A method, system and related device for detecting safety status of train doors

By automatically identifying the status of train doors through high-definition video capture and deep learning models, the problem of low efficiency and insufficient accuracy of manual inspection has been solved, realizing automated door safety detection and ensuring train operation safety.

CN114067254BActive Publication Date: 2024-11-05SHANDONG MATRIX SOFTWARE ENG +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111401381.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-19
Publication Date
2024-11-05
Estimated Expiration
2041-11-19

AI Technical Summary

Technical Problem

The existing methods for inspecting railway train doors mainly rely on manual visual inspection, which is labor-intensive and prone to fatigue, resulting in low detection accuracy, the risk of missed inspections, and affecting train operation safety.

Method used

High-definition video acquisition and deep learning models are used to automatically identify the safety status of train doors. Video data is transmitted to back-end processing equipment via the RTMP protocol for preprocessing and feature point recognition. ResNet backbone network and deep learning models are used for door detection.

Benefits of technology

It enables automatic identification of the loading and reinforcement status of train doors, reducing the time required for manual inspection, improving the accuracy of detection, and ensuring the safe operation of freight trains.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114067254B_ABST
    Figure CN114067254B_ABST
Patent Text Reader

Abstract

The present application provides a method, system and related device for detecting the safety status of train doors, which relates to the field of train safety detection, including: receiving a detection request and collecting high-definition video data of train doors; transmitting the high-definition video data to a back-end processing device via the RTMP protocol; in the back-end device, pre-processing the high-definition video data to obtain image data containing feature points; inputting the image data containing feature points into a preset deep learning model to obtain recognition results of each feature point; integrating the recognition results of each feature point to obtain door detection results. The present application can automatically identify the loading and reinforcement status of freight train doors by calling a preset deep learning model for detection and processing of the high-definition video data of train doors, without the need for manual on-site inspection, saving departure time, and effectively ensuring the driving safety of freight trains.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of train safety detection, and in particular to a method and system for detecting the safety status of a train door. Background Art

[0002] Door loading reinforcement is an important measure to ensure operational safety and cargo safety. In order to ensure the driving safety of freight trains, it is necessary to check the door loading reinforcement status of freight trains after the cargo is loaded and before departure. The existing railway train inspection method mainly relies on the operation mode of manual visual inspection, and checks are carried out car by car. The workload is large. It takes about 40 minutes to inspect a train with 50 cars. Due to the limitation of human energy, fatigue is inevitable, and it is easy to bring the risk of missed inspection due to negligence, especially for problems that are not easy to find. The detection accuracy is low, causing hidden dangers to driving safety. Summary of the invention

[0003] The purpose of the present application is to provide a safety status detection method and a safety status detection system for train doors, which can automatically identify the safety status of the doors.

[0004] In order to solve the above technical problems, the present application provides a method for detecting the safety status of a train door, and the specific technical solution is as follows:

[0005] receiving a detection request and collecting high-definition video data of a train door; the train door comprises a middle door and a small door, the middle door comprises a door bolt and a middle door latch, and the small door comprises a small door lock;

[0006] Transmitting the high-definition video data to a back-end processing device via the RTMP protocol;

[0007] In the back-end device, the high-definition video data is pre-processed to obtain image data containing feature points; the feature points include the door bolt, the middle door latch and the small door lock;

[0008] Inputting the image data containing the feature points into a preset deep learning model to obtain recognition results of each feature point;

[0009] The recognition results of each feature point are integrated to obtain a door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least includes a car number and a corresponding abnormal feature point.

[0010] Optionally, in the backend device, preprocessing the high-definition video data to obtain image data containing feature points includes:

[0011] Parsing the high-definition video data, distinguishing carriages according to axles, and identifying the carriage number of each carriage;

[0012] Carriage identification is performed on each of the carriages, and a middle door and a small door are marked; and image data containing the middle door and / or the small door is determined in the high-definition video data;

[0013] The middle door and the small door are used as features, feature point recognition is performed on the image data, and feature points are obtained by marking.

[0014] Optionally, before inputting the image data containing the feature points into a preset deep learning model to obtain the recognition results of each feature point, the method further includes:

[0015] Get learning samples;

[0016] Performing at least one of image resizing, image flipping, random erasing, automatic enhancement, and random patching on the learning sample to obtain a training sample;

[0017] ResNet is used as the backbone network, connected to the common feature aggregation network, and common features are extracted from the convolutional network containing feature point images in different scenes to obtain feature maps; the loss function of the preset deep learning model includes cross entropy loss, face recognition algorithm loss and triplet loss;

[0018] Aggregating the feature maps to obtain a global vector;

[0019] The global vector is normalized, subjected to head linear processing and head simplification processing to obtain a preset deep learning model.

