Elevator door lock short circuit state recognition method based on YOLO algorithm

Through the combination of YOLO algorithm and pyramid-enhanced network, the problems of long detection time of elevator door locks and large manpower investment are solved, and fast and accurate identification of short-connection status of elevator door locks is achieved, which improves elevator safety.

CN119067932APending Publication Date: 2024-12-03CHONGQING SPECIAL EQUIP INSPECTION & RES INST +1
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411115778.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing elevator door lock detection methods are long and labor-intensive, making it difficult to quickly identify the short-connected state of elevator door locks, which poses safety hazards.

Method used

The elevator door lock short-circuit status recognition method based on the YOLO algorithm is adopted. Through offline training and online detection steps, combined with the pyramid enhancement network and low-frequency enhancement filter, a dark object detection framework is built, and the elevator door lock short-circuit situation is monitored in real time using image data.

Benefits of technology

It realizes the quick and accurate identification of the short-connected state of the elevator door lock, reduces detection time and manpower investment, and improves the safety of the elevator.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067932A_ABST
    Figure CN119067932A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of elevator door lock detection, and discloses an elevator door lock short circuit state recognition method based on a YOLO algorithm, and the method comprises an offline training step and an online detection step, a first image with an elevator cabinet control lamp in a normal light environment and a second image with an elevator cabinet control lamp in a dim light environment are collected; s20, the first image serves as an input layer parameter, the elevator door lock short circuit state serves as an output layer parameter, the first image serves as a training set, and a data preprocessing step, a network detection step and a loss calculation step are sequentially carried out for training to obtain a first model; the method has the beneficial effects that target image detection is carried out by adopting the YOLO algorithm on the basis of continuously collected image data of the elevator cabinet control lamp so as to identify the short-circuit state of the elevator door lock. According to the method, a dark object detection framework is constructed by combining the pyramid enhanced network with the YOLO algorithm, so that the aim of accurately detecting the short-circuit state of the elevator door lock in a dark light environment is fulfilled.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of elevator door lock detection, and in particular to a method for identifying a short-circuited state of an elevator door lock based on a YOLO algorithm. Background Art

[0002] As an important means of transportation in modern cities, the safety of elevators has attracted much attention. As one of the key components of elevators, elevator door locks are directly related to the safety of passengers. Therefore, the elevator industry has extremely high requirements for the reliability and safety of elevator door locks. The main function of elevator door locks is to confirm the locking and closing of elevator door leaves to ensure the safety of passengers when riding the elevator. Once the elevator door lock short-circuits, it may cause the elevator door to open unexpectedly or close loosely, posing a serious safety hazard. If the elevator door lock electrical switch is artificially short-circuited, it will cause the elevator door to open and the elevator to walk, which may cause serious accidents such as shearing, squeezing or falling of people. Every year, provinces and cities across the country carry out special spot checks including door lock short-circuiting to discover and punish such serious accident hazards.

[0003] In the prior art, elevator door lock detection often uses electrical parameter detection and vibration sensor detection. Electrical parameter detection is to determine whether a short circuit occurs by detecting the electrical parameters of the elevator door lock. This method usually requires professional instruments and operators, and the detection process is relatively cumbersome and takes a long time. Vibration sensor detection is to install a vibration sensor on the elevator door lock to monitor the vibration of the door lock in real time, and then determine whether there is a short circuit or abnormality. This method is critical to the selection and location of the sensor, and requires accurate data analysis and judgment.

[0004] It can be seen that the traditional door lock short-circuit detection method and the existing detection instruments are time-consuming and require a large amount of manpower. Therefore, a method for quickly detecting elevator door lock short-circuit is urgently needed. Summary of the invention

[0005] In order to overcome the shortcomings of long detection time and large manpower investment in elevator door lock detection in the prior art, the present invention proposes a method for identifying the short-circuit state of an elevator door lock based on the YOLO algorithm, which specifically includes the following technical solutions.

