Method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models

Abstract

Dynamic point cloud action recognition method based on static models, comprising: generating and preprocessing three-dimensional point cloud data; constructing a deep learning model associated with a

Description

Technical Field The present invention belongs to the field of computer vision and pattern recognition, and specifically relates to a method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models. Background Art In the fields of computer vision and pattern recognition, dynamic point cloud action recognition is an important research direction, widely applied in areas such as robotics, video surveillance, autonomous driving, and virtual reality. Dynamic point cloud data can capture the dynamic changes of objects or scenes in the real world, making effective processing and recognition of such data highly significant. Methods Based on4D Convolution: Many existing methods use complex 4D convolution to process dynamic point cloud data. These methods capture the dynamics of 3D spatial points in the temporal channel through space-time convolution. Although they achieve high performance, they come with significant computational overhead and complex models, making them difficult to apply in scenarios that require real-time processing. Dual-Stream Structure Methods: Another class of methods uses a dual-stream structure to decouple spatial and temporal features for processing. For example, the Kinet method captures 3D actions through scene flow and proposes a jointly trained dual-stream network that integrates static spatial features with dynamic temporal features. While this approach improves flexibility to some extent, concatenating all frames into a single point cloud results in the loss of temporal information. Despite the achievements of existing 4D convolution and dual-stream structure methods in action recognition, they have the following limitations: 1. High Computational Complexity: 4D convolution methods have large computational requirements and many parameters, making them difficult to apply in real-time processing scenarios. 2. Loss of Temporal Information: Dual-stream structure methods lose temporal information when all frames are concatenated into a single point cloud, requiring an additional dynamic stream to compensate for this deficiency. In summary, existing dynamic point cloud action recognition technologies need further improvement. Summary of the Invention The present invention addresses the current state of the art and overcomes the above-mentioned deficiencies by providing a method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models. The primary objective of the method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models disclosed in the present invention is to enhance temporal modeling through implicit and explicit temporal dependencies, enabling efficient processing of dynamic point cloud data. This can be widely applied in household robots, virtual reality, augmented reality, and other scenarios requiring efficient dynamic point cloud processing. Another objective of the method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models disclosed in the present invention is to specifically process and analyze human actions in three-dimensional point cloud data. Another objective of the method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models disclosed in the present invention is to introduce a novel temporal rolling operation that enhances implicit temporal dependency through weight sharing in batch processing and explicitly assimilates features of adjacent frames through temporal rolling operations. Another objective of the method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models disclosed in the present invention is to propose two different padding methods to fill empty frames at the end of sequences, maintaining the continuity and integrity of temporal information. Another objective of the method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models disclosed in the present invention is to accurately and efficiently extract information about human actions from high-dimensional point cloud data and to process and interpret temporal changes in dynamic scenes. The method for dynamic point cloud action recognition based on static models disclosed in the present invention comprises the following steps: Step SI: Generating and preprocessing three-dimensional point cloud data; Step S2: Constructing a deep learning model associated with a deep learning network; Step S3: Training the TimeNet framework, which includes a PointNet-like encoder. As a further preferred technical solution of the above technical scheme, Step SI is specifically implemented as the following steps: Step Sl.l: Acquiring a depth image sequence comprising T frames, where T is the length of the time series data; Step SI.2: Performing normalization processing on each frame of the depth image, wherein Step S1.2 includes Steps Sl.2.1, Sl.2.2, Sl.2.3, and Sl.2.4: Step Sl.2.1: Acquiring the length xx and width yy of the depth image, which represent the number of pixels in the length and width directions, respectively, and obtaining the corresponding empty grid matrix; Step SI.2.2: Defining the depth mapping value focal, with focal = 280; Step SI.2.3: Selecting the pixel positions with depth values d >0 in each empty grid matrix as the initial spatial values; Step SI.2.4: Calculating the spatial coordinates of the three-dimensional point cloud data. As a further preferred technical solution of the above technical scheme, Step SI is specifically implemented as the following steps: Step Sl.l: Acquiring a point cloud video comprising Tframes of depth images, where T is the length of the time series data; Step SI.2: Performing normalization processing on each frame of the depth image; Step SI.3: Classifying the dataset into training set, validation set, and test set; Step S1.4: Each training set includes multiple point cloud videos, each point cloud video containing T frames of three-dimensional point cloud data composed of N (x, y, z) spatial points; Step SI.5: For each T*N*3-dimensional data, setting a sampling step length step, an initial frame t, and a frame length c, with step = 1 and c = 24, where t <(T — c); Step SI.6: Segmenting each point cloud video into multiple overlapping video clips, each video clip having a length of Qframes; Step SI.7: Determining the shape of the video clip as 24*N*3; Step SI.8: Calculating the geometric center value (centroid) of each point cloud pc, calculating the maximum 2-norm (m) of each point cloud frame deviating from the center value, subtracting the geometric center value (centroid) from the