Learning scene real-time action recognition method based on spatiotemporal action detection technology

Abstract

A learning scene real-time action recognition method based on space-time action detection technology, comprising data preprocessing, constructing a learning scene space-time action detection network, training the learning scene space-time action detection network, saving a weight file, and testing the learning scene space-time action detection network; the shuffernetv2 model is adopted, the embedded DNL self-attention module can improve the dependence between channels, so that the features of different branches can be reasonably and smoothly aggregated, thereby highlighting the context relationship and enhancing the distinguishability of the features. The 2D features in the video frame and the context information between frames are fully utilized, and the balance between recognition accuracy and recognition speed is achieved. Compared with the prior art, the present application is simple, has few parameters, is fast and has high accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and particularly relates to a learning scene real-time action recognition method based on a space-time action detection technology. BACKGROUND

[0002] In order to maintain a good learning order, teachers need to monitor the actions of learners in real time and give real-time warnings for abnormal actions in learning scenes such as classrooms and examination rooms. With the development of deep neural networks and monitoring devices, the cost of obtaining monitoring video is reduced. It is challenging to use monitoring video data and computer vision related technologies to recognize the real-time actions of learners in learning scenes, which has a wide range of application scenarios and research value and has attracted more and more attention in the academic and industrial circles.

[0003] At present, the mainstream space-time action detection method is developing rapidly, and the recognition accuracy on public data sets is getting higher and higher, but it cannot balance the accuracy and speed. In the task of recognizing the actions of learners in learning scenes, the actions of learners need to be monitored in real time, and timely warnings are needed for abnormal actions.

[0004] Therefore, in the learning scene, a technical problem to be solved at present is to propose a real-time action recognition method that takes into account the accuracy and speed. SUMMARY

[0005] The technical problem to be solved by the application is to provide a real-time action recognition method in a learning scene with fast recognition speed and high recognition accuracy.

[0006] The technical scheme adopted to solve the above technical problem is: a learning scene real-time action recognition method based on a space-time action detection technology, comprising the following steps:

[0007] S1. Data preprocessing

[0008] Learning scene learner video is collected, a learner action recognition data set is generated according to the learning scene learner video, the learner action recognition data set is composed of original pictures and label files corresponding to the pictures, the original pictures are extracted from the learner action recognition data set, and the original pictures are adjusted to a uniform size by using a bilinear interpolation method and then divided into a training set and a test set according to a proportion;

[0009] S2. Constructing a learning scene space-time action detection network

[0010] The learning scene space-time action detection network is composed of a 2D branch network, a 3D branch network, a feature fusion network, and a classification regression network, the 2D branch network and the 3D branch network are connected to the feature fusion network in parallel, and the feature fusion network is connected to the classification regression network;

[0011] The 2D branch network is a ShuffleNetV2 network, which is used to extract 2D features of an input picture and generate a 2D basic feature map set {C1, C2, C3}, C1 is a 2D feature map set of a first convolutional layer in the ShuffleNetV2 network, C2 is a 2D feature map set of a second convolutional layer in the ShuffleNetV2 network, and C3 is a 2D feature map set of a third convolutional layer in the ShuffleNetV2 network;

[0012] The 3D branch network is a 3D-ShuffleNetV2 network, which is used to extract 3D features of an input picture and generate a 3D basic feature map set, and then perform dimension compression on the 3D basic feature map set to obtain a 3D basic feature compression map set {D1, D2, D3}, D1 is a 3D feature compression map set of a first convolutional layer in the 3D-ShuffleNetV2 network, D2 is a 3D feature compression map set of a second convolutional layer in the 3D-ShuffleNetV2 network, and D3 is a 3D feature compression map set of a third convolutional layer in the 3D-ShuffleNetV2 network;

[0013] The feature fusion network is composed of two convolutional layers and a DNL self-attention module, which is used to correspondingly stack the 2D basic feature map set {C1, C2, C3} and the 3D basic feature compression map set {D1, D2, D3} through channels, and then generate a fusion feature map set {E1, E2, E3} through two convolutional layers, and finally input the fusion feature map set {E1, E2, E3} into the DNL self-attention module to generate a new feature map set G;