[0006] A method for identifying the short-circuit state of an elevator door lock based on a YOLO algorithm comprises an offline training step and an online detection step. The offline training step comprises

[0007] S10, collecting a first image with an elevator cabinet control light in a normal light environment and a second image with an elevator cabinet control light in a dim light environment;

[0008] S20, taking the first image as an input layer parameter, taking the elevator door lock short-circuited state as an output layer parameter, taking the first image as a training set, and sequentially performing a data preprocessing step, a detection network step, and a loss calculation step to obtain a first model;

[0009] S30, using the second image as an input layer parameter, using the elevator door lock short-circuit state as an output layer parameter, and using the second image as a training set, first decomposing it into components of different resolutions through a pyramid enhancement network, and then using a detail processing module and a low-frequency enhancement filter to enhance the components of each scale to obtain a third image with deep feature fusion, and then sequentially performing a data preprocessing step, a detection network step, and a loss calculation step on the third image to obtain a second model;

[0010] Online detection steps:

[0011] S40, collect images with elevator cabinet control lights and the light intensity of the environment in which the elevator cabinet control lights are located. If the light intensity is less than or equal to 5lux, the collected image is used as the input layer parameter, and the short-circuited state of the elevator door lock is output through the second model; if the light is greater than 5lux, the collected image is used as the input layer parameter, and the short-circuited state of the elevator door lock is output through the first model.

[0012] Further, in S30, a pyramid enhancement network is combined with the YOLO algorithm to construct a dark object detection framework. The second image is taken as input, a Gaussian pyramid is used to obtain sub-images of different resolutions, and a Laplacian pyramid is used to reconstruct the image. The sampling formula of the Gaussian pyramid is:

[0013] G(x)=Down(Gaussian(x)) (11)

[0014] Among them, Down means down sampling, Gaussian means Gaussian filter, and the size of Gaussian kernel is 5×5; Gaussian expression is:

[0015]

[0016] The Gaussian function can easily implement Gaussian blur processing, and the degree of blur can be controlled by adjusting the size and standard deviation of the filter as needed. As x increases, the blurring effect becomes more obvious.

[0017] The Laplacian pyramid is defined as follows:

[0018] DATA-D i =G i -Up(G i+1 ), (12)

[0019] Among them, DATA_D irepresents the i-th level of the Laplacian pyramid, G i represents the i-th layer of the Gaussian pyramid, and Up represents the bilinear upsampling operation.

[0020] Further, the detail processing module is composed of a context branch and an edge branch;

[0021] The context branch uses the first residual block to change the channels of the features of the second image from 3 to 32, and uses the second residual block to change the channels of the features of the second image from 32 to 3, which is defined as follows:

[0022]

[0023] In the formula, Represents the output of the context branch, the expression is: F1 is a convolutional layer with a 3×3 convolution kernel, γ is a Leaky ReLU activation function, and σ is a Softmax function;

[0024]

[0025] x is a vector or array of confidences about the detected categories.

[0026] The edge branch uses a Sobel operator in the horizontal and vertical directions of the second image to re-extract edge information through a convolution filter, and then fuses the edge information with features in the second image to obtain a third image;

[0027] The process of edge branching is expressed as:

[0028] EB(x)=F3(Sobel h (x)+Sobel w (x))+x (14)

[0029] In the formula, the Sobel operator is a discrete operator that uses Gaussian filtering and differential derivatives at the same time; F3 is a convolution layer with a 3×3 convolution kernel.

[0030] Furthermore, the low-frequency enhancement filter is used to capture the low-frequency information in the second image; in S30, the second image is recorded as f∈R h×w×3 , the second image is transformed into f∈R through a convolutional layer h×w×32 , a dynamic low-pass filter is used to capture low-frequency information, and average pooling is used for feature filtering.

[0031] Furthermore, the third image is obtained by decomposing the second image by a Gaussian pyramid, fusing the edge information obtained by the detail processing module or the low-frequency information captured by the low-frequency enhancement filter, and reconstructing it by a Laplacian pyramid.

[0032] Furthermore, in S30, adaptive average pooling of sizes 1×1, 2×2, 3×3, and 6×6 is used, and upsampling is used at the end of each scale to restore the original size of the feature; the average pooling of different convolution kernel sizes forms a low-pass filter

[0033] Furthermore, the normal light environment is a light environment with a light intensity greater than 5lux; the dim light environment is a light environment with a light intensity less than or equal to 5lux.

[0034] Furthermore, the data preprocessing step includes flipping, scaling, mean normalization and Y label assignment, wherein the formula for mean normalization is:

[0035]

[0036] In the formula, Bx and By represent the normalized coordinate values, X s and Y s is the position of the original pixel, X w and Y h are the width and height of the original image;

[0037] Y label assignment refers to taking the first image and its corresponding Y label as input, where the Y label is a multidimensional vector containing the probability of the defined object existing in the grid, the normalized coordinate value, the normalized length and width, and the elevator door lock short-circuit state.

