Intelligent laboratory shank exposure detection system and method based on YOLOv8 and multi-modal judgment

Abstract

The invention provides an intelligent laboratory shank exposure detection system and method based on YOLOv8 and multi-modal determination, and the method comprises the steps: collecting a laboratory video stream through a multi-path camera, inputting a specially trained YOLOv8 model after image preprocessing, and synchronously detecting the lower limb and exposed shank regions of a human body; an area proportion calculation algorithm is innovatively adopted, the proportion r = B / A of the area of the bare shank area to the area of the lower limb area is calculated and compared with a preset threshold value T, and violation is judged; meanwhile, a multi-frame tracking consistency judgment mechanism is introduced, and instantaneous false alarms are eliminated by analyzing the violation proportion p = k / N of continuous N frames. The system comprises a video acquisition module, an image preprocessing module, a deep learning detection module, a judgment alarm module and a data management module. According to the method, the deep learning technology and the safety management requirement are deeply fused, the technical problems that traditional manual supervision is low in efficiency and different in standard are solved, and accurate, real-time and automatic monitoring of laboratory shank exposure behaviors is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and intelligent monitoring, and particularly relates to a smart laboratory small leg exposure detection system and method based on YOLOv8 and multi-modal judgment. BACKGROUND

[0002] Laboratories are important places for scientific research and teaching experiments, and standard dress is one of the basic requirements for ensuring experimental safety (such as preventing chemical splashing and high-temperature burns). At present, laboratory safety management mainly relies on manual patrol or ordinary video monitoring, which has the following defects: Low efficiency and high cost: manual patrol cannot achieve all-weather and dead-angle-free monitoring, and consumes a large amount of human resources.

[0003] Strong subjectivity and inconsistent standards: different managers may have different standards for judging "dressing standards", which is difficult to unify.

[0004] Post-tracing and lack of early warning: ordinary monitoring systems can only be used for post-tracing and cannot provide real-time early warning and intervention when violations occur.

[0005] Technical challenges: the small leg exposure area is a small target in the entire image, and in complex lighting, multiple poses, and occlusion environments, traditional visual algorithms have low detection accuracy and high false positive rates.

[0006] Although existing general target detection models (such as the YOLO series) have strong performance, direct application to the detection task of small leg exposure in the specific laboratory scene, which is a fine-grained and small target, has problems such as model inadaptability and missing judgment logic, and cannot directly meet the actual management needs. SUMMARY

[0007] The present application relates to the technical field of computer vision and intelligent monitoring, and particularly relates to a smart laboratory small leg exposure detection system and method based on YOLOv8 and multi-modal judgment.

[0008] To achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: A smart laboratory small leg exposure detection method based on YOLOv8 and multi-modal judgment, comprising the following steps: S1: Real-time acquisition of laboratory video streams by deploying multiple high-definition cameras in the laboratory, and extraction of video frames; Real-time acquisition of laboratory video streams by deploying multiple high-definition cameras in the laboratory at a frame rate of not less than 15fps, covering each experimental bench, passageway and public area, to ensure that the monitoring has no dead angle.

[0009] S2: image pre-processing and enhancement on the extracted video frames, the enhancement including geometric enhancement and color space enhancement; Geometric enhancement: including random crop (crop ratio 0.8-1.0), random rotation (±15°) and affine transformation; Color enhancement: hue perturbation (±10%), saturation and lightness perturbation (±30%) in HSV space; Normalization: scaling pixel values to the range of 0-1.

[0010] S3: input the pre-processed images into a pre-trained YOLOv8 deep learning detection model, which simultaneously outputs human lower limb detection boxes and exposed calf detection boxes, as well as their respective confidence levels; S4: based on the lower limb detection boxes and exposed calf detection boxes, execute a calf exposure violation judgment algorithm: calculate the ratio r of the exposed calf area B to its corresponding lower limb area A, i.e. r=B / A; if r is greater than a preset threshold T, it is determined to be a violation; the threshold T has a value range of 20% to 40%; S5: when a violation is determined, trigger an audible and light alarm mechanism, and upload relevant data of the violation event, including the violation frame image, timestamp, camera number and exposure ratio R, to a management platform.