[0020] Optionally, also include:

[0021] An attention-type non-local module and an IBN module are added to the backbone network.

[0022] Optionally, preprocessing the high-definition video data includes:

[0023] Acquire target image data including the middle door and / or the small door in the high-definition video data;

[0024] Determine the image quality of each of the target image data by determining the clarity of the middle gate and / or the small gate according to an image quality scoring rule;

[0025] The target image data whose image quality is higher than a preset quality score is retained.

[0026] The present application also provides a train door safety status detection system, comprising:

[0027] The acquisition module is used to receive the detection request and acquire high-definition video data of the train door; the train door includes a middle door and a small door, the middle door includes a bolt and a middle door latch, and the small door includes a small door lock;

[0028] A data transmission module, used to transmit the high-definition video data to a back-end processing device via the RTMP protocol;

[0029] The back-end device is used to pre-process the high-definition video data to obtain image data containing feature points; the feature points include the door bolt, the middle door latch and the small door lock; the image data containing the feature points is input into a preset deep learning model to obtain the recognition result of each feature point; the recognition results of each feature point are integrated to obtain the door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least includes the car number and the corresponding abnormal feature point.

[0030] Optionally, the backend device includes:

[0031] The image analysis unit is used to analyze the high-definition video data, distinguish the carriages according to the axles, and identify the carriage number of each carriage; identify the carriage of each carriage and mark the middle door and the small door; determine the image data containing the middle door and / or the small door in the high-definition video data; use the middle door and the small door as features, identify feature points of the image data, and mark the feature points.

[0032] Optionally, also include:

[0033] The model training module is used to obtain learning samples; perform at least one of image resizing, image flipping, random erasing, automatic enhancement and random patching on the learning samples to obtain training samples; use ResNet as the backbone network, connect to the common feature aggregation network, and extract common features from the convolution network containing feature point images in different scenes to obtain feature maps; the loss function of the preset deep learning model includes cross entropy loss, face recognition algorithm loss and triplet loss; aggregate the feature maps to obtain a global vector; normalize, head linearize and head simplify the global vector to obtain a preset deep learning model.

[0034] The present application also provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the steps of the above-mentioned method when executed by a processor.

[0035] The present application also provides a server, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps of the above-mentioned method when calling the computer program in the memory.

[0036] The present application provides a method for detecting the safety status of a train door, comprising: receiving a detection request and collecting high-definition video data of the train door; the train door comprises a middle door and a small door, the middle door comprises a door bolt and a middle door latch, and the small door comprises a small door lock; the high-definition video data is transmitted to a back-end processing device via an RTMP protocol; in the back-end device, the high-definition video data is pre-processed to obtain image data containing feature points; the feature points comprise the door bolt, the middle door latch and the small door lock; the image data containing the feature points is input into a preset deep learning model to obtain recognition results of each feature point; the recognition results of each feature point are integrated to obtain a door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least comprises a car number and a corresponding abnormal feature point.

[0037] This application calls a preset deep learning model to perform detection and processing on the high-definition video data of the train doors, and can automatically identify the loading and reinforcement status of the freight train doors. This eliminates the need for manual on-site inspections, saves departure time, and effectively ensures the safety of freight train operations.

[0038] The present application also provides a train door safety status detection system, which has the above-mentioned beneficial effects and will not be described in detail here. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0040] Figure 1 A flow chart of a method for detecting the safety status of a train door provided in an embodiment of the present application;

[0041] Figure 2 A schematic diagram of the structure of a train door safety status detection system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0042] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0043] Please refer to Figure 1 , Figure 1 A flow chart of a method for detecting the safety status of a train door provided in an embodiment of the present application, the method comprising:

[0044] S101: receiving a detection request and collecting high-definition video data of a train door;

