Double-moving-target defect detection method and system

Abstract

The invention discloses a double-motion target defect detection method and system, and the method comprises the steps: collecting a clear frame image under a double-motion condition through an industrial camera and a synchronous stroboscope lamp which are carried on a first motion carrier; inputting the image into a defect detection model based on deep learning for real-time reasoning to identify defects; a position reference is provided for the detection process based on a plurality of RFID tags arranged on a motion path, and starting and stopping of a task and interval management of a detection result are controlled according to the position reference; and finally, associating a defect result in each detection interval with geographic position information, and uploading the defect result to a server through a wireless network for displaying and recording. According to the method, the technical problems of unclear image acquisition, insufficient detection precision and difficulty in realizing reliable positioning and automatic closed-loop management in long-distance detection in a dynamic industrial scene are solved, and high-reliability and traceable automatic defect detection is realized.

Description

Technical Field

[0001] This application relates to the field of machine vision technology, and in particular to an intelligent defect detection method and system for dual moving targets. Background Technology

[0002] In industrial production lines, such as washing machine factories and air conditioner factories, intelligent production lines are usually built in industrial production sites to transport semi-finished or finished products more quickly and efficiently. These intelligent production lines involve all production floors and can reach several thousand to tens of thousands of meters in length.

[0003] Intelligent production lines typically consist of a track and trolleys suspended below it. The track is powered by a motor belt, which drives metal rollers in the middle and flexible rollers at bends to continuously move the trolleys. However, since the metal rollers in the middle cover almost the entire production line and rotate at high speed continuously, wear and tear usually occurs over time, leading to defects and potentially serious safety accidents. Therefore, it is necessary to continuously inspect the high-speed rotating metal rollers in the middle of the entire production line for defects and replace them in a timely manner.

[0004] However, intelligent production lines may involve multiple floors and distances ranging from thousands to tens of thousands of meters. Workers need to constantly inspect the entire production line using lighting equipment to detect defects. However, firstly, industrial sites generally have high temperatures, high noise, high pressure, and high radiation. Workers exposed to such harsh environments for extended periods will suffer significant health damage. Furthermore, manual defect detection methods are susceptible to subjective human factors and face the challenge of continuous, 24 / 7 inspections. Secondly, with the continuous increase in labor costs, more workers are needed to maximize inspections, which significantly increases the company's operating costs. Finally, due to the high-speed rotation of the metal rollers in the middle of the track, even if defects are present, they are sometimes difficult for the human eye to see clearly.

[0005] In recent years, with the rapid development of computer technology, deep learning technology, real-time video stream inspection technology, and 5G technology, and with the rapid iteration and cost reduction of various hardware technologies, the difficulties of manual inspection can be overcome to some extent. However, through technological iteration and long-term observation and analysis of real-world environments, the inventors have recognized that existing technical solutions still face a series of systemic challenges in long-distance, complex defect detection applications on industrial production lines. First, in dynamic inspection scenarios, especially when both the inspection platform and the target are in relative motion, conventional imaging methods struggle to capture stable and clear images, affecting the reliability of subsequent analysis. Second, image analysis methods relying on traditional rules or fixed thresholds often lack sufficient accuracy and adaptability in industrial environments with varying lighting and complex backgrounds, easily leading to missed or false positives for minor defects or inconspicuous textures. Furthermore, to achieve coverage and positioning for long-distance inspection, existing solutions still need improvement in areas such as the reliability of positioning beacons, automated management of inspection tasks, and real-time processing and structured feedback of massive image data, making it difficult to support efficient, accurate, and traceable closed-loop quality control. Summary of the Invention

[0006] This application provides a method and system for detecting defects in dual moving targets, aiming to solve the problems of unclear image acquisition, insufficient detection accuracy and adaptability, and difficulty in achieving reliable positioning and automated closed-loop management for long-distance detection in dynamic and complex industrial scenarios.

[0007] Firstly, a method for detecting defects in dual moving targets is provided, including:

[0008] S1, during the process of the industrial camera mounted on the first moving vehicle acquiring video streams of the second moving target, the strobe light linked to the industrial camera is simultaneously triggered to provide supplementary lighting in order to obtain clear frame images;

[0009] S2, the clear frame image is input into the defect detection model trained based on deep learning for real-time inference to identify defects on the second moving target;

[0010] S3, a position reference is provided for the detection process by multiple radio frequency identification (RFID) tags arranged on the motion path of the first moving vehicle;

[0011] S4, based on the reading of the RFID tag, control the start and stop of the defect detection task and the interval management of the detection results;

[0012] S5 associates the defect detection results within each detection range with the corresponding geographical location information and uploads them to the server via wireless network for display and recording.

