set of all subsequence representations;

[0117] Step 4.2.4) Use the temporal mutual attention network to learn the correlation between the query and support feature sequences and calculate the similarity.

[0118] Step 5: Use the two outputs of the spatiotemporal interaction alignment network as category similarity scores to calculate the loss function, including:

[0119] Step 5.1) Use the negative distances obtained in steps 4.2.4) and 4.1.5) as the class scores, first use the softmax function to obtain the class probabilities, and then use the standard cross entropy loss;

[0120] 5.2) During inference, use the weighted sum of the above negative distances as the category score to classify the query into the category with the closest distance.

[0121] Step 6: Optimize network parameters to improve the accuracy of skeletal action recognition under small sample sizes, including:

[0122] Step 6.1) Iteratively optimize the neural network parameters based on the loss function obtained in step 5;

[0123] Step 6.2) Take the results of multiple random seeds as the average.

[0124] In summary, the embodiments of the present invention disclose a skeletal action recognition method based on parallel spatiotemporal interactive alignment under small sample conditions. In view of the problem that the skeletal sequence features generated by the existing methods ignore the semantics and temporal nature of the bones, a parallel spatiotemporal interactive alignment dual-branch network is introduced. Through mutual attention and self-attention, the correlation between the query set and the support set samples between sequences and within sequences is focused on, and the temporal and spatial misalignment is effectively solved. At the same time, attention based on local body topology and semantic association of joint points is proposed, which can more accurately capture the spatial position association.

[0125] See also Figures 1 to 5 , a skeleton action recognition method based on parallel spatiotemporal interactive alignment under a small sample size according to an embodiment of the present invention comprises the following specific steps:

[0126] Step 1: Divide the small sample data set and perform sample category and length sampling, including:

[0127] Step 1.1) Divide the dataset into training set, validation set, and test set according to the action category; each training (test) iteration randomly samples N samples from each category to form an input task And randomly divide the samples of each category into N q query samples and N s Support samples to form the query set and support set x and c represent sample and label information respectively. If the support set S randomly samples K samples from each of N categories, it is called an N-way-k-shot problem.

[0128] Step 1.2) Sequences of different lengths are uniformly sampled into sequences of 30 frames in length; each sample sequence is divided into 30 sub-segments of equal length using a uniform sampling method, and then one frame is randomly sampled from each sub-segment, so that each sample sequence is ultimately sampled for 30 frames; in this embodiment of the present invention, the skeleton graph consists of 25 joint points and 18 joint points, respectively.

[0129] Step 2: Use the skeleton-based graph convolutional model to build an embedding network, including:

[0130] Step 2.1) Preprocessing of skeletal coordinates: The skeletal coordinate input is converted into three sets of information: joint position, bone length vector, and bone velocity vector through a preprocessing network. These are input in parallel to provide multi-angle motion patterns. The preprocessing network includes subtraction between coordinates, subtraction between previous and next frames, and angle calculation between node coordinates.

[0131] Step 2.2) Construct a graph convolution encoder based on front fusion to generate joint level features; wherein, a graph convolution model based on skeleton input (STGCN) is constructed as the encoder Ψ E (Encoder), respectively encodes the node, length, and speed information of the three inputs in parallel; among them, there are L = 10 cascaded STGCN encoders to gradually generate the skeleton encoding output. Each STGCN encoder is composed of a spatial graph convolution module and a temporal convolution module cascaded, and when L = 3, the three parallel encoding outputs are pre-fused, and the fused node features are used as the input of the subsequent cascade module; finally, the query and support joint level features based on the graph convolution model are obtained. and Where T represents the time length, V represents the number of joint points, and C represents the number of channel dimensions.

