Crawling tactical action evaluation method based on head posture capture and orientation calculation

Abstract

The present invention relates to an action evaluation method, and in particular to a crawling tactical action evaluation method based on head posture capture and orientation calculation. The method solves the technical problem that the existing training for head posture evaluation does not have a set of objective and quantitative common standards and cannot replicate experience on a large scale, resulting in low evaluation efficiency. The method of the present invention includes the following steps: 1) constructing a first backbone network to collect and identify the student's head image in real time; 2) calculating the head azimuth; 2.1) constructing a second backbone network, which is a ResNet50 backbone network; 2.2) training the ResNet50 backbone network and obtaining actual labeled data; 2.3) transmitting the identified student's head image to the trained ResNet50 backbone network for head azimuth calculation; 3) judging whether the yaw angle, roll angle or pitch angle of the head at the current moment falls within the corresponding set range; if not, performing corresponding deduction operations according to the deduction rules and giving a score; if so, directly giving a score.

Description

Technical Field

[0001] The present invention relates to an action evaluation method, in particular to a crawling tactical action evaluation method based on head posture capture and orientation calculation. Background Art

[0002] Head posture refers to the position and orientation of the head. Currently, there is no objective, universally accepted standard for evaluating head posture in training. For example, during crawl training, whether a trainee's forward gaze meets the standard is currently judged primarily by the training officer based on prior experience. This human judgment leads to subjective results, lacks quantification, and is difficult to replicate on a large scale, resulting in low evaluation efficiency. Summary of the Invention

[0003] The present invention aims to address the technical issues of the lack of a set of objective and quantitative standards for judging head posture in existing training, which often relies on the training officer's judgment based on previous experience. This human judgment easily leads to subjective judgment results, the inability to quantify conclusions, and the inability to replicate experience on a large scale, resulting in low evaluation efficiency. The present invention provides an evaluation system and method based on human head posture capture and orientation calculation. The present invention uses image recognition to determine head posture parameters within a global coordinate system. Using human head posture capture and orientation calculation technology, the trainee's head posture information is first collected. The angle presented in this posture is then analyzed in horizontal and vertical sections, and the yaw angle is calculated. Finally, a threshold range is set to limit or determine whether the yaw angle of the trainee's head posture meets the requirements of the training syllabus. The present invention can objectively evaluate the trainee's head posture and quantify the trainee's assessment standards.

[0004] The technical solution of the present invention is:

[0005] A method for evaluating crawling tactical actions based on head posture capture and orientation calculation is special in that it includes the following steps:

[0006] 1) Build the first backbone network to collect and recognize the student's head image in real time;

[0007] 2) Calculate the head orientation angle

[0008] 2.1) Constructing a second backbone network, which is a ResNet50 backbone network; the ResNet50 backbone network includes three fully connected layers, each corresponding to the yaw angle, roll angle, and pitch angle of the head orientation angle; the mapping interval of the ResNet50 backbone network is in the range of [-99, +99]; each fully connected layer makes a separate prediction within the mapping interval, and the fully connected numbers for the separate predictions are all binary numbers;

[0009] 2.2) Train the constructed ResNet50 backbone network to obtain the trained ResNet50 backbone network and obtain the actual labeled data;

[0010] 2.3) The student's head image identified in step 1) is fed into the ResNet50 backbone network trained in step 2.2) to calculate the head orientation angle;

[0011] 2.3.1) Map the yaw, roll, or pitch angles of the student's head orientation in the image to the interval [-99, +99]. Classify the 199 values in the mapped interval from -99 to +99 into groups of three, with the intervals between them as discrete categories.

[0012] 2.3.2) The discrete categories obtained in step 2.3.1) are normalized into probabilities by applying the softmax function. The sum of the probabilities of all discrete category data in the fully connected layer is 1.

[0013] 2.3.3) Multiply the probability value obtained in step 2.3.2) by the corresponding discrete category and sum them to obtain the expected value. Calculate the mean square error between the expected value and the angle of the actual annotated data, and finally output the head orientation angle through end-to-end training.

[0014] 3) Determine whether the yaw angle, roll angle, or pitch angle of the head azimuth at the current moment falls within the corresponding set range; if not, perform the corresponding deduction operation according to the deduction rules and give a score; if so, give the score directly.

[0015] Furthermore, the step 1 is specifically as follows:

[0016] 1.1) Constructing a first backbone network, which is a Darknet-53 backbone network; the Darknet-53 backbone network comprises a plurality of convolutional blocks and fully connected layers connected sequentially from the input to the output; each convolutional block comprises a convolutional layer, a batch normalization layer, and an activation layer connected in sequence; the convolutional layer comprises 1x1 and 3x3 convolution kernels; when the convolution step size defaults to 1, the padding defaults to uniform; when the convolution step size is 2, the padding is enabled;

[0017] 1.2) Downsample the Darknet-53 backbone network with a sampling step of 1 or 2. Then, concatenate the upsampled layers of the middle and penultimate layers of the Darknet-53 backbone network to form a new optimized network, which is used to iterate new sample training to obtain the head model.