[0013] Optionally, in the above scheme, step S4 includes:

[0014] When the initial RFID tag is read, the real-time inference of the defect detection model is initiated;

[0015] When the first moving vehicle moves within the detection range defined by two adjacent RFID tags, if a defect is detected, the defect result and corresponding image within that range are uploaded; if no defect is detected, any normal image within that range and the result of no defect are uploaded.

[0016] When a termination RFID tag is read, the real-time inference of the defect detection model is stopped.

[0017] In the above scheme, optionally, in step S4, the location information of the starting RFID tag and the ending RFID tag, as well as the start and stop instructions of the detection task, are transmitted between the software control terminal and the computing device performing inference through the Message Queuing Telemetry Transmission (MQTT) protocol.

[0018] Optionally, in the above scheme, the defect detection model is a single-stage target detection model based on the YOLO-V10 architecture.

[0019] Optionally, the above scheme may further employ a multi-threaded processing mechanism to perform the real-time inference, including: a first thread continuously extracting frame images from the video stream and storing them in a queue, and a second thread continuously obtaining frame images from the queue and calling the defect detection model for inference.

[0020] In the above scheme, optionally, after the second thread continuously obtains frame images from the queue and calls the defect detection model for inference, a defect region screening step is further included: only when the detected defect region is within a preset image horizontal range is it determined to be a valid defect.

[0021] Optionally, in the above scheme, the uploaded detection results and image data paths are encapsulated in JSON format.

[0022] Secondly, a dual-moving-target defect detection system is provided for implementing the above-described method, the system comprising:

[0023] The image acquisition module includes an industrial camera mounted on a first moving vehicle and a strobe light that is triggered synchronously with the industrial camera, for acquiring clear frame images of a second moving target;

[0024] The computing module is equipped with a defect detection model trained based on deep learning, which is used to perform real-time reasoning on the clear frame image to identify defects.

[0025] The positioning and triggering module includes multiple RFID tags arranged on the movement path of the first moving vehicle, a reader for reading tag information, and an interface for communicating with the software control terminal, which is used to provide a position reference for detection and realize the interval control of the detection task.

[0026] The communication and display module is used to upload the detection results and images associated with location information to the server for display and storage via a wireless network.

[0027] Optionally, in the above scheme, the computing module adopts a multi-threaded architecture and utilizes a graphics processing unit (GPU) to accelerate the inference process of the defect detection model.

[0028] Optionally, in the above scheme, the positioning and triggering module receives task start / stop commands from the software control terminal via the MQTT protocol and feeds back the location information read by the RFID tag to the software control terminal.

[0029] Compared with the prior art, this application has at least the following beneficial effects:

[0030] Based on further analysis and research of existing technologies, this application recognizes that existing technologies suffer from problems such as unclear image acquisition, insufficient detection accuracy and adaptability in dynamic and complex industrial scenarios, and difficulty in achieving reliable positioning and automated closed-loop management for long-distance detection. Therefore, this application proposes a systematic solution: by deploying an industrial camera and synchronous strobe illumination on the first moving vehicle, image blurring caused by relative motion is overcome from the imaging source, ensuring input quality; a deep learning-based defect detection model is introduced to replace traditional rule-based methods, improving robustness and accuracy in identifying complex lighting and minute defects; furthermore, RFID tags are deployed along the movement path as reliable spatial references, and a task start / stop and interval-based result management mechanism is constructed based on these tags to achieve automatic positioning, segmented traceability, and closed-loop control during long-distance detection. The synergistic effect of these components achieves stable and clear imaging under continuous motion and changing environmental conditions, high-precision real-time defect identification, and fully traceable automated detection, ultimately achieving comprehensive technical effects such as improved detection reliability, enhanced system adaptability, closed-loop operation and maintenance, and reduced reliance on manual labor.

[0031] This application also has the following advantages based on the technical design of RFID positioning and MQTT communication linkage:

[0032] 1. Achieve blind-spot-free and highly reliable spatiotemporal synchronization: By continuously deploying RFID tag groups along the path, an absolute position reference is provided for each detection interval, solving the problem of position loss or confusion caused by visual tag recognition failure or cumulative error in long-distance continuous detection. This ensures that each defect can be bound to a unique and accurate geographical location, achieving precise alignment and full path coverage of detection results in the temporal and spatial dimensions.