[0132] Step 3: Extract body parts and semantic association features, including:

[0133] Step 3.1) Extract body part related features; divide the human body structure into P = 5 body parts according to the human skeletal structure, namely left arm, right arm, left leg, right leg, and torso; construct the part-node adjacency matrix based on prior knowledge Where V represents the number of joints, P represents the number of locals, and each column consists of one-hot vector encodings represented by 0 and 1, indicating that the node belongs to a specific local of the human skeleton topology;

[0134] Then use mutual attention to learn joint level features F and local node features The semantic interaction between them makes the joints contain rich body local and overall movement patterns, which can be expressed as:

[0135]

[0136] in, is the scaling factor, is the adjacency matrix normalized using the inverse degree matrix, W is the linear mapping weight, F part For local body features, such as Figure 5 Schematic diagram of the local-node mutual attention sub-network structure;

[0137] Step 3.2) Extract the semantic association features of the joint points; considering that each node has a unique physical meaning, semantic information is given to each node in the spatial interaction, that is, a set of trainable parameters is assigned to each node pair as a bias item to form a semantic bias adjacency matrix As a semantic association feature, it represents the semantic association between V×V pairs of nodes.

[0138] Step 4: Use attention to build a parallel spatiotemporal interaction alignment network, and build two parallel networks based on the first interaction dimension of query and support;

[0139] Step 4.1) Construct query and support alignment network branches based on spatial interactions, including:

[0140] Step 4.1.1) Use bidirectional spatial mutual attention to adjust the frame-by-frame node response between the query and support features; wherein, a symmetrical single-layer single-head mutual attention network encoding is used to consider the different strengths and positions of the associated nodes at the query joint level feature. and support joint level features Matrix multiplication is used to perform spatial mapping association between them; in addition, the semantic association feature A is introduced sm and body local features F part , which enhances the flexibility of matching between queries and supports:

[0141]

[0142]

[0143] in, represents the linear mapping weight, FFN represents the feedforward neural network, GAP represents the global average pooling in the spatial dimension, A q←s represents the weight matrix of the support pair query, F q←s Represents the aligned query bone-level features. Similarly, we obtain the bone-level features of the query and support sequences aligned in the spatial dimension. represents the linear mapping weight, A s←q represents the weight matrix of the query pair support, Represents the supporting bone-level features after alignment;

[0144]

[0145]

[0146] 4.1.2) Divide the skeleton-level features into a set of all subsequence features;

[0147] A complete action can be represented by a sequence of sub-actions at different positions and speeds, so the skeleton-level representation is decomposed into a set of all paired temporal frame representations. For example, is the representation of the i-th frame, the time series frame pair Represents a specific sequence semantic part, where 1≤i<j≤T.

[0148] By traversing all paired time frame combinations, the feature sequence tuple representation is obtained Where T' represents the number of frame pairs. Through the above binary frame pair division scheme, the combined and Where N represents the number of query samples and K represents the number of supporting samples in each action category.

[0149] Step 4.1.3) uses the one-way average symmetric surface criterion (ASSM) to measure the similarity between the query and the supporting temporal features; wherein, after adaptive adjustment of spatial mutual attention, the support features with good spatial alignment characteristics relative to the specific query are used, which reduces the measurement range and noise influence compared to using all supporting features. and supporting characterization To achieve the optimal overall matching similarity by performing detailed pairwise comparisons, we use a one-way average symmetric surface criterion to reflect temporal similarity, i.e., we use the maximum distance metric. We calculate the distance D(Q,c) by accumulating all similarity scores of the query set's frame pairs in temporal order. We then take the average of these scores in the temporal and sample dimensions as the similarity distance from the query feature to category c. This is expressed as:

[0150]

[0151] Where D(q,c) is the spatial category distance score from the query sequence to the support sequence, c is the action category, and are the combined query and support time series features, i and j are the indexes of query and support features in the time series dimension, and T′ is the length of the time series subsequence.