three-dimensional point cloud data, and dividing the obtained value by the norm (m) to obtain the intermediate value (X); Step SI.9: Setting a random scaling value (scale) between [0.9,1.1], performing the corresponding random scaling on each intermediate value (X) to obtain the initial output data. As a further preferred technical solution of the above technical scheme, the PointNet-like encoder in Step S3 performs the following steps: Input and Preprocessing Steps: - Input Data Structure: Given a point cloud containing N points, represented as P = {pl, p2,..., pN}, where each point pi G RA3 is specified by geometric coordinates {x, y, z}; - Video Frame Representation: The point cloud video contains T (here, T = 24) frames of point clouds, represented as V = {Pl, P2,..., PT). Extending the Spatial Encoder to the Temporal Domain Steps: - Spatial Feature Capture Based on PointNet++: For static point cloud P, using a multi-layer perceptron for spatial set abstraction, expressed as Feature = f(P', P), where f represents standard PointNet++ layers, and P' is the downsampled query points generated from the original P; - Reshaping for Batch Processing: Reshaping the point cloud video V through batch processing, expanding the formula to Feature = f(V, V), where V and V respectively represent sets of support points and query points; - Rolling Operation: Applying a rolling operation in the time dimension to aggregate information between consecutive frames, effectively capturing dynamic changes in the time series: - Frame Misalignment Through Rolling Operation: Adding a rolling operation in the time dimension so that support points and query points come from different point cloud frames, thereby aggregating temporal features from Pt to Pt+1; - Temporal Feature Aggregation Formula: Expressed as Feature = f(Pt, Pt+1'), where Pt are the support points, and Pt+1' are the downsampled query points from the next point cloud frame. By performing operations on adjacent frames, the spatial set abstraction is extended to the temporal domain. The present invention discloses a dynamic point cloud action recognition system based on static models, executing the steps of any of the technical solutions of the dynamic point cloud action recognition method based on static models. The present invention also discloses an electronic device, comprising: a processor and a memory, wherein the memory stores executable instructions for the processor; the processor is configured to execute the steps of any of the technical solutions of the dynamic point cloud action recognition method based on static models by executing the executable instructions. The present invention further discloses a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of any of the technical solutions of the dynamic point cloud action recognition method based on static models. The beneficial effects of the dynamic point cloud action recognition method, system, electronic device, and storage medium based on static models disclosed by the present invention are as follows: 1. Implicit Temporal Awareness: Sequence data is treated as batch processing, utilizing the shared weight paradigm in batch training to achieve implicit temporal awareness. 2. Explicit Temporal Assimilation: Time information is further enhanced for explicit assimilation through plug-and-play temporal rooling operations and prediction-based padding modules. 3. Superior Performance: Utilizing only 0.17M parameters and 0.31G FLOPs, significantly outperforming existing methods, demonstrating superior performance in terms of model parameters, computational complexity, and data storage requirements. Description of Drawings Figure 1 is a schematic diagram of the basic concept of TimeNet and a comparison diagram of dynamic point cloud processing in static models with shared weight training. Figure 2 is a schematic diagram of the temporal rooling operation and its application in PointNet++. Detailed Description of Embodiments The present invention discloses a method, system, electronic device, and storage medium for dynamic point cloud action recognition based on static models. The specific embodiments are further described below in conjunction with the preferred embodiment (Embodiment 1) and with reference to Figures 1-2. Embodiment 1 Referring to Figures 1-2. Figure 1 shows the basic concept of TimeNet and the comparison of dynamic point cloud processing in static models with shared weight training. Figure 2 shows the temporal rooling operation and its application in PointNet++. Preferably, the method for dynamic point cloud action recognition based on static models includes the following steps: Step SI: Generating and preprocessing three-dimensional point cloud data; Step S2: Constructing a deep learning model associated with a deep learning network; Step S3: Training the TimeNet framework, which includes a PointNet-like encoder. Wherein, Step SI is specifically implemented as the following steps: Step Sl.l: Acquiring a depth image sequence comprising T frames, where T is the length of the time series data; Step SI.2: Performing normalization processing on each frame of the depth image, wherein Step SI.2 includes Steps SI.2.1, SI.2.2, SI.2.3, and SI.2.4: Step SI.2.1: Acquiring the length xx and width yy of the depth image, which represent the number of pixels in the length and width directions, respectively, and obtaining the corresponding empty grid matrix; Step SI.2.2: Defining the depth mapping value focal, with focal = 280; Step SI.2.3: Selecting the pixel positions with depth values d >0 in each empty grid matrix as the initial spatial values; Step SI.2.4: Calculating the spatial coordinates of the three-dimensional point cloud data. Wherein, Step SI is specifically implemented as the following steps: Step Sl.l: Acquiring a point cloud video comprising Tframes of depth images, where T is the length of the time series data; Step SI.2: Performing normalization processing on each frame of the depth image; Step SI.3: Classifying the dataset into training set, validation set, and test set; Step S1.4: Each training set includes multiple point cloud videos, each point cloud video containing T frames of three-dimensional point cloud data composed of N (x, y, z) spatial points; Step SI.5: For each T*N*3-dimensional data, setting a sampling step length step, an initial frame t, and a frame length c, with step = 1 and c = 24, where t <(T - c); Step SI.6: Segmenting each point cloud video into multiple overlapping video clips, each video clip having a length of Q.frames; Step SI.7: Determining the shape of the video clip as 24*N*3; Step SI.8: Calculating the geometric center value (centroid) of each point cloud pc, calculating the