[0014] The classification regression network is composed of a 1x1 convolutional layer, which is used to process the new feature map set G to obtain a recognition result, and the specific processing process is to generate a recognition feature map set with a feature map size of (3x(NumCls+5))x through the 1x1 convolutional layer, where 3 is the corresponding three prior anchor boxes, NumCls is the number of action recognition categories in the learning scene, is the height of the feature map, is the width of the feature map;

[0015] S3. Training the learning scene spatiotemporal action detection network

[0016] The training set and the label file corresponding to the training set are input into the learning scene spatiotemporal action detection network for training, and an SGD optimizer and a loss function are used to iteratively optimize the network to update the parameters until a set upper limit of iteration times is reached, that is, the training of the learning scene spatiotemporal action detection network is completed;

[0017] The loss function includes a CIOU loss function and a cross-entropy loss function, and the regression task adopts the CIOU loss function , the classification task adopts a cross-entropy loss function ;

[0018] =1-IOU+ +αν

[0019]

[0020] In the formula, , is the center point of the two rectangular frames, is the Euclidean distance between the two rectangular frames, is the distance of the diagonal line of the closure region of the two rectangular frames, α is a balance parameter, ν is a parameter measuring the consistency of the aspect ratio, IOU is the ratio of the intersection area to the union area, is the true value, is the estimated value;

[0021] S4. Save the weight file

[0022] The learning scene spatio-temporal action detection network saves the optimized weight file after F iterations of learning;

[0023] S5. Test the learning scene spatio-temporal action detection network

[0024] The weight file saved in step S4, the test data set and its corresponding label file are input into the learning scene spatio-temporal action detection network for testing, the detection results are post-processed by using the non-maximum suppression and confidence threshold method, and the accuracy and parameter quantity of the detection results are quantitatively evaluated.

[0025] As a preferred technical solution, in the step S1, the frame rate of the learner video of the learning scene is 30FPS; the learner action recognition data set generation method is to extract one frame image every 10 frames from the learner video of the learning scene as the original image, and to obtain the label file by classifying and labeling the extracted image; the number ratio of the training set and the test set is 8:2 or 7:3, and the picture size in the training set and the test set is 448x448.

[0026] As a preferred technical solution, the DNL self-attention module of the feature fusion network in the step S2 includes a pairwise NL branch, a unary NL branch, and an addition operation unit.

[0027] The pairwise NL branch includes a first 1x1 convolution block, a second 1x1 convolution block, a first whitening operation unit, a second whitening operation unit, a point multiplication unit, and a first normalization processing unit.

[0028] The first 1x1 convolution block is used for converting the input feature map into a local information feature map theta with a size of CxWxH and outputting to a first whitening operation unit, C is a plane channel number, W is a width of the feature map, and H is a height of the feature map, and the first whitening operation unit outputs a new local information feature map theta with a size of KxC / , K is a feature map side length, K=WH;

[0029] The second 1x1 convolution block is used for converting the input feature map into a global information feature map phi with a size of CxWxH and outputting to a second whitening operation unit, and the second whitening operation unit outputs the global information feature map phi with a size of KxC / ;

[0030] The point multiplication unit is used for multiplying the results output by the first whitening operation unit and the second whitening operation unit;

[0031] The first normalization processing unit is used for performing normalization processing on the result output by the point multiplication unit to obtain an intra-class information feature map with a size of KxK;

[0032] The unary NL branch includes a third 1x1 convolution block and a second normalization processing unit, the third 1x1 convolution block is used for converting the input feature map into a boundary information feature map with a size of 1xHxW and outputting to the second normalization processing unit, and the second normalization processing unit performs normalization processing to obtain a new boundary information feature map with a size of KxK;

[0033] The addition operation unit is used for adding the results output by the pairwise NL branch and the unary NL branch to obtain a final output result of the learning scene and space action detection network.

[0034] As a preferred technical solution, in the training process of step S3, the initial learning rate lr is set to 0.1, the number of pictures is 36 each time, the total number of iterations is 40, and the learning rate is decayed to 0.01, 0.001, 0.0001 and 0.00001 after 20, 25, 30 and 35 iterations.

[0035] As a preferred technical solution, in the training process of step S3, the initial learning rate lr is set to 0.5, the number of pictures is 24 each time, the total number of iterations is 48, and the learning rate is decayed to 0.05, 0.005, 0.0005 and 0.00005 after 20, 25, 30 and 35 iterations.