[0033] 2. Constructing an adaptive, lightweight task management closed loop: Utilizing the MQTT protocol to achieve low-latency, reliable command and status synchronization between the software control terminal and the computing device. The system can automatically and accurately start and stop detection tasks based on real-time location events triggered by RFID, and autonomously manage the detection cycle and result reporting according to preset intervals. This not only significantly reduces the need for manual intervention and the complexity of operation and maintenance, but also enables the system to flexibly adapt to pipeline layouts of different lengths and segment divisions, improving the versatility and scalability of deployment.

[0034] 3. Supports fine-grained, structured quality traceability: Using RFID-defined physical intervals as the smallest management unit for result collection and reporting enables the structuring and segmentation of massive amounts of inspection data. This not only facilitates efficient storage, querying, and visualization on the server side, but more importantly, provides a data foundation with precise spatial context for subsequent defect statistical analysis, trend prediction, and maintenance decisions, achieving end-to-end quality traceability capabilities from "inspection" to "analysis-decision". Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a dual-moving-target defect detection method provided in one embodiment of this application.

[0036] Figure 2 This is a schematic diagram of the overall process of a dual-moving-target defect detection method provided in one embodiment of this application.

[0037] Figure 3 This is an example image of the defect image detected in steps 2-4 of one embodiment of this application, along with the left width threshold (Left_Width_Threshold) and right width threshold (Right_Width_Threshold). Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0039] In the description of this application: unless otherwise stated, "a plurality of" means two or more. The terms "first," "second," "third," etc., in this application are intended to distinguish the objects referred to and do not have any special meaning in terms of technical connotation (e.g., they should not be construed as an emphasis on importance or order). Expressions such as "including," "comprising," and "having" also mean "not limited to" (certain units, components, materials, steps, etc.).

[0040] In one embodiment, combined Figure 1 A method for detecting defects in dual moving targets is provided, including:

[0041] S1, during the process of the industrial camera mounted on the first moving vehicle acquiring video streams of the second moving target, the strobe light linked to the industrial camera is simultaneously triggered to provide supplementary lighting in order to obtain clear frame images;

[0042] S2, the clear frame image is input into the defect detection model trained based on deep learning for real-time inference to identify defects on the second moving target;

[0043] S3, a position reference is provided for the detection process by multiple radio frequency identification (RFID) tags arranged on the motion path of the first moving vehicle;

[0044] S4, based on the reading of the RFID tag, control the start and stop of the defect detection task and the interval management of the detection results;

[0045] S5 associates the defect detection results within each detection range with the corresponding geographical location information and uploads them to the server via wireless network for display and recording.

[0046] In this embodiment, the terms "first moving vehicle" and "second moving target" are not simply sequential designations, but rather serve to clearly define two physical entities in relative motion with distinct functional roles within the technical solution. The "first moving vehicle" refers to a platform carrier that carries the detection equipment (such as an industrial camera and strobe light) and can move along a specific path; its core function is to move and support the detection equipment. The "second moving target" refers to the object or component to be observed or detected, which is itself in motion. The distinction between "vehicle" and "target" emphasizes their active and observed relationship in the detection task, avoiding role confusion caused by both being referred to as "targets."

[0047] In this application, "dual moving target detection" specifically refers to a complex scenario where both the detection platform (first moving vehicle) and the object being detected (second moving target) are in motion during the operation of the detection system. This term emphasizes the dynamic conditions encountered during the detection process, rather than referring to the detection of two independent moving targets. Therefore, "dual motion" is essentially "a dual relative motion between the detection platform and the object being detected," which constitutes the key technical challenge this application aims to address—namely, how to achieve clear imaging, accurate identification, and reliable positioning under dual motion conditions.

[0048] In this embodiment, the method systematically constructs a defect detection process suitable for dynamic industrial scenarios through five core steps (S1-S5), covering the entire process from clear image acquisition, intelligent defect recognition, precise positioning benchmark setting, automated task control to structured result reporting.

[0049] The main methods for detecting defects in dual moving targets on industrial production lines include: acquiring high-definition visible light images or videos with a camera, detecting defects across the entire production line, determining the location of defects, and displaying and saving defect data.