maximum 2-norm (m) of each point cloud frame deviating from the center value, subtracting the geometric center value (centroid) from the three-dimensional point cloud data, and dividing the obtained value by the norm (m) to obtain the intermediate value (X); Step SI.9: Setting a random scaling value (scale) between [0.9,1.1], performing the corresponding random scaling on each intermediate value (X) to obtain the initial output data. Wherein, the PointNet-like encoder in Step S3 performs the following steps: Input and Preprocessing Steps: Input Data Structure: Given a point cloud containing N points, represented as P = {pl, p2,..., pN}, where each point pi e RA3 is specified by geometric coordinates {x, y, z}; Video Frame Representation:*The point cloud video contains T (here, T = 24) frames of point clouds, represented as V = {Pl, P2,..., PT}; Extending the Spatial Encoder to the Temporal Domain Steps: Spatial Feature Capture Based on PointNet++: For static point cloud P, using a multi-layer perceptron for spatial set abstraction, expressed as Feature = f(P', P), where f represents standard PointNet++ layers, and P' is the downsampled query points generated from the original P; Reshaping for Batch Processing: Reshaping the point cloud video V through batch processing, expanding the formula to Feature = f(V', V), where V and V respectively represent sets of support points and query points; Rolling Operation: Applying a rolling operation in the time dimension to aggregate information between consecutive frames, effectively capturing dynamic changes in the time series: Frame Misalignment Through Rolling Operation: Adding a rolling operation in the time dimension so that support points and query points come from different point cloud frames, thereby aggregating temporal features from Pt to Pt+1; Temporal Feature Aggregation Formula: Expressed as Feature = f(Pt, Pt+1'), where Pt are the support points, and Pt+1' are the downsampled query points from the next point cloud frame; By performing operations on adjacent frames, the spatial set abstraction is extended to the temporal domain. Preferably, this embodiment discloses a dynamic point cloud action recognition system based on static models, executing the steps of any of the technical solutions of the dynamic point cloud action recognition method based on static models. Preferably, this embodiment discloses an electronic device, comprising: a processor and a memory, wherein the memory stores executable instructions for the processor; the processor is configured to execute the steps of any of the technical solutions of the dynamic point cloud action recognition method based on static models by executing the executable instructions. Preferably, this embodiment discloses a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the steps of any of the technical solutions of the dynamic point cloud action recognition method based on static models. The working principle of the dynamic point cloud action recognition method, system, electronic device, and storage medium based on static models disclosed in this embodiment is as follows. 1. Implicit Temporal Awareness: Sequence data is treated as batch processing, utilizing the shared weight paradigm in batch training to achieve implicit temporal awareness. By reshaping the input tensor from {B, T, N, C} to {BxT, N, C}, implicit temporal awareness is maintained during batch processing. This approach allows temporal information to be implicitly incorporated during the training phase without adding extra computational burden. 2. Explicit Temporal Assimilation: Temporal information assimilation is further enhanced through plug-and-play temporal rooling operations and prediction-based padding modules. 3. Temporal Rolling Operation: To further explicitly enhance temporal information, a temporal rolling operation is introduced, which rolls the dynamic point cloud sequence along the time channel, enabling interaction between features of adjacent frames. 4. TimeNet Framework: The TimeNet framework proposed in this invention enhances the temporal modeling capability of static 3D models through lightweight operations. Combined with the above points 1, 2, and 3, the TimeNet framework is specifically described as follows. 4.1 Implicit Temporal Awareness: Referring to Figure 1, the dynamic point cloud sequence is treated as batch processing, enabling implicit temporal awareness through the shared weight paradigm in batch processing. This method avoids the issue of temporal information loss caused by concatenation operations in existing methods. For example, the input tensor is reshaped from {B, T, N, C} to {B x T, N, C}, where B represents the batch size, T represents the number of sequence frames, N represents the number of points, and C represents the feature channels. This adjustment allows implicit inclusion of temporal information during training. 4.2 Explicit Temporal Assimilation: Referring to Figure 2, a temporal rooling operation is proposed to enhance explicit temporal awareness. Specifically, SAM (Spatial Aggregation Module) aggregates spatial local features of static point cloud frames to selected center points through sampling and grouping operations. The temporal rooling operation aggregates the spatial local features of frame t to the next frame t+1, enabling interaction of adjacent frame features during network inference and enhancing explicit temporal information. 4.3 Padding Mechanism: The temporal rooling operation rolls the point cloud sequence forward by one frame each time, resulting in a vacancy in the last frame. To address this issue, two padding methods are proposed: 4.3.1 Neighbor Frame Padding: Simply padding with the adjacent frame, but this method can only focus on its own spatial information and may affect network performance; 4.3.2 Prediction Padding: Padding by inferring the next frame point cloud. Specific methods include calculating the average change of the sequence and using this change to infer the next frame, or using PointLSTM as a learnable component for inference. Prediction padding specifically includes the following methods. 4.3.2.1 Average Change Prediction: This method infers the next frame data by calculating the average change of the sequence, as follows: Calculating the average change between adjacent frames in the sequence: 6 = Mean({Qt}t=2T - {Qt}t=l T-l); Inferring the padding frame using the calculated average change: Padding Frame = QT + 6; This method is simple and direct but may be affected by noise points in the sequence. 