Laser targeting identification method based on YOLOV5

Abstract

The invention belongs to the field of laser targeting training and intelligent target range detection, and discloses a laser targeting recognition method based on YOLOV5. A YOLO target detection framework is adopted, YOLOV5n-7. Pt with high detection speed is taken as a core model, and the requirements of target surface image rapid detection and score real-time calculation are met. The image quality is guaranteed through system initialization and camera parameter configuration, and a target surface, laser and a bull's-eye are accurately recognized through preprocessing, model feature extraction, combination with algorithms such as prior frame scanning and the like. The score is quickly calculated based on the recognition result, and the marked image is cropped for analysis. Model training adopts multi-scale enhancement, CIoU is introduced into a loss function to improve precision, and end-to-end closed loop is realized. The system can stably operate in a complex scene, solves the problems of low efficiency and poor precision of manual target judgment, feeds back a result in real time, stores data, guarantees fairness and justice, and promotes scientific, practical and digital upgrading of shooting training.

Description

Technical Field

[0001] This invention, which belongs to the field of laser target training and intelligent range detection, employs computer vision and machine learning techniques to calculate shooting scores through image preprocessing. This solves the problems of low efficiency and poor accuracy in manual target judgment, automates and accurately detects shooting results, improves training feedback speed and data traceability, and provides technical support for the standardization and digitalization of shooting training and the optimization of range management, enabling the scientific improvement of shooting skills. Background Art

[0002] In laser target training, military shooting and competitive shooting scenarios, traditional manual target judging has defects such as low efficiency and poor accuracy, which makes it difficult to meet the requirements of digital training for automatic recognition of target surface images, precise target detection and rapid calculation of scores. To this end, the present invention is based on computer vision and machine learning technology to build a process automation system: data adaptability is enhanced through image preprocessing, multi-target features are extracted by pre-training models, and an end-to-end closed loop of target classification, coordinate regression and score calculation is achieved by combining prior frame scanning and loss function optimization. This solution solves the core problems of traditional target judging, provides technical support for real-time data feedback and fair evaluation of scores in shooting training, promotes the upgrading of training to scientific and practical training, effectively improves detection accuracy and efficiency, reduces labor costs and errors, and enables the construction of intelligent shooting ranges and quantitative training of shooting skills. Summary of the Invention

[0003] In order to solve the problems existing in the background technology, the present invention proposes a laser target recognition method based on YOLO_V5 to achieve real-time feedback on shooting training results, thereby helping to improve the scientificity and practical level of shooting training.

[0004] A laser target recognition method based on YOLO_V5, characterized by comprising the following steps:

[0005] S1: First, complete the initialization of system parameters and optimize the operating parameters of each module to ensure that the system enters the optimal operating state; simultaneously complete the fine configuration of camera parameters and ensure that the captured image quality meets the subsequent algorithm processing standards by adjusting parameters such as focal length, exposure, and resolution.

[0006] S2: After receiving the shooting signal, the system immediately captures the shooting image and performs a series of preprocessing operations on the image: the image size is adjusted to 800x600 pixels and the image data is numerically normalized to improve the efficiency and accuracy of subsequent image recognition and processing.

[0007] S3: The pre-processed image is then transferred to the pre-trained Target_Laser.pt model. This model has been trained with a large amount of data and has the ability to accurately identify the target surface, laser, and bull's eye features.

[0008] S4: The Target_Laser.pt model uses a pre-trained convolutional neural network to extract features from the input image. These features will be used for subsequent target classification and coordinate regression tasks.

[0009] S5: Based on the feature extraction results, the system uses a predefined priori frame to scan the feature map, and gradually performs target classification and coordinate regression on each priori frame to determine the position and size of the target surface, laser and bull's eye.

[0010] S6: The system evaluates the prediction boxes using a loss function and uses a non-maximum suppression algorithm to filter and merge overlapping prediction boxes to eliminate redundant results, ensuring that the final output prediction results are both accurate and concise.