[0152] Step 4.2) Construct the query and support alignment network branches based on temporal interactions, including:

[0153] Step 4.2.1) Use the above joint semantic association features and body local features to adjust the joint response frame by frame within the query and support; Among them, a single-layer single-head self-attention encoder is used to consider the different strengths and positions of the associated nodes and convert the query joint level features into and support joint level features Matrix multiplication is used for spatial mapping association respectively. In addition, the semantic association feature A is introduced as above. sm and body part features X part , which enhances the flexibility of feature adjustment in spatial dimensions:

[0154]

[0155] The query and support joint level features F are passed through a learnable weight matrix The mapping obtains different projections Q g , K g 、V g , and then use the self-attention encoder to obtain the representation of the weighted interaction:

[0156]

[0157] G=GAP(FFN(A g V g )).

[0158] The FFN consists of a three-layer point-by-point convolutional network, and then uses spatial global average pooling (GAP) to obtain query and support skeleton level features G.

[0159] Step 4.2.2) The query and support skeleton level features G are first interacted in the time dimension and then in the channel dimension; wherein, a hybrid encoder is used to interact with G in the time dimension and the channel dimension in turn, and the hybrid encoder W T and W C They are composed of two multi-layer perceptrons. Here, a standard fully connected layer W is used for feature interaction. The first layer interacts with G in the time dimension, and performs temporal integration and interaction between local features with the same channel position along the time dimension:

[0160] U=G T +σ(LN(G T )W T ),

[0161] Where σ represents the GELU nonlinear activation function, LN represents the layer normalization method, and U is the feature after time interaction. Then the second layer of hybrid encoder is used to perform channel interaction on the feature U, that is, channel integration and interaction are performed between local features with the same time position along the channel dimension:

[0162] Z=U T +σ(LN(U T )W C ).

[0163] According to the above formula, we get the skeleton-level features Z after the query and support are interacted in the time and channel dimensions through the hybrid encoder. Step 4.2.3) divides the skeleton-level features into a set of all subsequence representations; among them, using the interactively enhanced feature map V, the subsequence binary frame pair partitioning scheme and the introduction of temporal position encoding (position encoding),

[0164] Get query and supporting time series feature sequences

[0165] Step 4.2.4) Use the temporal mutual attention network to learn the correlation between the query and the supporting feature sequence and calculate the similarity; the actions of the supporting feature sequence have different speeds and positions, and use the temporal features relative to the specific query The correlation of the sequence to selectively adjust the supporting time series feature sequence The mutual attention network consists of a single-head mutual attention layer to obtain the support features of category c weighted relative to a specific query:

[0166]

[0167]

[0168] Where, and are the learnable matrices, represents the temporal feature sequence of the query, represents the correlation matrix of query to supporting time series features, represents the supporting temporal features of the corresponding action category c, LN is the standard layer normalization method, represents the weighted supporting temporal features relative to a specific query, represents the weighted query time series feature; the expression of the time category distance score from the query sequence to the support sequence is, and then the distance between all subsequences of the query sequence and its specific weighted support time series feature sequence is calculated, and the average value in the time series dimension is taken as the similarity index:

[0169]

[0170] The distance metric uses the Euclidean distance function. According to the above formula, the time category distance score T(q,c) from the query sequence to the support sequence is obtained. t is the index of the query and support features in the time series dimension, and T′ is the length of the time series subsequence.

[0171] Step 5: Use the two outputs of the spatiotemporal interaction alignment network as category similarity scores to calculate the loss function, including:

[0172] Step 5.1) Use the negative distances obtained in steps 4.2.4) and 4.1.5) as the class scores, first use the Softmax function to obtain the class probabilities, and then use the standard cross entropy loss;

[0173] L=L cs (-D(q,·),y)+λL ct (-T(q,·),y),

[0174] Where λ is the balanced spatial interaction alignment loss L cs and temporal interaction alignment loss L ct The commonly used weights are , and y is the true category label of the query sequence.

[0175] Step 5.2) During inference, use the weighted sum of the above negative distances as the category score T(q,c)+λD(q,c) to classify the query sequence into the category represented by the closest support sequence.