4.3.2.2 Prediction Based on PointLSTM: To more accurately predict the next frame point cloud data, PointLSTM is used as a learnable component. The specific steps are as follows: Using an MLP-based encoder to convert the point cloud sequence into feature space: point features = Encoder({Qt}t=l T); Using PointLSTM to infer temporal sequence features: point features = PointLSTM(point features); Using a decoder to convert the results from feature space back to geometric space and using the Tanh function to limit the range of inter-frame changes: 6 = 0.1 x Tanh(Decoder(point features)); Inferring the padding frame using the calculated 6 value: Padding Frame = QT + 6; This method introduces a learnable component to more accurately capture changes in the temporal sequence, improving the accuracy of padding. Through these padding methods, the TimeNet framework effectively resolves the issue of vacant frames caused by the temporal rooling operation, maintaining the integrity of temporal information while improving the precision and efficiency of dynamic point cloud data processing. The following describes the pseudocode example for the dynamic point cloud action recognition method, system, electronic device, and storage medium based on static models in this embodiment. Example 1: PyTorch implementation code example for the PointNet++ layer: def StaticPointNet2Layer(input): support_points = query_points = input query_points = Sampling(query_points) local_regions = Grouping(query_points, support_points) features = PointNetLayer(local_regions) features = MaxPooling(features) return features Require: p<stes ity«t iss te shape (8, A?, C) def Static >WntNet2 teyerfisspsi): p<iiw "■ <x«y pstelx « input x wnsphng •*> (8, M. C) query pates * tempi mgtqmy paixisi S steXpUSg - S' teC) local region® »: Oroupiugtesery points, support points) feateres ® FwintNei layers teal regiom) features - Max IMolisteltetete retest fealum Rc^tern pohiis- input in the shape (8 x T, &V''f def Dynamic PaiatNetS Layer(fepsi> satori insists;, qssry poisriS ® get layer iupwt(Input) » sssixgUng -> (>x T. M C| <ysi?ry points » &napBM»g(<tuery pnhib) a stesnis® ■?' {8 M T, M X, C| heal regksns ■» Gr<»»pin^(<|nery piasd.y ssoppiMt points) features - PotatNet Iayer{Wa: regions) tevam ~ Max Poolin^t eater s) retent features .SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSVVVVVXSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSVVVVVXSSSSSSSSSSSSSSSSSSSW.'.'.'.'.'.'.'.'.'.'.'.'.'.'.S- Example 2: PyTorch implementation code example for the temporal rooling operation, def shift_point(input): input = input.reshape(B, T, N, -1) output = torch.roll(input, stride=-l, dims=l) output[:, -1] = padding(input) output = output.reshape(B x T, N, -1) return output def getjayerjnput(input): support_points = input query_points = shift_point(in put) return support_points, query_points Example 3: PyTorch implementation code example for padding based on average change S lyforeMite for shift m the fttftx- IS x T, ft', ft) smpod tm«h 4ef shift mpal::: T H - ft » ks>'eh.«h!(foyai.; shkfes^' L. dhsss" ft smtiwlU -I | - ^a4kR»g(fopfo) wtp^i« ^xiLmh^e|B x T, M -ft return <fofget layer mputsmgex}: support pebm * query ~ shift trta wpjwt }»SUs( misery petals s padklingo is the fwetfos wd for ^whfotg fo the fest in the ftftst efoud setjnenee. prediction. def avg_padding(input): delta = torch.mean(input[:, 1:] - input]:, :-l], dim=l) padding_frame = input[:, -1] + delta return padding_frame Example 4: PyTorch implementation code example for padding based on PointLSTM prediction, class PointLSTM(nn.Module): def __init__(self): super(PointLSTM, self).__init__() self.encoder = nn.Sequential nn.Linear(input_dim, hidden_dim), nn.BatchNormld(hidden_dim), nn.ReLU() self.lstm = nn.LSTM(hidden_dim, hidden_dim, batch_first=True) self.decoder = nn.Sequential nn.Linear(hidden_dim, input_dim), nn.Tanh() def forward(self, input): point_features = self.encoder(input) point_features, _ = self.lstm(point_features) delta = 0.1 * self.decoder(point_features[:, -1]) padding_frame = input]:, -1] + delta return padding_frame The key technical means of the dynamic point cloud action recognition method, system, electronic device, and storage medium based on static models in this embodiment are described as follows: 1. The general framework of the TimeNet framework is specifically proposed. 2. The TimeNet framework enhances the temporal modeling capability of static 3D models in two ways through lightweight operations. 3. It is proposed to treat sequence data as batch processing, thereby maintaining implicit temporal awareness based on the shared weight paradigm during batch processing. 4. A plug-and-play temporal shift operation and prediction-based padding module are proposed to achieve additional explicit temporal assimilation. The technical characteristics of the TimeNet framework in this embodiment are as follows: 1. High parameter and computational efficiency: The TimeNet framework contains only 0.17M parameters and 0.31G FLOPs, and each frame input requires only 128 points. On the MSRAction-3D dataset, the TimeNet framework achieved 97.35% accuracy, surpassing the current state-of-the-art methods, and demonstrating excellent performance in terms of model parameters, computational complexity, and data storage requirements. 2. Wide applicability: The TimeNet framework can be applied to most existing static 3D models and achieves efficient dynamic point cloud processing capabilities through simple batch processing and temporal shift operations without relying on additional complex dynamic networks. 3. Implicit and explicit temporal awareness: By treating sequence data as batch processing, the TimeNet framework achieves implicit temporal awareness. At the same time, explicit temporal information assimilation is further enhanced through temporal shift operations and predictionbased padding modules. It is worth mentioning that the technical features such as the selection of processors involved in this patent application should be considered prior art. The specific structure, working principle, and possible control methods and spatial arrangement methods can be chosen using conventional choices in the field and should not be considered the inventive points of this patent. Therefore, this patent does not provide further detailed descriptions of these aspects. For those skilled in the art, it is still possible to modify the technical solutions described in the preceding embodiments or to make equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic point cloud action recognition method based on static models, characterized by comprising the following steps:Step SI: Generating and preprocessing three-dimensional point cloud data;Step S2: Constructing a deep learning model associated with a deep learning network;Step S3: Training the TimeNet framework, which includes a PointNet-like encoder.