[0011] S7: After completing the image recognition task of the current frame, the system calculates the distance from the bull's eye to the laser and the distance from the bull's eye to the outermost ring of the target surface based on the positions of the target surface, laser and bull's eye obtained in step S5, and calculates the shooting result of this time.

[0012] S8: Based on the image processed in step S2 and the position and size information obtained in step S5, the system crops the target surface shooting image and draws rectangular frames at the corresponding positions in the image to mark the target surface, laser and bull's eye. At the same time, the shooting score is displayed in the upper left corner of the image and saved.

[0013] S9: Finally, the system outputs the detection results, including the predicted category, confidence level, and box location. This information is displayed to the user in an appropriate manner or used for subsequent data analysis. When a new shooting signal is received, the system returns to step S2 for a new round of image processing.

[0014] Furthermore, the camera in step S1 is used for real-time monitoring and transmits real-time data to a computer.

[0015] Furthermore, the Target_Laser.pt model in the S3 step requires dataset collection, category labeling, and model training.

[0016] Furthermore, the preprocessing operation of the Target_Laser.pt model training in the S3 step adopts Multi-Scale Augmentation. By dynamically adjusting the input resolution during the training process, the model learns the features of small objects at different scales to reduce the risk of overfitting to a single scale.

[0017] Furthermore, the loss function of S6 and the method of non-maximum suppression are:

[0018] There are two bounding box areas, B is the predicted border area, and B gtis the actual border area, measuring B and B gt The distance, overlap rate and shape consistency between them are calculated by the following formula:

[0019]

[0020]

[0021]

[0022] Where v is the normalized difference between the aspect ratio of the predicted box and the true box, and α is a balance factor that weighs the loss caused by the aspect ratio and the loss caused by the IoU part.

[0023] Furthermore, the output detection result of step S9 includes the predicted category (indicating whether the target was hit), the confidence level (indicating the reliability of the prediction result), and the position of the box (indicating the location of the target category in the image).

[0024] The present invention uses YOLOV5n-7.pt with fast detection speed as the research object to construct a detection network to meet the needs of rapid detection of target surface images and real-time calculation of scores in laser target shooting training and competitive shooting scenes. During training or competition, the system can monitor the status of the target surface, laser and bull's eye in real time, and accurately identify their positions and sizes. Once it is recognized that the laser hits the target surface, the system can quickly calculate the distance from the bull's eye to the laser and the distance from the bull's eye to the outermost ring of the target surface, and then quickly and accurately calculate the shooting score, and display the results to the trainers or referees in a timely manner. At the same time, the system will also crop the target surface shooting image, mark the position of the target surface, laser and bull's eye, and save the image for subsequent analysis, thereby improving the training feedback speed and data traceability, ensuring the fairness and justice of target shooting training and competitions, and promoting the upgrading of shooting training to scientific and practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Schematic diagram of the method and hardware configuration of the present invention;

[0026] Figure 2 This is a flow chart of the laser target recognition program of the method of the present invention;

[0027] Figure 3 Flowchart of the training procedure of the method of the present invention; DETAILED DESCRIPTION

[0028] In order to more clearly present the technical features and practical application scenarios of the present invention, it will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] Example 1.

[0030] like Figure 1The figure shows a schematic diagram of the method and hardware configuration of the present invention. This embodiment provides a laser target recognition method based on YOLO_V5, the hardware of which includes a laser emitting device, a chest ring target surface, a computer, and a camera.

[0031] The laser emitting device is used for:

[0032] Emitting infrared laser and sending transmission signal to computer at the same time.

[0033] The computer is used for:

[0034] The image obtained from the camera is analyzed and processed to obtain the predicted category, confidence level, and location of the frame.

[0035] Example 2.

[0036] like Figure 2 The following is a flowchart of a laser target recognition program. This invention uses the YOLOV5 algorithm to collect and identify the characteristics of the target surface, laser, and bull's-eye. Using Python to build a software interface, it can accurately identify the target surface, laser, and bull's-eye information in the image, and the final recognition results are fed back to the program interface.