[0050] High-definition visible light images or videos are captured by cameras, including combinations of ordinary visible light cameras and supplementary lighting, and combinations of industrial cameras and strobe lights; full-coverage production line defect detection includes defect detection methods based on digital image processing and defect detection methods based on deep learning; defect location determination includes affixing numbers or Chinese / English labels at fixed positions on the track and determining the location by recognizing the numbers or Chinese / English labels, and placing RFID (Radio Frequency Identification) cards at fixed positions on the track and determining the location information by reading the RFID card information; defect data display and storage includes storing data on local hard drives and transmitting data to a server for intuitive display via wireless networks such as 5G networks.

[0051] For dual-motion defect detection where both the trolley and the intermediate metal rollers are in motion, conventional visible light cameras and supplementary lights typically produce blurry videos that are unsuitable for defect detection. This application employs a high-definition industrial camera and strobe light solution, ensuring that every frame acquired is clear and meets the requirements for subsequent defect detection.

[0052] For defect detection solutions, since some defects are small and the lighting conditions in industrial environments are complex, defect detection using digital image processing cannot guarantee detection accuracy after being affected by lighting. This application adopts a target detection solution based on deep learning for defect detection.

[0053] To determine the location of defects, the current method of attaching numbers or Chinese / English labels at fixed positions on the track and identifying the location by recognizing the numbers or labels is easily affected by complex lighting and shooting angles, and may fail to detect defects. This application adopts a method of placing RFID cards at fixed positions on the track and determining the location information by reading the information on the RFID cards, which ensures the accurate reading of the location information.

[0054] In summary, this application discloses a solution that employs an industrial camera and strobe light to ensure clear video capture; utilizes a real-time video stream based on deep learning to ensure accurate real-time detection throughout the entire process; and employs a solution that uses RFID cards placed at fixed positions on a track to determine location information by reading the RFID card information, thus ensuring accurate reading of location information. Together, these solutions guarantee full coverage, high precision, and accurate positioning for defect detection of dual moving targets in complex industrial lighting environments.

[0055] In one embodiment, step S4 includes:

[0056] When the initial RFID tag is read, the real-time inference of the defect detection model is initiated;

[0057] When the first moving vehicle moves within the detection range defined by two adjacent RFID tags, if a defect is detected, the defect result and corresponding image within that range are uploaded; if no defect is detected, any normal image within that range and the result of no defect are uploaded.

[0058] When a termination RFID tag is read, the real-time inference of the defect detection model is stopped.

[0059] In one embodiment, the RFID tag-based task control mechanism defined in step S4 embodies the design of this method for structured and refined management of the detection process in long-distance, continuous detection scenarios. By deploying start, intermediate, and end RFID tags, the system can divide the continuous detection path into multiple manageable logical intervals. This design not only enables automatic start and stop of detection tasks but, more importantly, establishes a strict correspondence between location and detection data, making each detection result traceable to a specific physical segment. This allows for rapid location and verification of defects during subsequent maintenance, improving the accuracy and efficiency of operational response.

[0060] In one embodiment, in step S4, the location information of the starting and ending RFID tags, as well as the start and stop instructions for the detection task, are transmitted between the software control terminal and the computing device performing inference via the Message Queuing Telemetry Transmission (MQTT) protocol.

[0061] Furthermore, in this embodiment, the Message Queuing Telemetry (MQTT) protocol is chosen as the communication method between the software control terminal and the computing device because of its lightweight, asynchronous nature and support for publish / subscribe patterns, making it particularly suitable for low-bandwidth, high-latency, or intermittently connected industrial wireless networks in mobile environments. Transmitting RFID tag information and task instructions via the MQTT protocol enables reliable and real-time transmission of control signals and reduces system coupling. This allows for decoupling and collaboration of the three key components: software scheduling, location identification, and real-time inference, enhancing the system's robustness and scalability in complex industrial communication environments.

[0062] In this embodiment, the combined use of RFID and MQTT is not a simple technology stacking, but rather a lightweight, traceable, loosely coupled closed-loop task control system suitable for long-distance mobile detection scenarios. It solves the problems of inaccurate positioning, rough task execution, and data and location disconnection in traditional detection methods from the perspective of process management.

[0063] In one embodiment, the defect detection model is a single-stage target detection model based on the YOLO-V10 architecture.