2. The dynamic point cloud action recognition method based on static models according to claim 1, characterized in that Step SI is specifically implemented as the following steps:Step Sl.l: Acquiring a depth image sequence comprising T frames, where T is the length of the time series data;Step SI.2: Performing normalization processing on each frame of the depth image, wherein Step S1.2 includes Steps Sl.2.1, Sl.2.2, Sl.2.3, and Sl.2.4:Step Sl.2.1: Acquiring the length xx and width yy of the depth image, which represent the number of pixels in the length and width directions, respectively, and obtaining the corresponding empty grid matrix;Step Sl.2.2: Defining the depth mapping value focal, with focal = 280;Step Sl.2.3: Selecting the pixel positions with depth values d >0 in each empty grid matrix as the initial spatial values;Step Sl.2.4: Calculating the spatial coordinates of the three-dimensional point cloud data.

3. The dynamic point cloud action recognition method based on static models according to claim 1, characterized in that Step SI is specifically implemented as the following steps:Step Sl.l: Acquiring a point cloud video comprising T frames of depth images, where T is the length of the time series data;Step SI.2: Performing normalization processing on each frame of the depth image;Step SI.3: Classifying the dataset into training set, validation set, and test set;Step SI.4: Each training set includes multiple point cloud videos, each point cloud video containing T frames of three-dimensional point cloud data composed of N (x, y, z) spatial points;Step SI.5: For each T*N*3-dimensional data, setting a sampling step length step, an initial frame t, and a frame length c, with step = 1 and c = 24, where t <(T - c);Step SI.6: Segmenting each point cloud video into multiple overlapping video clips, each video clip having a length of Q frames;Step SI.7: Determining the shape of the video clip as 24*N*3;Step SI.8: Calculating the geometric center value (centroid) of each point cloud pc, calculating the maximum 2-norm (m) of each point cloud frame deviating from the center value, subtracting the geometric center value (centroid) from the three-dimensional point cloud data, and dividing the obtained value by the norm (m) to obtain the intermediate value (X);Step SI.9: Setting a random scaling value (scale) between [0.9, 1.1], performing the corresponding random scaling on each intermediate value (X) for data augmentation to obtain the initial output data.