[0011] Further, the judgment algorithm in step S4 further includes multi-frame tracking consistency judgment: For a continuously tracked target individual, let its judgment results in consecutive N frames be the sequence {r i ,i=1,2,...,N}; Calculate the proportion p of the frame number k of r i in the sequence to the total frame number N, i.e. p=k / N; If p is greater than a consistency threshold p t , output the final violation judgment; Wherein, the value range of the tracking frame number N is 5 to 15, and the value range of the consistency threshold p t is 60% to 80%.

[0012] Further, the YOLOv8 deep learning detection model in step S3 is obtained by training the following method: Construct a dedicated training data set, the image annotations in the data set include two types of boundary boxes "lower limbs" and "exposed calves"; Use a data enhancement strategy optimized for small targets to enhance the training data, the strategy including Mosaic enhancement and copy_paste; The YOLOv8 network was trained using the dataset described above. Its loss function L is a composite loss function, consisting of bounding box regression loss L1, classification loss L2, and confidence loss L3 weighted by coefficients. The specific formula is as follows: L = w1·L1 + w2·L2 + w3·L3 The weight coefficients w1, w2, and w3 have the following ranges: w1∈[4.0,6.0], w2∈[0.3,0.7], and w3∈[0.3,0.7].

[0013] Furthermore, the bounding box regression loss L1 adopts the GIoU loss, which is calculated as follows: Let the predicted bounding box be P, the ground truth bounding box be G, and the smallest closed box containing both P and G be C. L1 = 1 - GIoU, Among them, GIoU=IoU-|C-(P∪G)| / |C|, IoU=|P∩G| / |P∪G|; The predicted box P and the ground truth box G are bounding boxes, respectively represented as: P=(x 1p ,y 1p ,x 2p ,y 2p ) G=(x 1g ,y 1g ,x 2g ,y 2g ) The coordinates of the smallest closed bounding box C are calculated as follows: x 1c =min(x 1p ,x 1g ), y 1c =min(y 1p ,y 1g ) x 2c =max(x 2p ,x 2g ), y 2c =max(y 2p ,y 2g ).

[0014] Furthermore, in step S4, areas A and B are respectively detected by the lower limb detection frame P. leg And the detection frame for bare calves P bare The calculation yielded: A=area(P leg )=(x 2leg -x 1leg )×(y 2leg -y 1leg ) B = area(P) bare )=(x 2bare -x 1bare )×(y 2bare -y 1bare ).

[0015] Among them, P leg and P bare These are all predicted bounding boxes output by the model.

[0016] An automated detection system for exposed lower legs in a smart laboratory includes: The video acquisition module is used to acquire panoramic video data of the laboratory through multiple high-definition cameras; The image preprocessing module is used to normalize, geometrically enhance, and enhance the color space of video frames. The deep learning detection module has a built-in YOLOv8 model trained by the method described in claim 3 or 4, which is used to identify the lower limbs and bare lower leg areas. The judgment and alarm module is used to perform area ratio calculation r=B / A, multi-frame tracking consistency judgment p=k / N, and trigger alarms; The data management and feedback module is used to store violation records, generate statistical reports, and support incremental learning of the model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention adopts a specially optimized YOLOv8 model, which can achieve an accuracy of over 95% for small target detection, and supports real-time processing of multiple video streams with rapid response.

[0018] This invention effectively filters out transient interference through multi-frame consistency determination, supports incremental learning, and adapts to environmental changes. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the steps and flow structure of a smart laboratory method for detecting exposed lower legs based on YOLOv8 and multimodal determination according to the present invention. Figure 2 This is a structural diagram of the framework of an automatic detection system for exposed lower legs in a smart laboratory, as described in this invention. Detailed Implementation

[0020] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0021] A smart laboratory method for detecting exposed lower legs based on YOLOv8 and multimodal decision-making includes the following steps: S1: Real-time acquisition of video streams within the laboratory and extraction of video frames via multiple high-definition cameras deployed in the laboratory; Multiple high-definition cameras deployed within the laboratory capture real-time video streams at a frame rate of at least 15fps, covering all workbenches, corridors, and public areas to ensure comprehensive monitoring. The multiple cameras provide panoramic monitoring of the laboratory, eliminating blind spots inherent in a single camera. The frame rate of at least 15fps captures continuous movements such as walking and turning, providing a smooth temporal data foundation for subsequent multi-frame tracking and preventing target loss or disjointed movements due to low frame rates. Furthermore, it provides a high-quality, continuous real-world data source for subsequent processing and model training.