[0038] Furthermore, the detection network step includes intersection-over-union calculation, precision calculation, recall calculation and average precision calculation;

[0039] The intersection-over-union calculation is used to measure the relative size of the overlap of two edge bounding boxes, and its formula is:

[0040]

[0041] Where A represents bounding box A, and B represents bounding box B;

[0042] The formula for calculating the accuracy is:

[0043]

[0044] The formula for calculating the recall rate is:

[0045]

[0046] In formula (3) and formula (4), TP is the number of positive examples predicted and the actual value is also positive; FN is the number of negative examples predicted and the actual value is positive; FP is the number of positive examples predicted and the actual value is negative.

[0047] The average precision calculation calculates the area enclosed by the precision curve and recall curve of each category and the coordinate axis. The calculation formula is:

[0048]

[0049] In the formula, AP is the average precision, P is the precision curve, and R is the recall curve.

[0050] Furthermore, the loss calculation step includes performing a loss on the predicted center coordinates, performing a loss on the width and height of the predicted bounding box, performing a loss on the predicted category, and performing a loss on the predicted confidence;

[0051] The formula for the loss of the predicted center coordinates is:

[0052]

[0053] λ is a set constant. The function calculates the sum of each bounding box prediction value (j=0,...,B) for each grid cell (i=0,...,S^2); is defined as follows:

[0054] 1. If there is an object in grid cell i, the jth bounding box prediction value is valid for that prediction;

[0055] 0, then there is no target in grid cell i;

[0056] The loss formula for the width and height of the predicted bounding box is:

[0057]

[0058] The formula for the loss of the predicted category is:

[0059]

[0060] The formula for loss of predicted confidence is:

[0061]

[0062] C is the confidence score, C is the intersection of the predicted bounding box and the ground truth; when there is an object in a cell, is equal to 1, otherwise it is 0.

[0063] The beneficial effects of the present invention are as follows: the present invention uses the YOLO algorithm to detect the target image based on the continuously collected image data of the elevator cabinet control light to identify the short-circuited state of the elevator door lock. At the same time, due to the dim light environment in the elevator cabinet, in order to eliminate the influence of dim light, a pyramid enhancement network is used in conjunction with the YOLO algorithm to construct a dark object detection framework, based on the original image, low-frequency information is captured through a low-frequency enhancement filter, edge information is obtained using a detail processing module, and the low-frequency information and edge information are deeply fused with the original image features to achieve the purpose of accurately detecting the short-circuited state of the elevator door lock.

[0064] The present invention is based on image data, can monitor the short circuit of elevator door lock in real time, and make rational use of resources. Compared with the existing elevator door lock detection method, it takes less time and wastes less manpower. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a schematic diagram of the process of the present invention.

[0066] Figure 2 This is a flow chart of the model training for a normal light environment of the present invention.

[0067] Figure 3 It is a flow chart of model training for normal light environment and dim light environment of the present invention.

[0068] Figure 4 It is a schematic diagram of the overlap of prediction frames of the present invention.

[0069] Figure 5 It is a structural schematic diagram of the detail processing module of the present invention.

[0070] Figure 6 It is a structural schematic diagram of the low-frequency enhancement filter of the present invention. DETAILED DESCRIPTION

[0071] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0072] A method for identifying the short-circuit state of an elevator door lock based on the YOLO algorithm, such as Figure 1 As shown, it includes offline training steps and online detection steps.

[0073] In the offline training step, first, in the case of the elevator door lock being short-circuited and the elevator door lock being not short-circuited, images with the elevator cabinet control light at different angles, different distances and different light environments are collected. In the present invention, different light environments mainly refer to normal light environments and dim light environments. Since the dim light environment will have a more obvious impact on image detection, different training steps are used in the present invention to perform model training for different light environments. In the present invention, a light environment with a light intensity less than or equal to 5lux is defined as a dim light environment, and a light environment with a light intensity greater than 5lux is defined as a normal light environment.

[0074] For images under normal light conditions, Figure 2 The training method shown in the figure takes the following approach for images with both normal light environment and dim light environment: Figure 3 The training method shown in Figure 2. The image training steps under normal light environment include 1. data preprocessing, 2. detection network, and 3. loss calculation.