[0045] This step is intended to receive a detection request and collect corresponding high-definition video data. The specific source of the detection request is not limited here, and it can be a detection request generated by an external device. For example, the detection request generated when the loading of a freight train is completed can be generated by a manual operation of the user, or it can be generated when a corresponding instruction is detected. It is easy to understand that this step is intended to require corresponding video acquisition equipment when it is executed, including but not limited to camera equipment, and can also be other equipment containing cameras, such as drones. The train doors include a middle door and a small door, the middle door includes a door bolt and a middle door latch, and the small door includes a small door lock;

[0046] S102: Transmitting the high-definition video data to a back-end processing device via the RTMP protocol;

[0047] RTMP is the abbreviation of Real Time Messaging Protocol. This step can achieve real-time synchronization of high-definition video data through the RTMP protocol. Of course, in the actual application of this step, the transmission of high-definition video data does not necessarily require real-time synchronization, and a corresponding delay can be set to ensure the stability of transmission.

[0048] S103: In the backend device, preprocessing the high-definition video data to obtain image data containing feature points;

[0049] This step is to pre-process the high-definition video data to obtain feature points, including the door bolt, middle door latch and small door lock.

[0050] This step does not limit how to perform preprocessing. The purpose of preprocessing is to achieve the conversion process of high-definition video data to image data. The image data that is blurred during the conversion process is optimized or discarded. A preferred preprocessing process can be as follows:

[0051] The first step is to obtain target image data including the middle door and / or the small door in the high-definition video data;

[0052] Step 2: Determine the clarity of the middle gate and / or small gate according to the image quality scoring rule to determine the image quality of each target image data;

[0053] Step 3: retain the target image data whose image quality is higher than the preset quality score.

[0054] That is, the high-definition video data is processed to obtain target image data containing the car door. Obviously, not all image frames in the high-definition video data contain the car door, so in the first step, the image frames that do not contain the car door need to be discarded to obtain the target image data. Thereafter, the image quality of each image frame is determined according to the image quality scoring rule. The image quality scoring rule is not specifically limited here, for example, it can be the following scoring rule:

[0055] 10 points: the marked area is complete, the door is in the middle of the picture, the whole picture is clear, there is no light spot or other impurities around it, and the marked area is clear;

[0056] 9 points: the marked area is complete, the door is basically located in the middle of the picture, each marked area can be clearly distinguished, and there are light spots or other impurities around it;

[0057] 8 points: the marked area is complete, the door position is slightly off, and the recognition area can be clearly identified;

[0058] 7 points: the marked area is complete, the door is slightly off, and there is a shadow in the identification area;

[0059] 6 points: the marked area is complete, the door position is biased, and the identification area is fuzzy;

[0060] 5 points: the door area is not fully marked, but the identification area is complete;

[0061] 4 points: the door area is not fully marked, the recognition area is complete, but the clarity of the recognition area is poor;

[0062] 3 points: the door area is not fully marked and the identification area is not complete;

[0063] 2 points: the door area is not fully marked, the identification area is not complete, and the clarity of the identification area is poor;

[0064] 1 point, the image is overall blurry;

[0065] 0 points, unable to identify.

[0066] Of course, those skilled in the art may also adopt other image quality scoring rules, which are not specifically limited here.

[0067] After that, the train carriages and doors in the high-definition video data are identified, that is, the middle door and the small door are identified. The specific process can be as follows:

[0068] S1031: parsing the high-definition video data, distinguishing carriages according to axles, and identifying the carriage number of each carriage;

[0069] S1032: Carriage identification is performed on each carriage, and the middle door and the small door are marked; image data containing the middle door and / or the small door is determined in the high-definition video data;

[0070] S1033: Taking the middle door and the small door as features, performing feature point recognition on the image data, and marking the obtained feature points.

[0071] That is, the high-definition video data is first analyzed to distinguish the carriages of the train and determine the corresponding carriage numbers. Then each carriage is identified to identify the middle door and small door of each carriage, and each middle door and small door is marked as a feature point. It should be noted that the identification method of each carriage and the marking method used to mark the feature points are not specifically limited here.

[0072] S104: inputting the image data containing the feature points into a preset deep learning model to obtain recognition results of each feature point;

[0073] S105: Integrate the recognition results of each feature point to obtain a door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least includes a car number and a corresponding abnormal feature point.

[0074] The embodiment of the present application calls a preset deep learning model to perform detection and processing on the high-definition video data of the train doors, so as to automatically identify the loading and reinforcement status of the freight train doors. This eliminates the need for manual on-site inspections, saves departure time, and effectively ensures the driving safety of freight trains.