[0022] S2: Perform image preprocessing and enhancement on the extracted video frames, including geometric enhancement and color space enhancement; Geometric augmentation includes random cropping (cropping ratio 0.8-1.0), random rotation (±15°), and affine transformation. This method simulates the imaging effect under different camera angles and different human postures (standing, walking, squatting), greatly improving the model's adaptability to different viewing angles and changes in human posture, and preventing the model from overfitting to a specific posture.

[0023] Color enhancement: Hue perturbation (±10%), saturation, and brightness perturbation (±30%) are performed in the HSV color space. This method simulates different color temperature lighting conditions (such as fluorescent lights and natural light), shadows, and overexposure that may occur in a laboratory by adjusting hue, saturation, and brightness. This significantly improves the model's robustness in recognizing different lighting conditions and skin tones.

[0024] Normalization: Scales pixel values ​​to the range of 0-1; this operation standardizes pixel values, accelerates the model convergence process, improves training stability, and eliminates the impact of differences in camera hardware.

[0025] S3: The preprocessed image is input into a pre-trained YOLOv8 deep learning detection model. The model simultaneously outputs bounding boxes for human lower limbs and bare calves, as well as their respective confidence scores. The model outputs the bounding boxes and confidence scores of the two targets in parallel, achieving efficient and real-time detection and providing direct input for subsequent ratio calculations. Thanks to the backbone network and feature pyramid of YOLOv8, the model can effectively extract multi-scale features, especially ensuring the detection rate of small targets such as "bare calves" in complex backgrounds.

[0026] S4: Based on the lower limb detection frame and the exposed calf detection frame, execute the calf exposure violation judgment algorithm: calculate the ratio r of the area B of the exposed calf region to the area A of its corresponding lower limb region, i.e., r = B / A; if r is greater than a preset threshold T, it is judged as non-compliant clothing; the threshold T ranges from 20% to 40%; transforming the vague "non-compliant clothing" into a precise, calculable mathematical ratio problem. This method overcomes misjudgments caused by different distances between the person and the camera and different image sizes. It provides an adjustable and clear decision boundary. This threshold range, determined through numerous experiments, can effectively distinguish between "short socks / exposed ankles" and "true exposed calves," effectively reducing the false alarm rate while ensuring the detection rate.

[0027] S5: When a violation is detected, an audible and visual alarm mechanism is triggered, and relevant data of the violation event, including the violation frame image, timestamp, camera number, and exposed area R, is uploaded to the management platform. The on-site audible and visual alarm immediately alerts experimenters to correct inappropriate attire, achieving in-process intervention rather than post-event tracing, and truly playing a role in safety prevention.

[0028] Furthermore, the determination algorithm in step S4 further includes multi-frame tracking consistency determination: For a target individual being continuously tracked, let its determination result in N consecutive frames be the sequence {r}. i ,i=1,2,...,N}; Calculate r in this sequence i The proportion of the number of frames k to the total number of frames N is p, i.e., p = k / N; If p is greater than the consistency threshold p t If so, the final violation judgment will be output; The number of tracking frames N ranges from 5 to 15, and the consistency threshold p t The value ranges from 60% to 80%.

[0029] Furthermore, the YOLOv8 deep learning detection model in step S3 is trained using the following method: Construct a dedicated training dataset in which image annotations include two types of bounding boxes: "lower limbs" and "bare calves"; The training data is augmented using data augmentation strategies optimized for small objectives, including Mosaic augmentation and copy_paste. The YOLOv8 network was trained using the dataset described above. Its loss function L is a composite loss function, consisting of bounding box regression loss L1, classification loss L2, and confidence loss L3 weighted by coefficients. The specific formula is as follows: L = w1·L1 + w2·L2 + w3·L3 The weight coefficients w1, w2, and w3 have the following ranges: w1∈[4.0,6.0], w2∈[0.3,0.7], and w3∈[0.3,0.7].