[0018] 1.3) Optimizing and training the head model using a loss function; the loss function is one or more of target positioning offset loss, target confidence loss, and target classification loss;

[0019] 1.4) Collect the head image information of the student to be tested, transmit it to the head model, and recognize the student's head image.

[0020] Furthermore, in step 1.2), the Darknet-53 backbone network is downsampled 5 times, with a sampling step of 2 each time, reducing the output feature map to 1 / 32 of the original image.

[0021] Furthermore, in step 2.1), the output of the ResNet50 backbone network is first mapped to the range of 0 to 99, and then the mapped value is multiplied by 3 and then subtracted by 99, so that the range is mapped to the range of [-99, +99].

[0022] Furthermore, in step 2.3.1), the 199 values in the mapping interval range from -99 to +99 are classified in sequence, specifically including the starting point -99 degrees, the end point is set to 102 degrees and does not include 102 degrees, and every 3 numbers are an interval, for a total of 66 intervals.

[0023] Beneficial effects of the present invention:

[0024] 1. The present invention can capture the position and direction information of the subject's head while crawling in real time, such as head yaw, pitch and roll posture, and can also quantify the assessment information and quickly analyze and judge whether the subject's crawling action is illegal.

[0025] 2. The present invention combines the Darknet-53 backbone network and the ResNet50 backbone network to obtain a human head posture detection and recognition backbone network model and a head azimuth angle calculation model, which strikes a good balance between computational complexity and accuracy. It can ensure accuracy and improve the efficiency of the network model without adding additional computational overhead. On this basis, the calculation rate of the method of the present invention is 2 to 3 times faster than that of the traditional method.

[0026] 3. The present invention collects head posture information and uses artificial intelligence to perform data analysis to objectively and accurately evaluate soldier training.

[0027] 4. The present invention is intelligent and efficient, and can quickly and accurately achieve multi-task head posture estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic diagram of the network structure of the ResNet50 backbone network, where MSE is the mean square error. DETAILED DESCRIPTION

[0029] The present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0030] The present invention provides a method for evaluating crawling tactical actions based on head posture capture and orientation calculation, comprising the following steps:

[0031] 1) Build the first backbone network to collect and recognize the student's head image in real time;

[0032] 1.1) Build the first backbone network, which is the Darknet-53 backbone network;

[0033] The Darknet-53 backbone network consists of multiple convolutional blocks and fully connected layers connected sequentially from the input to the output. Each convolutional block consists of a convolutional layer, a batch normalization layer, and an activation layer connected in sequence. The convolutional layers consist of 1x1 and 3x3 kernels. When the convolution stride is set to 1 by default, the padding is uniform by default; when the convolution stride is set to 2, padding is enabled. The Darknet-53 backbone network structure has 52 convolutional layers, the last of which is a fully connected layer, for a total of 53 layers, as shown in Table 1.

[0034] Table 1 Darknet-53 backbone network structure

[0035]

[0036] 1.2) Downsample the Darknet-53 backbone network five times, with a sampling step of 2 each time, reducing the output feature map to 1 / 32 of the original image. This ensures that the feature map is neither too large nor too small, which is conducive to model training. Then, the upsampled middle layer and the penultimate layer of the Darknet-53 backbone network are spliced together to form a new optimized network, which is used for iterative training and learning with new samples to obtain the head model.

[0037] 1.3) Optimizing and training the head model through a loss function; the loss function is one or more of target positioning offset loss, target confidence loss, and target classification loss.

[0038] When an object's center falls within the grid, we calculate the target classification loss, target confidence loss, and the loss of the coordinate information of the predicted box with the largest IOU (image segmentation metric) between the predicted box and the actual annotated data. When no object's center falls within the grid, we only calculate the confidence loss. A high confidence level indicates that the object's center is within the grid, and calculating the target localization loss and target classification loss makes sense.

[0039] 1.4) Collect the head image information of the student to be tested, transmit it to the head model, and recognize the student's head image;

[0040] 2) Calculate the head orientation angle;

[0041] 2.1) Build the second backbone network, which is the ResNet50 backbone network;

[0042] like Figure 1 As shown in the figure, the ResNet50 backbone network includes three fully connected layers, which correspond to the yaw angle, roll angle and pitch angle of the head azimuth respectively; the mapping interval range of the ResNet50 backbone network is [-99, +99]; each fully connected layer makes a separate prediction within the mapping interval, and the fully connected numbers for separate predictions are all bin numbers (binary numbers); the output of the ResNet50 backbone network is first mapped to the interval range of 0 to 99, and then the mapping value is multiplied by 3 and then subtracted by 99, so that the interval range is mapped to the interval range of [-99, +99].