[0075] The following describes the generation process of a preset deep learning model, which may include:

[0076] S201: Obtain learning samples;

[0077] S202: performing at least one of image resizing, image flipping, random erasing, automatic enhancement, and random patching on the learning sample to obtain a training sample;

[0078] S203: ResNet is used as the backbone network, connected to the common feature aggregation network, and common features are extracted from the convolution network containing feature point images in different scenes to obtain feature maps; the loss functions of the preset deep learning model include cross entropy loss, face recognition algorithm loss and triplet loss;

[0079] S204: Aggregate the feature graph to obtain a global vector;

[0080] S205: Normalize the global vector, perform head linear processing and head simplification processing to obtain a preset deep learning model.

[0081] The preset deep learning model uses ResNet as the backbone network, connects to the public feature aggregation network, and sends different detection item pictures from different scenes into the convolutional network to extract common features and aggregate and train network parameters. Of course, the preset deep learning model can also use ResNext and ResNeSt as the backbone network, which is not specifically limited here. Network parameters are important parameters required for model reasoning, and the trained network parameters are part of the model. Before reasoning, the registration of typical scene detection items can be completed first, and the pictures are matched in real time with the pictures in the pre-sorted standard picture library through model calculation, and the corresponding detection item category is inferred, and combined with the door loading reinforcement status standard, it is judged whether the detection item is a normal category or an abnormal category. Reasoning refers to sending the picture that needs to be inferred into the training model for reasoning, and a feature map can be obtained, so as to perform feature matching with the registered picture. It should be noted that the "model" described in this embodiment refers to the preset deep learning model.

[0082] The loss functions of the preset deep learning model include Cross-entropy loss, Arcface loss and Focalloss. The loss function is used to reflect the gap between the model prediction value and the true value. The network parameter training is adjusted by the size of the loss value to make the loss value smaller and smaller. The smaller the loss value, the better the model training. Then the three losses of Cross-entropy loss, ArcLoss and FocalLoss are calculated in parallel. The calculation process can also use a hyperparameter to determine the corresponding weight of each loss, thereby adjusting the network parameter training to make the model more accurate.

[0083] Cross-entropy loss can be defined as

[0084] Where C is the total number of categories.

[0085] (1) is the model prediction value, y i represents the true value.

[0086] in The cross entropy loss makes the logit predictions close to the ground truth. This makes the difference between the maximum logit and all other logits larger, which combined with the bounded gradient reduces the adaptability of the model, causing the model to be overconfident in its predictions, which in turn leads to overfitting. In order to improve the robustness of the model, a label smoothing method is introduced. The ground truth label of the cross entropy loss can be defined as y i (j=c)=1-δ and Among them, δ is a very small constant. This constant is introduced to make yi Not equal to 1 or 0.

[0087] Arcface loss maps Cartesian coordinates to spherical coordinates. Arcface loss converts logit into where θ i is the weight W i The angle between the feature f and the single weight ||W is fixed by l2 regularization i ||=1, and fix the embedding feature f through l2 regularization, and change the size to s. So In order to enhance both intra-class compactness and inter-class differences, Arcface adds an angular edge penalty m to the intra-class metric, that is, other factors are added to the intra-class metric rule (formula), so is rewritten as

[0088] Focal loss is mainly used to solve the problem of sample imbalance. The Focal Loss formula is: Where γ is the sample adjustment factor, which can be set by technicians in this field. When y=1, it is a positive sample (positive), and when y=0, it is a negative sample (negative). α is an adjustment constant. Generally, the more categories there are, the smaller α is set. For example, it can be 0.25. γ is an adjustment factor that adjusts the rate at which the weight of simple samples decreases. When γ increases, the impact of the adjustment factor also increases. γ can be set to 2. Where y′ represents the predicted value. Focal Loss reduces the weight of a large number of simple negative samples in model training, thereby making the model more accurate.

[0089] The recognition process of the preset deep learning model mainly consists of four processes: image preprocessing, backbone, aggregation, and head processing.

[0090] 1) Image preprocessing:

[0091] Image preprocessing mainly refers to image enhancement. The recognition model obtains more abundant learning samples through algorithms such as resizing, flipping, random erasing, auto-augmentation, random patching, and cutout, and inputs them into the network to obtain a more robust recognition model. For any image, each image enhancement operation is random and determined by a probability parameter.