[0036] As a preferred technical solution, in the step S4 of saving the weight file, the F is 1000.

[0037] The beneficial effects of the present application are as follows:

[0038] The shuffernetv2 model of the present application is lightened, the embedded DNL self-attention module can improve the dependence of channel features, so that the features of different branches can be reasonably and smoothly aggregated, thereby highlighting the context relationship and enhancing the distinguishability of the features. The present application makes full use of the 2D features in the video frame and the context information between frames, and achieves a balance between recognition accuracy and recognition speed. Compared with the prior art, the present application is simple, has less parameters, is fast and has high accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a flowchart of embodiment 1 of the present application.

[0040] Figure 2 is nine pictures in a learner action recognition dataset.

[0041] Figure 3 is a structure diagram of the spatiotemporal action detection network in embodiment 1 of the present application.

[0042] Figure 4 is a structure diagram of the DNL self-attention module in embodiment 1 of the present application. DETAILED DESCRIPTION

[0043] The present application will be further described in detail below in combination with the drawings and embodiments, but the present application is not limited to the following embodiments.

[0044] Embodiment 1

[0045] In Figure 1 , the learning scene real-time action recognition method based on the spatiotemporal action detection technology of the present embodiment comprises the following steps:

[0046] S1. Data preprocessing

[0047] A learner video in a learning scene is collected, the frame rate of the video is 30FPS, and every 10 frames of images in the learner video in the learning scene are extracted as original pictures, such as Figure 2 The extracted images are classified and labeled to obtain a label file, and the original pictures and the label file constitute a learner action recognition dataset, the original pictures are extracted from the learner action recognition dataset, the size of the original pictures is uniformly adjusted to 448x448 by using a bilinear interpolation method, and then the original pictures are divided into a training set and a test set according to a ratio of 8:2 or 7:3;

[0048] S2. Constructing a learning scene spatiotemporal action detection network

[0049] The learning scene spatio-temporal action detection network is composed of a 2D branch network, a 3D branch network, a feature fusion network and a classification regression network, the 2D branch network and the 3D branch network are connected to the feature fusion network in parallel, and the feature fusion network is connected to the classification regression network, as shown in Figure 3 .

[0050] The 2D branch network is a ShuffleNetV2 network, which is used to extract 2D features of the input picture and generate a 2D basic feature set {C1, C2, C3}, C1 is a 2D feature set of the first convolutional layer in the ShuffleNetV2 network, C2 is a 2D feature set of the second convolutional layer in the ShuffleNetV2 network, and C3 is a 2D feature set of the third convolutional layer in the ShuffleNetV2 network;

[0051] The 3D branch network is a 3D-ShuffleNetV2 network, which is used to extract 3D features of the input picture and generate a 3D basic feature set, and then perform dimension compression on the 3D basic feature set to obtain a 3D basic feature compression set {D1, D2, D3}, D1 is a 3D feature compression set of the first convolutional layer in the 3D-ShuffleNetV2 network, D2 is a 3D feature compression set of the second convolutional layer in the 3D-ShuffleNetV2 network, and D3 is a 3D feature compression set of the third convolutional layer in the 3D-ShuffleNetV2 network;

[0052] The feature fusion network is composed of two convolutional layers and a DNL self-attention module, which is used to correspondingly stack the 2D basic feature set {C1, C2, C3} and the 3D basic feature compression set {D1, D2, D3} through channels, and then generate a fusion feature set {E1, E2, E3} through two convolutional layers, and finally input the fusion feature set {E1, E2, E3} into the DNL self-attention module to generate a new feature set G;

[0053] As Figure 4 , the DNL self-attention module of the embodiment includes a pairwise NL branch, a unary NL branch and an addition operation unit; the pairwise NL branch includes a first 1x1 convolutional block, a second 1x1 convolutional block, a first whitening operation unit, a second whitening operation unit, a point multiplication unit and a first normalization processing unit; the first 1x1 convolutional block is used to convert the input feature map into a local information feature map θ with a size of CxWxH and output to the first whitening operation unit, C is the number of plane channels, W is the width of the feature map, H is the height of the feature map, the first whitening operation unit outputs a new local information feature map θ / , K is the edge length of the feature map, K = WH. The second 1x1 convolution block is used to convert the input feature map into a global information feature map φ with a size of CxWxH and output to the second whitening operation unit, and the second whitening operation unit outputs a global information feature map φ with a size of KxC / ; a point multiplication unit for multiplying the results output by the first whitening operation unit and the second whitening operation unit; a first normalization processing unit for normalizing the result output by the point multiplication unit to obtain an intra-class information feature map with a size of KxK;