[0064] In this embodiment, the choice is primarily based on the excellent performance balance demonstrated by the YOLO series models in real-time object detection tasks, especially the further optimization of YOLO-V10 in terms of detection speed and accuracy. Compared to two-stage detection models, single-stage models do not require a region proposal network, have shorter inference paths, and are more suitable for continuous video stream processing in industrial scenarios with high real-time requirements. Furthermore, YOLO-V10 itself possesses strong multi-scale feature fusion and a lightweight design, which facilitates deployment on computing devices, ensuring both detection accuracy and real-time response requirements.

[0065] In one embodiment, the real-time inference is performed using a multi-threaded processing mechanism, including: a first thread continuously extracting frame images from the video stream and storing them in a queue, and a second thread continuously obtaining frame images from the queue and calling the defect detection model for inference.

[0066] In this embodiment, the multi-threaded processing mechanism is a decoupled and buffered architecture designed to address the differences in time consumption and stability between the two major tasks of "acquisition" and "inference" in industrial real-time video stream processing. Specifically, the first thread is dedicated to the continuous and stable acquisition of the video stream, and its core task is to ensure that no frame images are lost due to fluctuations in subsequent processing speed; the second thread focuses on computationally intensive model inference, and its execution time may fluctuate due to factors such as image content and model load. By using a shared queue as an asynchronous buffer channel, this mechanism decouples the two threads, ensuring that high-speed and stable frame acquisition is not blocked by the relatively slow or fluctuating inference process, while also ensuring that the inference thread always has image data to process, thereby achieving stable pipelined processing and higher system throughput overall.

[0067] In one embodiment, after the second thread continuously retrieves frame images from the queue and calls the defect detection model for inference, a defect region filtering step is further included: a defect region is only determined to be a valid defect if the detected defect region is within a preset image horizontal range.

[0068] In one embodiment, the defect region screening step introduced after the second thread completes model inference is a spatial prior application and result optimization strategy tailored to a specific industrial inspection scenario. This step is based on a key scenario prior knowledge: in applications such as assembly line roller inspection, the target object being inspected (the second moving target) typically presents itself in the image as a feature occupying a specific horizontal region (e.g., a strip-shaped region located in the middle of the image). Therefore, this step constructs a preset effective horizontal region of the image by setting left and right horizontal width thresholds.

[0069] Its operating mechanism is as follows: For each raw detection result (usually represented by a bounding box) output by the model inference, its horizontal coordinate range is calculated, and it is determined whether it falls completely within the preset valid horizontal range. Only when it falls completely within the range is the detection result judged as a valid defect and retained; otherwise, it is discarded as an invalid detection (such as background interference or false alarms in non-target areas). By directly filtering out false alarms that are unlikely to belong to the target area being tested, the overall false alarm rate of the system is significantly reduced, making the reported defect results more business-specific and credible.

[0070] In one embodiment, the uploaded detection results and image data paths are encapsulated in JSON format.

[0071] JSON, as an open standard and lightweight data exchange format, supports nested key-value pair structures. This facilitates the encapsulation of heterogeneous information from multiple sources, such as detection results (e.g., defect type, confidence level, coordinates), image metadata (e.g., frame ID, timestamp), location information (RFID tag ID, interval number), and possible device status, into a single structured data packet, improving data readability and parsing. Furthermore, JSON is a text format, natively or through libraries supported by the vast majority of programming languages ​​and platforms, reducing the complexity of system integration and expansion.

[0072] In one embodiment, combined Figure 2 The complete description of the solution in this application is as follows: This application provides a real-time video stream-based method for detecting defects in two moving targets based on deep learning. It employs an industrial camera and strobe light to ensure clear video capture; it utilizes a real-time video stream based on deep learning to ensure accurate detection throughout the entire process; and it uses a scheme that places RFID cards at fixed positions on a track and determines the location information by reading the RFID card information, ensuring accurate location reading. Together, these methods guarantee full coverage, high precision, and accurate positioning for detecting defects in two moving targets in complex industrial lighting environments. Specifically, the method includes the following steps:

[0073] Step 1, using a resolution of Industrial cameras capture real-time video of the high-speed rotating metal rollers on the entire production line and extract defective frames to facilitate the training of defect detection algorithm models.

[0074] Step 2, the implementation of the defect detection algorithm based on deep learning, is as follows:

[0075] Step 2-1, Sample image annotation: Labelme was used to annotate the defect samples of the transmission metal roller shaft, and a total of 1,000 sample images were annotated.