[0092] 2) Backbone:

[0093] The ResNet backbone network of the preset deep learning model also adds attention non-local modules and instance batch normalization (IBN) modules to the backbone network to make the model more robust. The backbone network is fine-tuned and each channel of the feature map is processed. The channel represents the vector dimension, for example, RGB represents the three dimensions of the image w*h*c, which enables the backbone network to extract features more subtly, so that the network can learn the features of the image more accurately.

[0094] 3) Aggregation

[0095] The aggregation process aims to aggregate the feature maps generated by the backbone into a global feature. The recognition model adopts the attention pooling aggregation method. Input, generate a vector As the output of the pooling process, X, f and R represent different vectors, and W, H, C represent the width, height and number of channels of the feature map respectively. Common pooling methods include maximum pooling and average pooling, which are not specifically limited here. In the case of maximum pooling, the global vector f = [f1, ..., f c , ..., f C ], average pooling, pooling (attention pooling) is given by the following formula:

[0096] ω is the pooling parameter.

[0097] 4) Head

[0098] Head is the part that processes the global vector generated by the aggregation process, including batch normalization (BN) Head, linear Head and reduction head. Among them, the linear head contains only one decision layer, the BN head contains a BN layer and a decision layer, and the reduction head obtains the reduction layer through conv operation, bn operation, ReLU operation and Dropout operation, and then obtains the decision layer through batch normalization operation.

[0099] Since it is difficult to train models with saturated nonlinearity, batch normalization is introduced to address internal covariance shifts. Given a batch of feature vectors m is the number of samples in a batch, then the feature vector of the bn layer It can be calculated as:

[0100]

[0101] Here γ and β are trainable scale and shift parameters, and ∈ is a constant added to the mini-batch variance for numerical stability.

[0102] The simplification layer aims to convert high-dimensional features into low-dimensional features, for example, 2048 dimensions into 512 dimensions, while the decision layer outputs the probabilities of different categories and distinguishes different categories for subsequent model training.

[0103] After the Head process, the feature vector is output and then input into the defined loss function for calculation, so as to adjust the network parameters and make the model more accurate.

[0104] A train door safety status detection system provided in an embodiment of the present application is introduced below. The safety status detection system described below and the train door safety status detection method described above can refer to each other.

[0105] join Figure 2 , Figure 2 This is a schematic diagram of the structure of a train door safety status detection system provided in an embodiment of the present application. The present application also provides a train door safety status detection system, including:

[0106] The acquisition module is used to receive the detection request and acquire high-definition video data of the train door; the train door includes a middle door and a small door, the middle door includes a bolt and a middle door latch, and the small door includes a small door lock;

[0107] A data transmission module, used to transmit the high-definition video data to a back-end processing device via the RTMP protocol;

[0108] The back-end device is used to pre-process the high-definition video data to obtain image data containing feature points; the feature points include the door bolt, the middle door latch and the small door lock; the image data containing the feature points is input into a preset deep learning model to obtain the recognition result of each feature point; the recognition results of each feature point are integrated to obtain the door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least includes the car number and the corresponding abnormal feature point.

[0109] Based on the above embodiment, as a preferred embodiment, the backend device includes:

[0110] The image analysis unit is used to analyze the high-definition video data, distinguish the carriages according to the axles, and identify the carriage number of each carriage; identify the carriage of each carriage and mark the middle door and the small door; determine the image data containing the middle door and / or the small door in the high-definition video data; use the middle door and the small door as features, identify feature points of the image data, and mark the feature points.

[0111] Based on the above embodiments, as a preferred embodiment, the following may also be included:

[0112] The model training module is used to obtain learning samples; perform at least one of image resizing, image flipping, random erasing, automatic enhancement and random patching on the learning samples to obtain training samples; use ResNet as the backbone network, connect to the common feature aggregation network, and extract common features from the convolution network containing feature point images in different scenes to obtain feature maps; the loss function of the preset deep learning model includes cross entropy loss, face recognition algorithm loss and triplet loss; aggregate the feature maps to obtain a global vector; normalize, head linearize and head simplify the global vector to obtain a preset deep learning model.

[0113] The present application also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed, the steps provided in the above embodiment can be implemented. The storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0114] The present application also provides a server, which may include a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps provided in the above embodiment may be implemented. Of course, the server may also include various network interfaces, power supplies and other components.