[0075] 1. Data preprocessing includes: flipping, scaling, and mean normalization. The coordinate normalization formula is:

[0076]

[0077] X, Y are normalized coordinate values, ranging from 0 to 1. s and Y s is the position of the original pixel, X w and Y h is the width and height of the original image. After normalization, the image and its corresponding Y label are taken as input, where the Y label is a multidimensional vector about (1) defining whether the object exists in the grid (probability of existence) P, (2) specifying the bounding box, and (3) representing the category C. Assume that the image has been divided into a grid of size 3X3, and there are only 2 categories in total, namely, the elevator door lock is normal (c1) and the elevator door lock is shorted (c2). Therefore, for each cell, the label Y will be a seven-dimensional vector, which is expressed as:

[0078]

[0079]

[0080] 2. Detection network, including IoU (intersection over union), precision, Recall, and AP (average precision).

[0081] IoU (intersection over union) measures the relative size of the overlap between two bounding boxes, and the accuracy of the predicted box is reflected by IoU. IoU is an important indicator in the object detection problem. It reflects the degree of overlap between the annotation box and the predicted box during the training phase and is used to measure the accuracy of the predicted box. Figure 4 As shown:

[0082] The formula is:

[0083]

[0084] The closer IOU is to 0, the less the two boxes overlap; the closer IOU is to 1, the higher the degree of overlap between the two boxes. When IOU is equal to 1, the two boxes overlap completely. IOU should be used together with the IOU threshold, which is the threshold of accuracy. The IOU threshold is generally set to 0.5. When the IOU of the two boxes is greater than the threshold, the predicted box is judged to be correctly predicted. The IOU threshold can be modified. The higher the IOU threshold, the stricter the condition for judging the prediction of the predicted box is correct. The accuracy formula is:

[0085]

[0086] Among them, TP is the number of positive examples predicted and the true value is also positive; FP is the number of negative examples predicted and the true value is positive. From the analysis formula, we can see that Precision is concerned with the predicted positive examples and the true positive and negative examples. When Precision is larger, FP is smaller, and the number of other categories predicted as this category is smaller. It can be understood that the purity of the predicted positive examples is higher. The higher the Precision, the fewer false positives.

[0087] Recall refers to the recall rate. The definition of Recall is as follows:

[0088]

[0089] According to the definition, the denominator of Recall is the sum of TP and FN. TP is the number of positive examples predicted and true positive examples; FN is the number of negative examples predicted and true positive examples. From the analysis formula, we can see that Recall is concerned with the predicted positive and negative examples, as well as the true positive examples. When Recall is larger, FN is smaller. At this time, the number of positive examples predicted as negative examples is smaller, which can be understood as picking out more positive examples. The higher the Recall, the fewer missed detections.

[0090] Although it is called average precision, the calculation method of AP is not to calculate the average value of Precision, but to calculate the area enclosed by the PR curve of each category and the coordinate axis, which can be calculated by integration. If the AP of a model is larger, that is, the area enclosed by the PR curve and the coordinate axis is larger, the Precision and Recall are relatively high overall. The expression of AP is:

[0091] AP = ∫0 1 P(R)dR (5)

[0092] mAP (mean of Average Precision): averages the AP values ​​of all categories. AP can reflect the accuracy of prediction for each category. mAP is the average of the AP values ​​of all categories, which is used to reflect the accuracy of the entire model. The larger the mAP, the larger the area between the PR curve and the coordinate axis. Usually, when we say that the accuracy of a certain target detection algorithm has reached a certain level, this accuracy generally refers to mAP.

[0093]

[0094] 3. Loss calculation. In the present invention, the loss function of YOLO consists of 4 parts:

[0095] 3.1. Loss on the predicted center coordinates:

[0096]

[0097] This equation calculates the loss value relative to the predicted bounding box position (x, y). Assume λ is a given constant. This function calculates the sum of each bounding box prediction value (j=0,...,B) for each grid cell (i=0,...,S2)(i=0,...,S^2)(i=0,...,S2). is defined as follows:

[0098] 1. If there is an object in grid cell i, then the jth bounding box prediction is valid for that prediction.

[0099] 0, then there is no target in grid cell i.