[0176] Step 6: Optimize network parameters to improve the accuracy of skeletal action recognition under small sample sizes, including:

[0177] Step 6.1) Iteratively optimize the neural network parameters according to the loss function obtained in step 5; use the AdamW optimizer for 100 iterations, randomly sample 1000 scenarios (eposide) for training in each iteration, use an adaptive learning strategy, and weight decay 0.0005.

[0178] Step 6.2) When the predetermined number of iterations is reached, steps 4 and 5 are used to implement small sample skeletal action recognition, and the average of multiple random seeds is taken as the final result in each experiment.

[0179] In summary, the embodiment of the present invention proposes an effective parallel spatiotemporal interaction alignment network for the task of skeletal action recognition under small sample conditions, which obtains key semantic features in time and space through the relative states between and within the query and support sequences. The spatial interaction alignment branch adjusts the necessary node dependencies between the query and support sequences, and implements the maximum temporal similarity matching between sequence frame pairs at the skeletal level. The temporal interaction alignment branch first enhances the distinguishability of node-level and frame-level features within the sequence, and utilizes the temporal mutual attention correlation based on frame pairs for temporal matching.

[0180] For fair comparison, the present invention also uses the mean of the three random seeds as the final experimental result. This method is qualitatively and quantitatively compared with existing methods on the NTU-T, NTU-S, and Kinetics public datasets to verify the effectiveness of the proposed method.

[0181] Table 1. Comparison of 5-way-1-shot accuracy (%) on NTU-S, NTU-T, and Kinetics datasets under different backbones

[0182]

[0183] Table 2. Comparison of 5-way-5-shot accuracy (%) on NTU-S, NTU-T, and Kinetics datasets under different backbones

[0184]

[0185] As can be seen from Table 1, under the 5-way-1-shot setting, the proposed method significantly exceeds the existing best performance in all datasets and the backbone used. It has a greater improvement on the NTU-S dataset with more complex action types, reflecting the better generalization of the method. Moreover, when using a graph convolutional network backbone with more parameters or a more advanced one (2s-AGCN, MS-G3D), it does not bring more performance improvement, which proves the potential of using a simple ST-GCN spatiotemporal graph convolutional network.

[0186] Table 2 shows that when using a 5-way-5-shot setting with more support samples, it also significantly exceeds the existing best performance. The improvement is even more obvious for the Kinetics dataset, proving that the method of the embodiment of the present invention better utilizes the information of multiple support samples.

[0187] See also Figure 6 and Figure 7 , Figure 6 and Figure 7 This is a visualization result analysis of an embodiment of the present invention; Figure 6 Heatmap results of temporal frame pair subsequence mutual attention matching for query and support action representations, from Figure 6 (b) shows that the association under the same temporal index is the most significant, reflecting the strongest semantic association. Moreover, due to the different speeds and positions of the actions, different combinations of temporal divisions also show certain associations. Figure 6 (a) and Figure 6 In (c), the activation responses of actions belonging to different classes are not obvious. Figure 7 Visualization of TSNE dimensionality reduction of features obtained before and after the spatiotemporal interaction alignment. Figure 7 (a) and (b) are the basic representations obtained by the graph convolution model. Figure 7(c) and (d) in the middle are the representations after temporal mutual attention alignment adaptation. It can be seen that the query and support features of the corresponding categories after improvement achieve accurate clustering effects.

[0188] In summary, the embodiment of the present invention proposes a new parallel spatiotemporal interactive alignment network, which successfully improves the similarity measurement of skeletal action sequences by aggregating the spatiotemporal context content of features within the sequence and associating the spatiotemporal states between sequences, while enhancing the specific correlation between joint points by utilizing the local division and semantic association features inherent in the skeleton; qualitative analysis and visualization result analysis fully demonstrate the effectiveness and superiority of the present invention, and realize accurate small-sample skeletal action recognition.

[0189] The following are device embodiments of the present invention, which can be used to perform the method embodiments of the present invention. For details not disclosed in the device embodiments, please refer to the method embodiments of the present invention.