4. The dynamic point cloud action recognition method based on static models according to claim 1, characterized in that the PointNet-like encoder in Step S3 performs the following steps:Input and preprocessing steps:Input data structure: Given a point cloud containing N points, represented as P = {pl, p2,..., pN}, where each point pi G R3 is specified by geometric coordinates {x, y, z};Video frame representation: The point cloud video contains T (here, T = 24) frames of point clouds, represented as V = {Pl, P2,..., PT};Extending the spatial encoder to the temporal domain steps:Spatial feature capture based on PointNet++: For static point cloud P, using a multi-layer perceptron for spatial set abstraction, expressed as Feature = f(P', P), where f represents standard PointNet++ layers, and P' is the downsampled query points generated from the original P;Reshaping for batch processing: Reshaping the point cloud video V through batch processing, expanding the formula to Feature = f(V', V), where V' and V respectively represent sets of support points and query points;Rolling operation: Applying a rolling operation in the time dimension to aggregate information between consecutive frames, effectively capturing dynamic changes in the time series:Frame misalignment through rolling operation: Adding a rolling operation in the time dimension so that support points and query points come from different point cloud frames, thereby aggregating temporal features from Pt to Pt+1;Temporal feature aggregation formula: Expressed as Feature = f(Pt, Pt+1'), where Pt are the support points, and Pt+1' are the downsampled query points from the next point cloud frame;Extending the spatial set abstraction to the temporal domain through operations on adjacent frames.

5. A dynamic point cloud action recognition system based on static models, characterized by executing the steps of the dynamic point cloud action recognition method based on static models according to any of claims 1-4.

6. An electronic device, characterized by comprising:a processor and a memory, wherein the memory stores executable instructions forthe processor; the processor is configured to execute the steps of the dynamic point cloud action recognition method based on static models according to any of claims 1-4 by executing the executable instructions.

7. A computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, performs the steps of the dynamic point cloud action recognition method based on static models according to any of claims 1-4.

Images