[0030] Furthermore, the bounding box regression loss L1 adopts the GIoU loss, which is calculated as follows: Let the predicted bounding box be P, the ground truth bounding box be G, and the smallest closed box containing both P and G be C. L1 = 1 - GIoU, Among them, GIoU=IoU-|C-(P∪G)| / |C|, IoU=|P∩G| / |P∪G|; The predicted box P and the ground truth box G are bounding boxes, respectively represented as: P=(x 1p ,y 1p ,x 2p ,y 2p ) G=(x 1g ,y 1g ,x 2g ,y 2g ) The coordinates of the smallest closed bounding box C are calculated as follows: x 1c =min(x 1p ,x 1g ), y 1c =min(y 1p ,y 1g ) x 2c =max(x 2p ,x 2g ), y 2c =max(y 2p ,y 2g ).

[0031] Furthermore, in step S4, areas A and B are respectively detected by the lower limb detection frame P. leg And the detection frame for bare calves P bare The calculation yielded: A=area(P leg )=(x 2leg -x 1leg )×(y 2leg -y 1leg ) B = area(P) bare )=(x 2bare -x 1bare )×(y 2bare -y 1bare ).

[0032] Among them, P leg and P bare These are all predicted bounding boxes output by the model.

[0033] An automated detection system for exposed lower legs in a smart laboratory includes: The video acquisition module is used to acquire panoramic video data of the laboratory through multiple high-definition cameras; The image preprocessing module is used to normalize, geometrically enhance, and enhance the color space of video frames. The deep learning detection module has a built-in YOLOv8 model trained by the method described in claim 3 or 4, which is used to identify the lower limbs and bare lower leg areas. The judgment and alarm module is used to perform area ratio calculation r=B / A, multi-frame tracking consistency judgment p=k / N, and trigger alarms; The data management and feedback module is used to store violation records, generate statistical reports, and support incremental learning of the model.

[0034] This application leverages existing camera hardware infrastructure, upgrading it from a "recorder" to a "sensing source," enabling 24 / 7 uninterrupted data acquisition without continuous human monitoring. The core detection and judgment processes are entirely automated by the YOLOv8 model and algorithm, freeing managers from tedious and repetitive visual inspections. This "machine-assisted" approach significantly reduces manual inspection costs, allowing human resources to be allocated to higher-level management decisions, thus solving the fundamental problem of inefficiency. Furthermore, by introducing the mathematical formula "exposed area ratio r = B / A," this application transforms the subjective question of "whether clothing is appropriate" into an objective and measurable area percentage. Simultaneously, the system sets a uniform judgment threshold (e.g., T = 30%) for all monitored areas, ensuring consistent and fair judgment standards for all personnel throughout the laboratory, completely eliminating the subjectivity and arbitrariness of human judgment. This invention does not directly detect "people," but first detects "lower limbs," and then detects "exposed calves" within the lower limb area. This coarse-to-fine strategy effectively narrows the search range, eliminates a large amount of irrelevant background interference, and directly improves the detection signal-to-noise ratio.

[0035] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A smart laboratory leg exposure detection method based on YOLOv8 and multi-modal judgment, characterized by: The method comprises the following steps: S1: Real-time collection of video streams in the laboratory by deploying multiple high-definition cameras in the laboratory, and extraction of video frames; S2: Image preprocessing and enhancement of the extracted video frames, the enhancement including geometric enhancement and color space enhancement; S3: Input of the preprocessed images into a pre-trained YOLOv8 deep learning detection model, the model synchronously outputting a human lower limb detection frame and a bare calf detection frame, and respective confidence levels; S4: Based on the lower limb detection frame and the bare calf detection frame, execution of a bare calf exposure violation judgment algorithm: calculation of a ratio r of an area B of the bare calf region to an area A of the corresponding lower limb region, i.e. r=B / A; if the ratio r is greater than a preset threshold T, it is determined that there is a violation; the threshold T has a value range of 20% to 40%; S5: When a violation is determined, a sound-light alarm mechanism is triggered, and relevant data of the violation event, including a violation frame image, a timestamp, a camera number, and a bare exposure ratio R, is uploaded to a management platform.