[0190] See also Figure 8 , an embodiment of the present invention provides a motion recognition system based on skeleton coordinate points, comprising:

[0191] A data acquisition module is used to acquire a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence;

[0192] A result acquisition module is used to perform action recognition based on the acquired video sequence of skeletal coordinate points using a pre-trained action recognition model to obtain an action recognition classification result;

[0193] The action recognition model includes:

[0194] The feature extraction module is used to input the video sequence of skeleton coordinate points for feature extraction and obtain joint point level features;

[0195] The topological encoding module is used to input joint-level features for feature extraction, obtaining body local features and semantic association features;

[0196] An attention interaction network module includes a parallel spatial interaction-based alignment network branch and a temporal interaction-based alignment network branch; wherein the spatial interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a spatial category distance score from the query sequence to the support sequence; the temporal interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a temporal category distance score from the query sequence to the support sequence; the attention interaction network module is used to obtain a weighted sum category distance score based on the spatial category distance score and the temporal category distance score;

[0197] The nearest neighbor classifier is used to input weighted and category distance scores and perform classification, and output action recognition classification results.

[0198] An embodiment of the present invention discloses a small-sample skeletal action recognition system, which is based on an encoder of multiple commonly used spatiotemporal graph convolutional networks and an attention-based parallel spatiotemporal interaction alignment network; the encoder models joint-level features through spatiotemporal association, and then introduces the skeleton's intrinsic local division and semantic association features to enrich the joint information; the parallel network improves the similarity measurement feature space of skeletal action sequences through intra-sequence feature spatiotemporal context content aggregation and inter-sequence spatiotemporal state association, thereby improving the matching accuracy through the fusion of complementary parallel modules.

[0199] In another embodiment of the present invention, a computer device is provided, which includes a processor and a memory, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method flow or corresponding function; the processor described in the embodiment of the present invention can be used for the operation of the action recognition method based on skeletal coordinate points.

[0200] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device for storing programs and data. It is understandable that the computer-readable storage medium here can include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the motion recognition method based on skeletal coordinate points in the above embodiment.

[0201] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0202] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0203] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0204] 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 it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A motion recognition method based on skeleton coordinate points, characterized in that: The following steps are involved: Obtaining a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence; Based on the obtained video sequence of skeletal coordinate points, performing action recognition using a pre-trained action recognition model to obtain an action recognition classification result; The action recognition model includes: The feature extraction module is used to input the video sequence of skeleton coordinate points for feature extraction and obtain joint point level features; The topological encoding module is used to input joint-level features for feature extraction, obtaining body local features and semantic association features; An attention interaction network module includes a parallel spatial interaction-based alignment network branch and a temporal interaction-based alignment network branch; wherein the spatial interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a spatial category distance score from the query sequence to the support sequence; the temporal interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a temporal category distance score from the query sequence to the support sequence; the attention interaction network module is used to obtain a weighted sum category distance score based on the spatial category distance score and the temporal category distance score; The nearest neighbor classifier is used to input weighted and category distance scores and perform classification, and output action recognition classification results; In the topological coding module, the steps of inputting joint level features for feature extraction and obtaining body part features and semantic association features include: dividing the human body structure into five body parts, namely the left arm, right arm, left leg, right leg and torso, according to the human skeletal structure; constructing a part-node adjacency matrix based on prior knowledge; V represents the number of joints, and P represents the number of locals. Each column of the local-node adjacency matrix consists of a one-hot vector encoding represented by 0 and 1, which is used to represent that the node belongs to a specific local of the human skeleton topology. The joint level feature F and the local node feature are learned using the mutual attention mechanism. The semantic interaction between It is the adjacency matrix normalized by the inverse degree matrix; semantic information is given to each joint point in the spatial interaction to obtain semantic association features; among them, a set of trainable parameters is assigned to each joint point pair as a bias item to form a semantic bias adjacency matrix As a semantic association feature, it is used to represent the semantic association between V×V pairs of nodes.