[0100] 3.2 Loss on the width and height of the predicted bounding box:

[0101]

[0102] This is the loss associated with the predicted bounding box width / height. The equation looks similar to the first one except for the square root. Not applying the square root would cause the error metric to reflect small deviations for large boxes less than for small boxes. To gradually fix this, we predict the square root of the bounding box width and height instead of predicting the width and height directly.

[0103] 3.3 Loss on the predicted category

[0104]

[0105] Apart from The equation looks similar to the normal summed squared error for classification except for the term . This term is used because we do not penalize classification errors when there is no object at the cell (the conditional class probabilities discussed earlier).

[0106] 3.4 Loss on the confidence of prediction

[0107]

[0108] Here we calculate the loss associated with the confidence score of each bounding box prediction. Where confidence is IOU*Precision. C is the confidence score and C is the intersection of the predicted bounding box and the ground truth. When there is an object in a cell, is equal to 1, otherwise it is 0. The λ parameter here and in the first part is used to weight the different parts of the loss function. This is very important to improve the stability of the model. The highest penalty is for the coordinate prediction (λ coord =5), when no target is detected, there is the lowest confidence prediction threshold (λ coord =0.5).

[0109] For image training in dim light environments, a dark object detection framework PE-YOLO is constructed by combining the pyramid enhancement network (PENet) with the YOLO algorithm. In PENet, the image is decomposed into components of different resolutions through the Laplacian pyramid, and the detail processing module (DPM) and low-frequency enhancement filter (LEF) are used to enhance the components of each scale.

[0110] Assume that image I∈R h×w×3 As input, a Gaussian pyramid is used here to obtain sub-images of different resolutions. The formula is:

[0111] G(x)=Down(Gaussian(x)) (11)

[0112] Among them, Down means down sampling, Gaussian means Gaussian filter, and the size of Gaussian kernel is 5×5. Gaussian expression is:

[0113]

[0114] The Gaussian function can easily implement Gaussian blur processing, and the degree of blur can be controlled by adjusting the size and standard deviation of the filter as needed. As x increases, the blurring effect becomes more obvious.

[0115] After each Gaussian pyramid operation, the width and height of the image are halved, which means that the resolution is 1 / 4 of the original image. Obviously, the downsampling operation of the Gaussian pyramid is irreversible. In order to restore the original high-resolution image after upsampling, it is necessary to restore the lost information, which constitutes the components of the Laplacian pyramid. The definition of the Laplacian pyramid is as follows:

[0116] DATA_D i =G i -Up(G i+1 ), (12)

[0117] Among them, DATA_D i represents the i-th level of the Laplacian pyramid, G i represents the i-th layer of the Gaussian pyramid. Up represents the bilinear upsampling operation. When reconstructing an image, only the inverse operation of formula (12) needs to be performed to restore the high-resolution image.

[0118] In dim lighting conditions, the structure of the detail processing module (DPM) is as follows Figure 5 As shown, the structure of the low frequency enhancement filter (LEF) is as follows Figure 6 As shown in Figure 2, the detail processing module (DPM) is used to enhance the components in the Laplacian pyramid, which consists of a context branch and an edge branch.

[0119] (1) Context branch:

[0120] Residual blocks (res blocks) are used to process features before and after acquiring long-range dependencies, and residual learning allows rich low-frequency information to be passed through skip connections. The first residual block increases the number of channels of features from 3 to 32, and the second residual block increases the number of channels of features from 32 to 3. Capturing global information in the scene has been shown to be beneficial for low-level visual tasks such as dark light enhancement. It is defined as follows:

[0121]

[0122] In the above formula, represents the output of the context branch, F is a convolutional layer with a 3×3 convolution kernel, γ is a Leaky ReLU activation function, and σ is a Softmax function.

[0123] (2) Marginal branches:

[0124] The Sobel operator is a discrete operator that uses Gaussian filtering and differential derivation at the same time. It finds edges by calculating gradient approximations. The Sobel operator is used in both horizontal and vertical directions to re-extract edge information through convolution filters and use the residual to enhance the flow of information. The process is expressed as:

[0125] EB(x)=F3(Sobel h (x)+Sobel w (x))+x (14)

[0126] In the above formula, Sobel h and Sobel w Represents the Sobel operation in the vertical and horizontal directions respectively. These two operations are applied to the input features to calculate the gradient approximation in the vertical and horizontal directions, thereby extracting the edge information in the image. In DPM, these two Sobel operations are used here to obtain edge information and add it to the original input features to enhance the feature representation.