[0037] First, complete the initialization of system parameters, and optimize the operating parameters of each module to make the system enter the best operating state; simultaneously complete the fine configuration of camera parameters, and ensure that the quality of the collected image meets the subsequent algorithm processing standards by adjusting parameters such as focal length, exposure, and resolution.

[0038] After receiving the shooting signal, the system immediately captures the shooting image and performs a series of preprocessing operations on the image: the image size is adjusted to 800x600 pixels and the image data is numerically normalized to improve the efficiency and accuracy of subsequent image recognition and processing.

[0039] The preprocessed image is then transferred to the pre-trained Target_Laser.pt model, which has been trained with a large amount of data and has the ability to accurately identify the target surface, laser and bull's eye features.

[0040] The Target_Laser.pt model uses a pre-trained convolutional neural network to extract features of the input image. These features will be used for subsequent target classification and coordinate regression tasks.

[0041] Based on the feature extraction results, the system uses a predefined priori frame to scan the feature map, and gradually performs target classification and coordinate regression on each priori frame to determine the position and size of the target surface, laser and bull's eye.

[0042] The system evaluates the prediction boxes through the loss function and uses the non-maximum suppression algorithm to filter and merge overlapping prediction boxes to eliminate redundant results and ensure that the final output prediction results are both accurate and concise.

[0043] After completing the image recognition task of the current frame, the system calculates the distance from the bull's eye to the laser and the distance from the bull's eye to the outermost ring of the target surface based on the positions of the target surface, laser and bull's eye obtained in step S5, and calculates the shooting result of this time.

[0044] Based on the image processed in step S2 and the position and size information obtained in step S5, the system crops the target shooting image and draws rectangular frames at the corresponding positions in the image to mark the target surface, laser and bull's eye. At the same time, the shooting score is displayed in the upper left corner of the image and saved.

[0045] Finally, the system outputs the detection results, which include the predicted category, confidence level, and box location. This information can be displayed to the user in an appropriate manner or used for subsequent data analysis and processing. When a new shooting signal is received, the system returns to step S2 for a new round of image processing.

[0046] Example 3.

[0047] like Figure 3 Shown is a flow chart of the training procedure of the method of the present invention.

[0048] When building the model, you need to collect the target shooting image dataset and train the pre-trained model yolov5n-7.pt.

[0049] Finally, the Target_Laser.pt model is obtained, which is sensitive to the target surface and laser recognition.

[0050] Output various parameter graphs representing model performance, such as confusion matrix, Precision growth curve, Recall growth curve, and F1Score curve.

[0051] Output the image to be verified from the verification set.

[0052] The verification model is the Target_Laser.pt model obtained after training.

[0053] Only by performing non-maximum threshold processing can the prediction boxes with higher consistency be screened out in one step.

[0054] Based on the prediction results, we get the post-processing bounding box coordinates, category labels, and confidence scores.

[0055] Then, for each picture, three different types of information can be obtained. After processing this information again, the confusion matrix, the growth curve of the precision rate, the growth curve of the recall rate and the F1Score curve are obtained.

[0056] Finally, by calculating these indicators, the final classification effect is obtained.

[0057] For examples 3 and 4, the growth curves of precision, recall and F1Score are supplemented

[0058] The confusion matrix comparison results include four cases: TP, FP, TN and FN.

[0059] TP: The predicted box satisfies IoU ≥ threshold and the category is correct.

[0060] FP: The predicted box does not meet IoU ≥ threshold, or the category is wrong.

[0061] FN: There is a real box, but there is no matching predicted box (i.e. missed detection).

[0062] TN: Less attention is paid to target detection (usually ignored) because negative samples are "background" and do not correspond to specific detection boxes.

[0063] Based on these four parameters, some performance evaluations of the trained model can be obtained.

[0064]

[0065] The above expression "Accuracy" represents the proportion of correct classification results to the total number of observations. It represents the probability that the model correctly predicts the objects labeled "Target," "Laser," and "Bull's-eye" in the training set. A higher accuracy indicates a better system.