[0043] 2.2) Train the constructed ResNet50 backbone network to obtain the trained ResNet50 backbone network and obtain the actual labeled data;

[0044] 2.3) The student's head image identified in step 1.4 is fed into the ResNet50 backbone network trained in step 2.2) to calculate the head orientation angle;

[0045] 2.3.1) Map the yaw, roll, or pitch angles in the participant's head image to the interval [-99, +99]. Classify each of the 199 values in the mapped interval (-99 to +99) into groups of three, with the intervals between them as discrete categories. The classification process includes a starting point of -99 degrees and an end point of 102 degrees, excluding 102 degrees. Each interval consists of three values, for a total of 66 intervals.

[0046] 2.3.2) The discrete categories obtained in step 2.3.1) are normalized into probabilities by applying the softmax function. The sum of the probabilities of all discrete category data in the fully connected layer is 1.

[0047] 2.3.3) Multiply the probability value obtained in step 2.3.2) by the corresponding discrete category and sum them to obtain the expected value; calculate the mean square error between the expected value and the angle of the actual annotated data; and use the cross-loss entropy to calculate the loss for the discrete categories obtained in step 2.3.1); finally, output the head orientation angle through end-to-end training;

[0048] 3) Determine whether the current head yaw, roll, or pitch angle falls within the corresponding set range. If not, deduct points according to the deduction rules and assign a score. If so, assign the score directly. The corresponding setting range is determined based on experience.

Claims

1. A method for evaluating crawling tactical actions based on head posture capture and orientation calculation, characterized in that: The following steps are involved: 1) Build the first backbone network to collect and recognize the student's head image in real time; 2) Calculate the head orientation angle 2.1) Constructing a second backbone network, which is a ResNet50 backbone network; the ResNet50 backbone network includes three fully connected layers, each corresponding to the yaw angle, roll angle, and pitch angle of the head orientation angle; the mapping interval of the ResNet50 backbone network is in the range of [-99, +99]; each fully connected layer makes a separate prediction within the mapping interval, and the fully connected numbers for the separate predictions are all binary numbers; 2.2) Train the constructed ResNet50 backbone network to obtain the trained ResNet50 backbone network and obtain the actual labeled data; 2.3) The student's head image identified in step 1) is fed into the ResNet50 backbone network trained in step 2.2) to calculate the head orientation angle; 2.3.1) Map the yaw, roll, or pitch angles of the student's head orientation in the image to the interval [-99, +99]. Classify the 199 values in the mapped interval from -99 to +99, starting from -99 degrees inclusive and ending at 102 degrees exclusive, with each interval consisting of three values, for a total of 66 intervals; use each interval as a discrete category. 2.3.2) The discrete categories obtained in step 2.3.1) are normalized into probabilities by applying the softmax function. The sum of the probabilities of all discrete category data in the fully connected layer is 1. 2.3.3) Multiply the probability value obtained in step 2.3.2) by the corresponding discrete category and sum them to obtain the expected value. Calculate the mean square error between the expected value and the angle of the actual annotated data, and finally output the head orientation angle through end-to-end training. 3) Determine whether the yaw angle, roll angle, or pitch angle of the head azimuth at the current moment falls within the corresponding set range; if not, perform the corresponding deduction operation according to the deduction rules and give a score; if so, give the score directly.

2. The method for evaluating crawling tactical actions based on head posture capture and orientation calculation according to claim 1, characterized in that: The step 1 is specifically as follows: 1.1) Constructing a first backbone network, which is a Darknet-53 backbone network; the Darknet-53 backbone network comprises a plurality of convolutional blocks and fully connected layers connected sequentially from the input to the output; each convolutional block comprises a convolutional layer, a batch normalization layer, and an activation layer connected in sequence; the convolutional layer comprises 1x1 and 3x3 convolution kernels; when the convolution step size defaults to 1, the padding defaults to uniform; when the convolution step size is 2, the padding is enabled; 1.2) Downsample the Darknet-53 backbone network with a sampling step of 1 or 2. Then, concatenate the upsampled layers of the middle and penultimate layers of the Darknet-53 backbone network to form a new optimized network, which is used to iterate new sample training to obtain the head model. 1.3) Optimizing and training the head model using a loss function; the loss function is one or more of target positioning offset loss, target confidence loss, and target classification loss; 1.4) Collect the head image information of the student to be tested, transmit it to the head model, and recognize the student's head image.

3. The method for evaluating crawling tactical actions based on head posture capture and orientation calculation according to claim 2, characterized in that: In step 1.2), the Darknet-53 backbone network is downsampled 5 times, with a sampling step of 2 each time, reducing the output feature map to 1 / 32 of the original image.

4. The method for evaluating crawling tactical actions based on head posture capture and orientation calculation according to claim 1, 2 or 3, wherein: In step 2.1), the output of the ResNet50 backbone network is first mapped to the range of 0 to 99, and then the mapped value is multiplied by 3 and then subtracted by 99, so that the range is mapped to the range of [-99, +99].

Images