2. The smart laboratory bare leg detection method based on YOLOv8 and multi-modal judgment of claim 1, wherein: The judgment algorithm in the step S4 further comprises multi-frame tracking consistency judgment: For a continuously tracked target individual, let its decision results in the consecutive N frames be the sequence {r i i=1,2,...,N}; The proportion p of the frame number k of r i in the sequence to the total frame number N, that is, p=k / N; If p is greater than the consistency threshold p t then output the final violation determination; The number of tracking frames N ranges from 5 to 15, and the consistency threshold p t The value ranges from 60% to 80%.

3. The smart laboratory bare leg detection method based on YOLOv8 and multi-modal judgment of claim 1, wherein: The YOLOv8 deep learning detection model in the step S3 is obtained by training in the following manner: A dedicated training data set is constructed, and the image annotations in the data set include two types of boundary boxes, i.e. "lower limb" and "bare calf"; A data enhancement strategy optimized for small targets is used to enhance the training data, and the strategy includes Mosaic enhancement and copy_paste; The YOLOv8 network is trained using the data set, and the loss function L is a composite loss function composed of a boundary box regression loss L1, a classification loss L2, and a confidence loss L3 according to weight coefficients, and the specific formula is: L=w1·L1+w2·L2+w3·L3 wherein the weight coefficients w1, w2, and w3 have value ranges of w1∈[4.0, 6.0], w2∈[0.3, 0.7], and w3∈[0.3, 0.7].

4. The smart laboratory bare leg detection method based on YOLOv8 and multi-modal judgment of claim 3, wherein: The boundary box regression loss L1 uses a GIoU loss, and the calculation method is as follows: Let P be a predicted box, G be a true box, and C be a minimum closed box containing P and G, L1=1-GIoU, wherein GIoU=IoU-|C-(P∪G)| / |C|, and IoU=|P∩G| / |P∪G|; The predicted box P and the true box G are boundary boxes, and are respectively represented as: P = (x 1p ,y 1p ,x 2p ,y 2p ) G = (x 1g ,y 1g ,x 2g ,y 2g ) The coordinate calculation method of the minimum closed box C is as follows: x 1c =min(x 1p ,x 1g ), y 1c =min(y 1p ,y 1g ) x 2c =max(x 2p ,x 2g ), y 2c =max(y 2p ,y 2g ).

5. The smart laboratory bare leg detection method based on YOLOv8 and multi-modal judgment of claim 1, wherein: The areas A and B in the step S4 are respectively calculated by the lower limb detection frame P leg and the exposed lower leg detection frame P bare The calculation result is: A = area (P leg )=(x 2leg -x 1leg )×(y 2leg -y 1leg ) B = area (P bare )=(x 2bare -x 1bare )×(y 2bare -y 1bare ) where P leg and P bare are the predicted boxes output by the model.

6. A smart lab bare leg automatic detection system for implementing the method of any one of claims 1-4, characterized in that: It comprises: a video acquisition module for acquiring panoramic video data in the laboratory through multiple high-definition cameras; an image preprocessing module for normalizing, geometrically enhancing, and color space enhancing video frames; a deep learning detection module having a YOLOv8 model trained by the method of claim 3 or 4, for identifying lower limbs and bare calf regions; a judgment and alarm module for calculating an area ratio r=B / A, performing multi-frame tracking consistency judgment p=k / N, and triggering an alarm; a data management and feedback module for storing violation records, generating statistical reports, and supporting model incremental learning.

7. The method of claim 5, wherein: The system also comprises a model iterative optimization module, which is configured to periodically collect new scene data, incrementally train the YOLOv8 model at a fine-tuning learning rate lower than the initial learning rate, and take the validation set mAP 0.5 as a core evaluation index, and automatically update the online model when the performance improvement exceeds a set threshold.

Images