[0066]

[0067] It indicates the proportion of samples that are actually positive among all samples predicted as positive by Precision. In other words, the proportion of samples that are actually Target among the samples predicted as Target by the system in this training is

[0068]

[0069] Recall represents the proportion of correctly predicted samples among all actually positive samples. In other words, the proportion of correctly detected samples among the real Target samples of the system in this training

[0070]

[0071] F1Score is the harmonic mean of precision and recall, and the curve shows its changing trend with the number of training rounds.

[0072] Because the original model pre-trained with the target shooting image dataset has better fit in target recognition, the F1Score of Target during training is always higher than that of Laser and bull's-eye, and the overall F1Score reaches 85%, indicating that the entire system has reached a high level.

[0073] The foregoing description is merely a preferred embodiment of the present invention. Furthermore, those skilled in the art will readily appreciate that various modifications and variations of the present invention may be made while maintaining the principles of the present invention. If such modifications and variations fall within the scope of the claims and their equivalents, such modifications and variations shall also be considered within the scope of protection of the present invention.

Claims

1. A laser target recognition method based on YOLO_V5, characterized in that: The following steps are involved: S1: First, complete the initialization of system parameters and optimize the operating parameters of each module to ensure that the system enters the optimal operating state; simultaneously complete the fine configuration of camera parameters and ensure that the captured image quality meets the subsequent algorithm processing standards by adjusting parameters such as focal length, exposure, and resolution. S2: After receiving the shooting signal, the system immediately captures the shooting image and performs a series of preprocessing operations on the image: the image size is adjusted to 800x600 pixels and the image data is numerically normalized to improve the efficiency and accuracy of subsequent image recognition and processing. S3: The pre-processed image is then transferred to the pre-trained Target_Laser.pt model. This model has been trained with a large amount of data and has the ability to accurately identify the target surface, laser, and bull's eye features. S4: The Target_Laser.pt model uses a pre-trained convolutional neural network to extract features from the input image. These features will be used for subsequent target classification and coordinate regression tasks. S5: Based on the feature extraction results, the system uses a predefined priori frame to scan the feature map, and gradually performs target classification and coordinate regression on each priori frame to determine the position and size of the target surface, laser and bull's eye. S6: The system evaluates the prediction boxes using a loss function and uses a non-maximum suppression algorithm to filter and merge overlapping prediction boxes to eliminate redundant results, ensuring that the final output prediction results are both accurate and concise. S7: After completing the image recognition task of the current frame, the system calculates the distance from the bull's eye to the laser and the distance from the bull's eye to the outermost ring of the target surface based on the positions of the target surface, laser and bull's eye obtained in step S5, and calculates the shooting result of this time. S8: Based on the image processed in step S2 and the position and size information obtained in step S5, the system crops the target surface shooting image and draws rectangular frames at the corresponding positions in the image to mark the target surface, laser and bull's eye. At the same time, the shooting score is displayed in the upper left corner of the image and saved. S9: Finally, the system outputs the detection results, including the predicted category, confidence level, and box location. This information is displayed to the user in an appropriate manner or used for subsequent data analysis. When a new shooting signal is received, the system returns to step S2 for a new round of image processing.

2. The method according to claim 1, characterized in that The camera in step S1 is used for real-time monitoring and transmits real-time data to the computer.

3. The method according to claim 1, characterized in that The Target_Laser.pt model in the S3 step requires dataset collection, category labeling, and model training.

4. The method according to claim 1, wherein The preprocessing operation of the Target_Laser.pt model training in the S3 step adopts Multi-Scale Augmentation. By dynamically adjusting the input resolution during the training process, the model learns the features of small objects at different scales to reduce the risk of overfitting to a single scale.

5. The method according to claim 1, wherein The loss function of S6 and the method of non-maximum suppression are: There are two bounding box areas, B is the predicted border area, and B gt is the actual border area, measuring B and B gt The distance, overlap rate and shape consistency between them are calculated by the following formula: Where v is the normalized difference between the aspect ratio of the predicted box and the true box, and α is a balance factor that weighs the loss caused by the aspect ratio and the loss caused by the IoU part.

Images