[0127] For the low-frequency enhancement filter (LEF): In each scale component, the low-frequency components contain most of the semantic information in the image, which are the key information for detector prediction. In order to enrich the semantic information of the reconstructed image, a low-frequency enhancement filter (LEF) is proposed to capture the low-frequency information in the component. Assume that the component is f∈R h×w×3 , first through a convolutional layer to transform it into f∈R h×w×32 , a dynamic low-pass filter is used here to capture low-frequency information, and average pooling is used for feature filtering, allowing only information below the cutoff frequency to pass. The low-frequency thresholds for different semantics are different. Considering the multi-scale structure of Inception, adaptive average pooling of sizes 1×1, 2×2, 3×3, and 6×6 is used here, and upsampling is used at the end of each scale to restore the original size of the feature. Average pooling of different convolution kernel sizes forms a low-pass filter. Through channel separation, f is divided into 4 parts, namely {f1,f2,f3,f4}. Each part is processed using pooling of different sizes, as described below:

[0128] Filter(f i )=Up(β s (f i )) (15)

[0129] where f i is the part of f divided on the channel, Up is the bilinear interpolation sampling, β sis an adaptive average pooling of different sizes s×s. Finally, concatenate each {f i ,i=1,2,3,4}, and restore them to f∈R h×w×3 .

[0130] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the short-circuit state of an elevator door lock based on the YOLO algorithm, comprising an offline training step and an online detection step. It is characterized in that The offline training step includes S10, collecting a first image with an elevator cabinet control light in a normal light environment and a second image with an elevator cabinet control light in a dim light environment; S20, taking the first image as an input layer parameter, taking the elevator door lock short-circuited state as an output layer parameter, taking the first image as a training set, and sequentially performing a data preprocessing step, a detection network step, and a loss calculation step to obtain a first model; S30, using the second image as an input layer parameter, using the elevator door lock short-circuit state as an output layer parameter, and using the second image as a training set, first decomposing it into components of different resolutions through a pyramid enhancement network, and then using a detail processing module and a low-frequency enhancement filter to enhance the components of each scale to obtain a third image with deep feature fusion, and then sequentially performing a data preprocessing step, a detection network step, and a loss calculation step on the third image to obtain a second model; Online detection steps: S40, collecting an image with an elevator cabinet control light and the light intensity of the environment where the elevator cabinet control light is located. If the light intensity is less than or equal to 5lux, the collected image is used as an input layer parameter, and the second model outputs the elevator door lock short-circuit state; If the illumination is greater than 5lux, the collected image is used as the input layer parameter, and the first model outputs the short-circuit state of the elevator door lock.

2. According to the method for identifying the short-circuit state of an elevator door lock based on the YOLO algorithm according to claim 1, It is characterized in that In S30, a pyramid enhancement network is used in conjunction with the YOLO algorithm to construct a dark object detection framework. The second image is used as input, a Gaussian pyramid is used to obtain sub-images of different resolutions, and the Laplacian pyramid is used to reconstruct the image. The sampling formula of the Gaussian pyramid is: G(x)=Down(Gaussian(x)) (11) Among them, Down means down sampling, Gaussian means Gaussian filter, and the size of Gaussian kernel is 5×5; Gaussian expression is: The Gaussian function can easily implement Gaussian blur processing, and the degree of blur can be controlled by adjusting the size and standard deviation of the filter as needed; as x increases, the blur effect becomes more obvious; The Laplacian pyramid is defined as follows: DATA_D i =G i -Up(G i+1 ), (12) Among them, DATA_D i represents the i-th level of the Laplacian pyramid, G i represents the i-th layer of the Gaussian pyramid, and Up represents the bilinear upsampling operation.

3. According to claim 2, the elevator door lock short circuit state recognition method based on the YOLO algorithm, It is characterized in that The detail processing module consists of a context branch and an edge branch; The context branch uses the first residual block to change the channels of the features of the second image from 3 to 32, and uses the second residual block to change the channels of the features of the second image from 32 to 3, which is defined as follows: In the formula, Represents the output of the context branch, the expression is: F 1 is a convolutional layer with a 3×3 convolution kernel, γ is the LeakyReLU activation function, and σ is the Softmax function; x is a confidence vector or array about the detection category; The edge branch uses the Sobe operator in the horizontal and vertical directions of the second image to re-extract edge information through a convolution filter, and then fuses the edge information with the features in the second image to obtain a third image; The process of edge branching is expressed as: EB(x)=F 3 (Sobel h (x)+Sobel w (x))+x (14) In the formula, the Sobe| operator is a discrete operator that uses Gaussian filtering and differential derivation at the same time; F3 is a convolution layer with a 3×3 convolution kernel.