[0054] The unary NL branch includes a third 1x1 convolution block and a second normalization processing unit, the third 1x1 convolution block is used to convert the input feature map into a boundary information feature map with a size of 1xHxW and output to the second normalization processing unit, and the normalization processing unit is used to normalize the boundary information feature map to obtain a new boundary information feature map with a size of KxK;

[0055] An addition operation unit is used to add the results output by the pairwise NL branch and the unary NL branch to obtain the final output result of the learning scene spatio-temporal action detection network;

[0056] The classification regression network is composed of a 1x1 convolution layer, which is used to process the new feature map set G to obtain the recognition result. The specific processing process is to generate a recognition feature map set with a size of (3x(NumCls+5))x through the 1x1 convolution layer, where 3 is the corresponding three prior anchor boxes, NumCls is the number of action recognition categories in the learning scene, and 5 is the regression value of the prior anchor box and the confidence score of the regression value, is the height of the feature map, is the width of the feature map;

[0057] S3. Training the learning scene spatio-temporal action detection network

[0058] The training set and the label file corresponding to the training set are input into the learning scene spatio-temporal action detection network for training. The initial learning rate lr is set to 0.1, the SGD optimizer and the loss function are used to iteratively optimize the network to update the parameters. The number of pictures is 36 each time, the total number of iterations is 40, the learning rate is decayed to 0.01, 0.001, 0.0001 and 0.00001 after 20, 25, 30 and 35 iterations, respectively. When the iteration reaches the set iteration number, the training of the learning scene spatio-temporal action detection network is completed.

[0059] The loss function includes a CIOU loss function and a cross-entropy loss function, and the regression task adopts the CIOU loss function The overlapping area, center point distance and aspect ratio information optimization model of the target is optimized to improve the detection accuracy; the cross entropy loss function is used for the classification task for alleviating the problem of gradient diffusion;

[0060] =1-IOU+ +αν

[0061]

[0062] In the formula, 、 is the center point of the two rectangular frames, is the Euclidean distance between the two rectangular frames, is the distance of the diagonal line of the closure area of the two rectangular frames, which is used to prevent the value of the loss function from being too large and improve the convergence speed, alpha is a balance parameter, nu is a parameter for measuring the consistency of the aspect ratio, IOU is the ratio of the intersection area to the union area, is the true value, is the estimated value;

[0063] S4. Save the weight file

[0064] The learning scene spatio-temporal action detection network saves the optimized weight file after F=1000 iterations;

[0065] S5. Test the learning scene spatio-temporal action detection network

[0066] The weight file saved in step S4, the test data set and the corresponding label file are input into the learning scene spatio-temporal action detection network for testing, the non-maximum suppression and confidence threshold method are used for post-processing of the detection result, and the accuracy and parameter quantity of the detection result are quantitatively evaluated.

[0067] Embodiment 2

[0068] In this embodiment, in the training of the learning scene spatio-temporal action detection network in step S3, the initial learning rate is set to lr as 0.5, the picture quantity is 24 for each iteration during training, the total number of iterations is 48, and the learning rate is attenuated to 0.05, 0.005, 0.0005 and 0.00005 after 20, 30, 40 and 45 iterations respectively.

[0069] The other operation steps are the same as those in embodiment 1.

[0070] Experiment 1

[0071] In order to verify the beneficial effects of the present application, the inventors tested the action recognition contrast experiment in the learner action recognition dataset using the learning scene real-time action recognition method based on the spatio-temporal action detection of example 1 and the prior art spatio-temporal action detection method, and the comparison results are as shown in Table 1 below.