[0115] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the system provided in the embodiment, since it corresponds to the method provided in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0116] Specific examples are used herein to illustrate the principles and implementation methods of the present application, and the description of the above embodiments is only used to help understand the method and core ideas of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present application, several improvements and modifications can be made to the present application, and these improvements and modifications also fall within the scope of protection of the claims of the present application.

[0117] It should also be noted that, in this specification, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

Claims

1. A method for detecting the safety status of a train door, characterized in that: include: Receive detection requests and collect high-definition video data of train doors; The train door comprises a middle door and a small door, the middle door comprises a door bolt and a middle door latch, and the small door comprises a small door lock catch; Transmitting the high-definition video data to a back-end processing device via the RTMP protocol; In the back-end device, the high-definition video data is pre-processed to obtain image data containing feature points; the feature points include the door bolt, the middle door latch and the small door lock; Inputting the image data containing the feature points into a preset deep learning model to obtain recognition results of each feature point; Integrate the recognition results of each feature point to obtain a door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least includes a car number and a corresponding abnormal feature point; The generation process of the preset deep learning model includes: Get learning samples; Performing at least one of image resizing, image flipping, random erasing, automatic enhancement, and random patching on the learning sample to obtain a training sample; ResNet is used as the backbone network, connected to the common feature aggregation network, and common features are extracted from the convolutional network containing feature point images in different scenes to obtain feature maps; the loss function of the preset deep learning model includes cross entropy loss, face recognition algorithm loss and triplet loss; Aggregating the feature maps to obtain a global vector; The global vector is normalized, subjected to head linear processing and head simplification processing to obtain a preset deep learning model.

2. The safety status detection method according to claim 1, characterized in that: In the back-end device, the high-definition video data is preprocessed to obtain image data containing feature points, including: Parsing the high-definition video data, distinguishing carriages according to axles, and identifying the carriage number of each carriage; Carriage identification is performed on each of the carriages, and a middle door and a small door are marked; and image data containing the middle door and / or the small door is determined in the high-definition video data; The middle door and the small door are used as features, feature point recognition is performed on the image data, and feature points are obtained by marking.

3. The safety status detection method according to claim 1, characterized in that: Also includes: An attention-type non-local module and an IBN module are added to the backbone network.

4. The safety status detection method according to claim 1, characterized in that: Preprocessing the high-definition video data includes: Acquire target image data including the middle door and / or the small door in the high-definition video data; Determine the image quality of each of the target image data by determining the clarity of the middle gate and / or the small gate according to an image quality scoring rule; The target image data whose image quality is higher than a preset quality score is retained.

5. A train door safety status detection system, characterized in that: include: The acquisition module is used to receive the detection request and collect the high-definition video data of the train door; The train door comprises a middle door and a small door, the middle door comprises a door bolt and a middle door latch, and the small door comprises a small door lock catch; A data transmission module, used to transmit the high-definition video data to a back-end processing device via the RTMP21024025 protocol; A back-end device is used to pre-process the high-definition video data to obtain image data containing feature points; the feature points include the door bolt, the middle door latch and the small door lock; Input the image data containing the feature points into a preset deep learning model to obtain recognition results of each feature point; integrate the recognition results of each feature point to obtain a door detection result; wherein, when there is a safety anomaly in the train door, the door detection result at least includes a car number and a corresponding abnormal feature point; The model training module is used to obtain learning samples; perform at least one of image resizing, image flipping, random erasing, automatic enhancement and random patching on the learning samples to obtain training samples; use ResNet as the backbone network, connect to the common feature aggregation network, and extract common features from the convolution network containing feature point images in different scenes to obtain feature maps; the loss function of the preset deep learning model includes cross entropy loss, face recognition algorithm loss and triplet loss; aggregate the feature maps to obtain a global vector; normalize, head linearize and head simplify the global vector to obtain a preset deep learning model.

6. The safety status detection system according to claim 5, characterized in that: The backend equipment includes: The image analysis unit is used to analyze the high-definition video data, distinguish the carriages according to the axles, and identify the carriage number of each carriage; identify the carriage of each carriage and mark the middle door and the small door; determine the image data containing the middle door and / or the small door in the high-definition video data; use the middle door and the small door as features, identify feature points of the image data, and mark the feature points.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the security status detection method according to any one of claims 1 to 4 are implemented.

8. A server, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor calls the computer program in the memory, the steps of the security status detection method according to any one of claims 1 to 4 are implemented.