4. According to claim 3, the elevator door lock short circuit state recognition method based on the YOLO algorithm, It is characterized in that The low-frequency enhancement filter is used to capture the low-frequency information in the second image; in S30, the second image is recorded as f∈R h×w×3 , the second image is transformed into f∈R through a convolutional layer h×w×32 , a dynamic low-pass filter is used to capture low-frequency information, and average pooling is used for feature filtering.

5. According to claim 4, the elevator door lock short circuit state recognition method based on the YOLO algorithm, It is characterized in that The third image is obtained by decomposing the second image by a Gaussian pyramid, fusing the edge information obtained by the detail processing module or the low-frequency information captured by the low-frequency enhancement filter, and reconstructing it by a Laplacian pyramid.

6. The elevator door lock short circuit state recognition method based on the YOLO algorithm according to claim 4, It is characterized in that In S30, adaptive average pooling of sizes 1×1, 2×2, 3×3, and 6×6 is used, and upsampling is used at the end of each scale to restore the original size of the feature; average pooling of different convolution kernel sizes forms a low-pass filter.

7. The elevator door lock short circuit state recognition method based on the YOLO algorithm according to claim 1, It is characterized in that The normal light environment is a light environment with a light intensity greater than 5lux; the dim light environment is a light environment with a light intensity less than or equal to 5lux.

8. The elevator door lock short circuit state recognition method based on the YOLO algorithm according to claim 1, It is characterized in that The data preprocessing steps include flipping, scaling, mean normalization and Y label assignment, where the formula for mean normalization is: In the formula, Bx and By represent the normalized coordinate values, X s and Y s is the position of the original pixel, X w and Y h are the width and height of the original image; Y label assignment refers to taking the first image and its corresponding Y label as input, where the Y label is a multidimensional vector containing the probability of the defined object existing in the grid, the normalized coordinate value, the normalized length and width, and the elevator door lock short-circuit state.

9. The elevator door lock short circuit state recognition method based on the YOLO algorithm according to claim 1, It is characterized in that The detection network step includes intersection-over-union calculation, precision calculation, recall calculation and average precision calculation; The intersection-over-union calculation is used to measure the relative size of the overlap of two edge bounding boxes, and its formula is: Where A represents bounding box A, and B represents bounding box B; The formula for calculating the accuracy is: The formula for calculating the recall rate is: In formula (3) and formula (4), TP is the number of positive examples predicted and the actual value is also positive; FN is the number of positive examples predicted and the actual value is positive; FP is the number of positive examples predicted and the actual value is negative; The average precision calculation calculates the area enclosed by the precision curve and recall curve of each category and the coordinate axis. The calculation formula is: In the formula, AP is the average precision, P is the precision curve, and R is the recall curve.

10. The elevator door lock short circuit state recognition method based on the YOLO algorithm according to claim 1, It is characterized in that The loss calculation step includes performing a loss on the predicted center coordinates, performing a loss on the width and height of the predicted bounding box, performing a loss on the predicted category, and performing a loss on the predicted confidence; The formula for the loss of the predicted center coordinates is: λ is a set constant. The function calculates the sum of each bounding box prediction value (j=0,...,B) for each grid cell (i=0,...,S^2); is defined as follows:

1. If there is an object in grid cell i, the jth bounding box prediction value is valid for that prediction; 0, then there is no target in grid cell i; The loss formula for the width and height of the predicted bounding box is: The formula for the loss of the predicted category is: The formula for loss of predicted confidence is: C is the confidence score, C is the intersection of the predicted bounding box and the ground truth; when there is an object in a cell, is equal to 1, otherwise it is 0.

Citation Information

Patent Citations

  • Method for detecting specific target in elevator based on YOLO neural network

    CN110298265A

  • Elevator floor display abnormity identification method and device and computer readable storage medium

    CN114663806A

  • Coal flow foreign matter identification method for coal mine belt conveyor based on machine vision

    CN116665011A

  • Target detection method under low illumination based on pyramid enhanced network

    CN117058019A