2. The method for motion recognition based on skeletal coordinate points according to claim 1, characterized in that: In the feature extraction module, the skeleton coordinate point video sequence is input for feature extraction, and the steps of obtaining joint point level features include: Convert the bone coordinate points into three sets of information: joint point position, bone length vector, and bone velocity vector; A graph convolutional encoder based on front fusion is used to encode three sets of information, namely joint position, bone length vector and bone velocity vector, in parallel to generate query joint level features and support joint level features.

3. The method for motion recognition based on skeletal coordinate points according to claim 1, characterized in that: Use mutual attention mechanism to learn joint-level features F and local node features When the semantic interaction between them is made, the joints contain rich body local and overall movement patterns, expressed as, Where, F part is the local feature of the body, Softmax() represents the normalized exponential function, is the scaling factor, and W is the linear mapping weight.

4. The method for motion recognition based on skeletal coordinate points according to claim 1, characterized in that: In the attention interaction network module, the step of inputting joint-level features, body-part features, and semantic association features into the alignment network branch based on spatial interaction and outputting the spatial category distance score between the query sequence and the support sequence includes: Bidirectional spatial mutual attention is used to adjust the frame-by-frame node responses between the query joint-level features and the support joint-level features to obtain query and support skeleton-level features. The obtained query and support skeleton-level features are divided into a set of all subsequence features in time series. The combined query and support time series features are obtained by traversing all subsequence combinations in time series. The similarity between the query and support time series features is measured using the one-way average symmetric surface criterion, and the cumulative sum of all similarity scores of the query time series features in time series is used as the spatial category distance score from the query sequence to the support sequence. in, The expression for querying bone-level features is: Where, F q←s represents the aligned query skeleton-level features; GAP represents the global average pooling in the spatial dimension; FFN represents the feedforward neural network; A q←s represents the weight matrix of the support pair query; represents the linear mapping weight; To query joint-level features, T represents the time length, V represents the number of joints, and C represents the number of channel dimensions; F is the support joint level feature; part It is a local feature of the body; is a semantic association feature; The expression supporting the bone-level features is, Where, represents the linear mapping weight, A s←q represents the weight matrix of the query pair support, Represents the supporting bone-level features after alignment; The expression of the spatial category distance score from the query sequence to the support sequence is, Where D(q,c) is the spatial category distance score from the query sequence to the support sequence, c is the action category, and are the combined query and supporting time series features, i and j are the indexes of query and supporting features in the time series dimension, T ′ is the length of the time series subsequence.

5. The method for motion recognition based on skeletal coordinate points according to claim 1, characterized in that: In the attention interaction network module, the step of inputting joint-level features, body-part features, and semantic association features into the alignment network branch based on temporal interaction and outputting a temporal category distance score from the query sequence to the support sequence includes: Joint semantic association features and body local features are used to adjust the joint response frame by frame within the query and support joint-level features to obtain query and support skeleton-level features. The query and support skeleton-level features are first interacted in the time dimension and then in the channel dimension to obtain the interacted skeleton-level features. The interacted skeleton-level features are divided into a set of all subsequence features in time series to obtain the temporal feature sequence of the query and support. The temporal mutual attention network is used to learn the correlation between the query and support feature sequences and calculate the similarity to obtain the temporal category distance score from the query sequence to the support sequence. in, The expressions for querying and supporting bone-level features are, G=GAP(FFN(A g V g )); Where G is the query and support skeleton level features; GAP represents the global average pooling in the spatial dimension; FFN represents the feedforward neural network, which consists of three layers of point-by-point convolutional networks; A sm is the semantic association feature; C represents the number of channel dimensions, X part is the local feature of the body; F is the query joint level feature or support joint level features and is the learnable weight matrix; The expression of the bone-level feature after interaction is: Z=U T +σ(LN(U T )W C ): UG T +σ(LN(G). T )W T )4 Where U is the feature after time interaction, σ represents the GELU nonlinear activation function, LN represents the layer normalization method, and W T and W C A hybrid encoder consisting of two multi-layer perceptrons; Where, and are the learnable matrices, represents the temporal feature sequence of the query, represents the correlation matrix of query to supporting time series features, represents the supporting temporal features of the corresponding action category c, LN is the standard layer normalization method, represents the weighted supporting temporal features relative to a specific query, Represents the weighted query time series features; The expression of the time category distance score from the query sequence to the support sequence is, Where T(q,c) is the temporal category distance score from the query sequence to the support sequence, t is the index of the query and support features in the temporal dimension, and T′ is the length of the temporal subsequence.