[0076] Step 2-2, Model Training: This application adopts a target detection model based on the YOLO-V10 architecture. This model is a single-stage target detection algorithm that can simultaneously complete target localization and classification through a single neural network inference. The pre-trained model uses yolov10m.pt, which is trained on large general datasets such as COCO and is a medium-sized target detection model that balances speed and accuracy. Efficient input detection is achieved through optimized feature extraction backbone networks (C2f, SCDown, SPPF, PSA) and detection heads (multi-scale fusion, v10Detect). In the backbone part, the C2f module enhances feature representation through multi-level feature fusion, the SCDown module implements feature downsampling and channel expansion, the SPPF module performs multi-scale feature aggregation, and the PSA module enhances the focus of key features, finally outputting feature maps of different scales (P3 / P4 / P5). In the head part, upsampling (nn.Upsample), feature concatenation (Concat), and the C2f / C2fCIB module are used to construct a feature pyramid structure to fuse and process the multi-scale features output by the backbone. Finally, the v10Detect module realizes the detection of targets at different scales.

[0077] In this application, the model training parameters are set such that the labeled samples are split into training, validation, and test datasets in a 7:2:1 ratio, and the image size is set accordingly. The number of epochs was set to 500, the learning rate to 0.01, the batch size to 3, and the graphics card to be used was an NVIDIA GeForce RTX2080 SUPER. After training, the model achieved an optimal mean accuracy (mAP) of 99.24%, and this model was exported as a defect inference model.

[0078] Steps 2-3 involve engineering application model inference. A combination of C++ multithreading and OpenCV (Open Source Computer Vision Library) is used. Two threads are created: one continuously extracts real-time video frames from the RTSP (Real-Time Streaming Protocol Uniform Resource Locator) video stream and stores them in a queue; the other continuously retrieves frames from the queue and calls the trained model for defect detection inference. This combination of CPU and GPU ensures that the inference time for one frame is approximately 50ms, and the real-time video stream has a frame rate of 10 frames per second, guaranteeing that each captured image can be fully covered for real-time detection.

[0079] Steps 2-4 involve defect detection and screening. Since defect detection is performed on the entire image... The image undergoes full-image detection. However, the high-speed rotating rollers in the industrial setting are mostly located in the center of the image. Therefore, left width thresholds (Left_Width_Threshold) and right width thresholds (Right_Width_Threshold) are set. The x-coordinate of the top-left corner of the detected bounding box is considered as a threshold. And the upper right x-coordinate If the defect is found to be valid, retain it; otherwise, discard it. (Reference) Figure 3 .

[0080] Step 3: Place RFID cards at fixed positions along the entire assembly line track, including the starting position. and termination position and other positions in the middle (i=1,2,3,4……N, N is the number of fixed locations determined in the industrial site).

[0081] Step 4: The software sends an inspection task to detect defects in the metal rollers of the entire production line.

[0082] Step 4-1: At the starting point of the cargo trolley, the software sends an inspection task for detecting defects in the metal rollers of the entire production line, and starts the defect detection algorithm of the real-time video stream. The software detects the starting position of the RFID card. At that time, location information is sent to the algorithm via MQTT, and the algorithm begins defect detection algorithm reasoning.

[0083] Step 4-2: If a defect is detected between two adjacent RFID cards, regardless of the number, the detection result and the corresponding image are returned to the software for display. If no defect is detected between two adjacent RFID cards, any normal image frame and the result are returned to the software for display. The return method is to return the result in JSON format via POST.

[0084] Step 4-3, the software detects the termination position of the RFID card. At that time, the location information is sent to the algorithm via MQTT, and the algorithm ends the defect detection algorithm reasoning.

[0085] Step 5: The detection results and images are returned to the server webpage for intuitive display via wireless network.

[0086] This application employs a combination of industrial cameras and strobe lights to ensure clear capture of dual moving targets. It utilizes a YOLOv10-based deep learning target detection algorithm, multi-threaded real-time video streaming, RFID positioning, and MQTT integration to guarantee full coverage, high precision, and accurate positioning of targets throughout the entire production line. This effectively solves the problem of clearly capturing dual moving targets in actual industrial production line applications, addresses the issues of full coverage and positioning of moving targets throughout the entire production line, and effectively solves the real-time detection problem, ensuring full coverage, high precision, and accurate positioning of defect detection for dual moving targets in complex industrial lighting environments.