[0072] Table 1. Results of learning scene spatio-temporal action detection network in BRL dataset

[0073]

[0074] The Residual module is proposed in the ResNet classification network of the prior art spatio-temporal action detection method, which solves the degradation problem of the neural network by fitting residual mapping, thereby deepening the network layer and further improving the recognition accuracy, and an aggregation transformation is proposed, which reduces the network parameter amount through grouped convolution, and at the same time, the feature layer is concatenated to ensure that the model can maintain a high accuracy; the MobileNetV1 in the prior art spatio-temporal action detection method greatly reduces the parameter amount of the model through deep separable convolution, and uses 1x1 convolution to adjust the channel number, and the MobileNetV2 adds an inverted residual network structure to improve the model accuracy on the basis of V1, and a linear activation function is added in the last convolution layer to prevent loss of too much feature information. The ShuffleNetV1 of the present application changes the dense 1x1 convolution in the resnet to grouped convolution, and reduces the calculation amount through DWConv, and finally uses channal shuffle to solve the boundary effect caused by multiple grouped convolutions, and the ShuffleNetV2 considers the memory access cost of different structures, and improves the running speed of the model by reducing the Element-wise operation.

[0075] Analysis of experimental results:

[0076] The spatio-temporal action detection model of the present application has the highest speed and the least parameter amount compared with the existing 3D backbone model, and can complete the real-time recognition requirement in the case of moderate accuracy and meet the needs of industrial deployment.

[0077] Test 2

[0078] In order to verify the beneficial effects of the feature fusion network of the present application, the inventors performed an ablation experiment in the learner action recognition dataset using the learning scene real-time action recognition method based on the spatio-temporal action detection of example 1, and the results are as shown in Table 2 below.

[0079] Table 2 Ablation experiment

[0080]

[0081] Analysis of experimental results:

[0082] The DNL self-attention module of the feature fusion network of the application can effectively improve the accuracy of the model without affecting the real-time detection performance of the model. Meanwhile, compared with the Non-Local module, it is shown that the fusion effect of the DNL self-attention module is better than that of the 2D and 3D modules, and the recognition accuracy of the whole model can be improved.

Claims

1. A learning scene real-time action recognition method based on a spatio-temporal action detection technology, characterized in that, The method comprises the following steps: S1. Data preprocessing Collecting a learner video of a learning scene, generating a learner action recognition dataset from the learner video of the learning scene, the learner action recognition dataset being composed of original pictures and a label file corresponding to the pictures, extracting the original pictures from the learner action recognition dataset, adjusting the original pictures to a uniform size by using a bilinear interpolation method, and dividing the adjusted pictures into a training set and a test set according to a proportion; S2. Constructing a learning scene spatiotemporal action detection network The learning scene spatiotemporal action detection network is composed of a 2D branch network, a 3D branch network, a feature fusion network, and a classification regression network, the 2D branch network and the 3D branch network being connected to the feature fusion network in parallel, and the feature fusion network being connected to the classification regression network; The 2D branch network is a ShuffleNetV2 network, which is used to extract 2D features of input pictures and generate a 2D basic feature map set {C1, C2, C3}, C1 being a 2D feature map set of a first convolutional layer in the ShuffleNetV2 network, C2 being a 2D feature map set of a second convolutional layer in the ShuffleNetV2 network, and C3 being a 2D feature map set of a third convolutional layer in the ShuffleNetV2 network; The 3D branch network is a 3D-ShuffleNetV2 network, which is used to extract 3D features of input pictures and generate a 3D basic feature map set, and then perform dimension compression on the 3D basic feature map set to obtain a 3D basic feature compression map set {D1, D2, D3}, D1 being a 3D feature compression map set of a first convolutional layer in the 3D-ShuffleNetV2 network, D2 being a 3D feature compression map set of a second convolutional layer in the 3D-ShuffleNetV2 network, and D3 being a 3D feature compression map set of a third convolutional layer in the 3D-ShuffleNetV2 network; The feature fusion network is composed of two convolutional layers and a DNL self-attention module, which is used to correspondingly stack the 2D basic feature map set {C1, C2, C3} and the 3D basic feature compression map set {D1, D2, D3} through channels, and then generate a fusion feature map set {E1, E2, E3} through the two convolutional layers, and finally input the fusion feature map set {E1, E2, E3} into the DNL self-attention module to generate a new feature map set G; The classification regression network is composed of a 1*1 convolution layer, which is used to process the new feature map set G to obtain a recognition result. The specific processing process is that the new feature map set G is generated into a recognition feature map set with a feature map size of (3*(NumCls+5))x , where 3 is the corresponding three prior anchor boxes, NumCls is the number of action recognition categories in the learning scene, is the height of the feature map, is the width of the feature map. S3. Training the learning scene spatiotemporal action detection network The training set and the label file corresponding to the training set are input into the learning scene spatiotemporal action detection network for training, and an SGD optimizer and a loss function are used to iteratively optimize the network to update parameters until a set upper limit of iteration times is reached, that is, the training of the learning scene spatiotemporal action detection network is completed; The loss function comprises a CIOU loss function and a cross-entropy loss function, the regression task adopts the CIOU loss function, and the classification task adopts the cross-entropy loss function . ​ = 1 - IOU + αv wherein, , is the center point of the two rectangular boxes, is the Euclidean distance between the two rectangular boxes, is the distance of the diagonal of the closure region of the two rectangular boxes, a is a balance parameter, v is a parameter measuring the consistency of the aspect ratio, IOU is the ratio of the intersection region to the union region, is the true value, is the estimated value; S4. Saving a weight file The learning scene spatiotemporal action detection network saves the optimized weight file after F iterations; S5. Testing the learning scene spatiotemporal action detection network The weight file saved in step S4, the test data set and its corresponding label file are input into the learning scene spatio-temporal action detection network for testing, the detection results are post-processed by the non-maximum suppression and confidence threshold method, and the accuracy and parameter quantity of the detection results are quantitatively evaluated.