6. The method for motion recognition based on skeletal coordinate points according to claim 1, characterized in that: In the attention interaction network module, in the step of obtaining a weighted sum category distance score based on the spatial category distance score and the temporal category distance score, The final weighted sum category distance score obtained is T(q,c)+λD(q,c); where T(q,c) is the temporal category distance score from the query sequence to the support sequence, λ is the weighting coefficient, and D(q,c) is the spatial category distance score from the query sequence to the support sequence.

7. The method for motion recognition based on skeletal coordinate points according to claim 1, characterized in that: The pre-trained action recognition model uses the standard cross entropy loss during training, which is expressed as: L=L cs (-D(q,·),y)+λL ct (-T(q,·),y); Where λ is the balanced spatial interaction alignment loss L cs and temporal interaction alignment loss L ct The commonly used weights are , and y is the true category label of the query sequence.

8. A motion recognition system based on skeleton coordinate points, characterized in that: include: A data acquisition module is used to acquire a video sequence of skeletal coordinate points for action recognition; wherein the video sequence of skeletal coordinate points includes a query sequence and a support sequence; A result acquisition module is used to perform action recognition based on the acquired video sequence of skeletal coordinate points using a pre-trained action recognition model to obtain an action recognition classification result; The action recognition model includes: The feature extraction module is used to input the video sequence of skeleton coordinate points for feature extraction and obtain joint point level features; The topological encoding module is used to input joint-level features for feature extraction, obtaining body local features and semantic association features; An attention interaction network module includes a parallel spatial interaction-based alignment network branch and a temporal interaction-based alignment network branch; wherein the spatial interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a spatial category distance score from the query sequence to the support sequence; the temporal interaction-based alignment network branch is used to input joint-level features, body-part features, and semantic association features, and output a temporal category distance score from the query sequence to the support sequence; the attention interaction network module is used to obtain a weighted sum category distance score based on the spatial category distance score and the temporal category distance score; The nearest neighbor classifier is used to input weighted and category distance scores and perform classification, and output action recognition classification results; In the topological coding module, the steps of inputting joint level features for feature extraction and obtaining body part features and semantic association features include: dividing the human body structure into five body parts, namely the left arm, right arm, left leg, right leg and torso, according to the human skeletal structure; constructing a part-node adjacency matrix based on prior knowledge; V represents the number of joints, and P represents the number of locals. Each column of the local-node adjacency matrix consists of a one-hot vector encoding represented by 0 and 1, which is used to represent that the node belongs to a specific local of the human skeleton topology. The joint level feature F and the local node feature are learned using the mutual attention mechanism. The semantic interaction between It is the adjacency matrix normalized by the inverse degree matrix; semantic information is given to each joint point in the spatial interaction to obtain semantic association features; among them, a set of trainable parameters is assigned to each joint point pair as a bias item to form a semantic bias adjacency matrix As a semantic association feature, it is used to represent the semantic association between V×V pairs of nodes.

9. An electronic device, characterized in that: include: at least one processor; as well as, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the motion recognition method based on skeleton coordinate points according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the motion recognition method based on skeletal coordinate points according to any one of claims 1 to 7 is implemented.

Images