[0087] In one embodiment, a dual-moving-target defect detection system is provided to implement the dual-moving-target defect detection method of the above embodiment, the system comprising:

[0088] The image acquisition module includes an industrial camera mounted on a first moving vehicle and a strobe light that is triggered synchronously with the industrial camera, for acquiring clear frame images of a second moving target;

[0089] The computing module is equipped with a defect detection model trained based on deep learning, which is used to perform real-time reasoning on the clear frame image to identify defects.

[0090] The positioning and triggering module includes multiple RFID tags arranged on the movement path of the first moving vehicle, a reader for reading tag information, and an interface for communicating with the software control terminal, which is used to provide a position reference for detection and realize the interval control of the detection task.

[0091] The communication and display module is used to upload the detection results and images associated with location information to the server for display and storage via a wireless network.

[0092] In one embodiment, the computing module employs a multi-threaded architecture and utilizes a graphics processing unit (GPU) to accelerate the inference process of the defect detection model.

[0093] In one embodiment, the positioning and triggering module receives task start / stop commands from the software control terminal via the MQTT protocol and feeds back the location information read by the RFID tag to the software control terminal.

[0094] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for detecting defects in dual moving targets, characterized in that, include: S1, during the process of the industrial camera mounted on the first moving vehicle acquiring video streams of the second moving target, the strobe light linked to the industrial camera is simultaneously triggered to provide supplementary lighting in order to obtain clear frame images; S2, the clear frame image is input into the defect detection model trained based on deep learning for real-time inference to identify defects on the second moving target; S3, a position reference is provided for the detection process by multiple radio frequency identification (RFID) tags arranged on the motion path of the first moving vehicle; S4, based on the reading of the RFID tag, control the start and stop of the defect detection task and the interval management of the detection results; S5 associates the defect detection results within each detection range with the corresponding geographical location information and uploads them to the server via wireless network for display and recording.

2. The method according to claim 1, characterized in that, Step S4 includes: When the initial RFID tag is read, the real-time inference of the defect detection model is initiated; When the first moving vehicle moves within the detection range defined by two adjacent RFID tags, if a defect is detected, the defect result and corresponding image within that range are uploaded; if no defect is detected, any normal image within that range and the result of no defect are uploaded. When a termination RFID tag is read, the real-time inference of the defect detection model is stopped.

3. The method according to claim 2, characterized in that, In step S4, the location information of the starting and ending RFID tags, as well as the start and stop instructions for the detection task, are transmitted between the software control terminal and the computing device executing inference via the Message Queuing Telemetry Transmission (MQTT) protocol.

4. The method according to claim 1, characterized in that, The defect detection model is a single-stage target detection model based on the YOLO-V10 architecture.

5. The method according to claim 4, characterized in that, The real-time inference is executed using a multi-threaded processing mechanism, including: a first thread continuously extracting frame images from the video stream and storing them in a queue, and a second thread continuously obtaining frame images from the queue and calling the defect detection model for inference.

6. The method according to claim 5, characterized in that, After the second thread continuously retrieves frame images from the queue and calls the defect detection model for inference, a defect region filtering step is also included: only when the detected defect region is within a preset image horizontal range is it determined to be a valid defect.

7. The method according to claim 2, characterized in that, The uploaded detection results and image data paths are encapsulated in JSON format.

8. A dual-moving-target defect detection system, characterized in that, The system for implementing the method as described in any one of claims 1 to 7 comprises: The image acquisition module includes an industrial camera mounted on a first moving vehicle and a strobe light that is triggered synchronously with the industrial camera, for acquiring clear frame images of a second moving target; The computing module is equipped with a defect detection model trained based on deep learning, which is used to perform real-time reasoning on the clear frame image to identify defects. The positioning and triggering module includes multiple RFID tags arranged on the movement path of the first moving vehicle, a reader for reading tag information, and an interface for communicating with the software control terminal, which is used to provide a position reference for detection and realize the interval control of the detection task. The communication and display module is used to upload the detection results and images associated with location information to the server for display and storage via a wireless network.

9. The system according to claim 8, characterized in that, The computing module adopts a multi-threaded architecture and utilizes a graphics processing unit (GPU) to accelerate the inference process of the defect detection model.

10. The system according to claim 8, characterized in that, The positioning and triggering module receives task start / stop commands from the software control terminal via the MQTT protocol and feeds back the location information read by the RFID tag to the software control terminal.

Images