2. The learning scenario real-time action recognition method based on the space-time action detection technology according to claim 1, characterized in that, In the step S1, the frame rate of the learner video in the learning scene is 30 FPS; the learner action recognition data set generation method is to extract one frame of image as a raw image every 10 frames from the learner video in the learning scene, and the extracted image is classified and labeled to obtain a label file; the number ratio of the training set and the test set is 8:2 or 7:3, and the picture size in the training set and the test set is 448x448.

3. The method of claim 1, wherein the learning scenario real-time action recognition method based on the space-time action detection technique is characterized by, The DNL self-attention module of the feature fusion network in the step S2 includes a pairwise NL branch, a unary NL branch and an addition operation unit. The pairwise NL branch includes a first 1x1 convolution block, a second 1x1 convolution block, a first whitening operation unit, a second whitening operation unit, a point multiplication unit and a first normalization processing unit. The first 1x1 convolution block is used for converting an input feature map into a local information feature map θ with a size of CxWxH and outputting to a first whitening operation unit, C is a plane channel number, W is a width of the feature map, and H is a height of the feature map, the first whitening operation unit outputs a new local information feature map θ with a size of KxC / , K is a feature map edge length, K=WH. The second 1x1 convolution block is configured to convert the input feature map into a global information feature map φ with a size of CxWxH and output to a second whitening operation unit, and the second whitening operation unit outputs a global information feature map φ with a size of KxC / ; The point multiplication unit is configured to multiply the results output by the first whitening operation unit and the second whitening operation unit. The first normalization processing unit is configured to normalize the result output by the point multiplication unit to obtain an intra-class information feature map with a size of KxK. The unary NL branch includes a third 1x1 convolution block and a second normalization processing unit, the third 1x1 convolution block is configured to convert the input feature map into a boundary information feature map with a size of 1xHxW, and output to the second normalization processing unit for normalization processing to obtain a new boundary information feature map with a size of KxK. The addition operation unit is configured to add the results output by the pairwise NL branch and the unary NL branch to obtain the final output result of the learning scene spatio-temporal action detection network.

4. The learning scenario real-time action recognition method based on the space-time action detection technology according to claim 1, characterized in that, In the training process of the step S3, the initial learning rate lr is set to 0.1, the number of pictures is 36 per iteration, the total number of iterations is 40, and the learning rate is decayed to 0.01, 0.001, 0.0001 and 0.00001 after 20, 25, 30 and 35 iterations.

5. The learning scenario real-time action recognition method based on the space-time action detection technology according to claim 1, characterized in that, In the training process of the step S3, the initial learning rate lr is set to 0.5, the number of pictures is 24 per iteration, the total number of iterations is 48, and the learning rate is decayed to 0.05, 0.005, 0.0005 and 0.00005 after 20, 25, 30 and 35 iterations.

6. The learning scenario real-time action recognition method based on the space-time action detection technology according to claim 1, characterized in that, In the step S4 of saving the weight file, the F is 1000.

Images