Elevator group control system scheduling method based on improved YOLOv10

Abstract

The invention relates to an elevator group control system scheduling method based on improved YOLOv10, and belongs to the technical field of intelligent elevator control and computer vision crossing. The elevator group control system scheduling method comprises the following steps that S1, pedestrian images of indoor elevator corridor scenes are collected, and a training set and a test set are constructed; s2, the YOLOv10 network is improved; s3, training the improved YOLOv10 network by using the training set to obtain a pedestrian detection model, and testing by using the test set; s4, the trained pedestrian detection model is used for actual pedestrian detection in the elevator floor; and S5, counting the number of pedestrians on each floor according to a detection result, and dynamically adjusting an elevator scheduling strategy according to the distribution of the number of pedestrians on high and low floors. According to the method, the detection capability of the network on the shielded pedestrians is enhanced, the pedestrian detection precision is improved, and the operation efficiency of the elevator and the riding experience of passengers are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of intelligent elevator control and computer vision, and relates to an elevator group control system scheduling method based on improved YOLOv10. Background Art

[0002] In modern buildings, as vertical transportation tools, the scheduling efficiency of elevators directly affects the usage experience and operation efficiency of buildings. Especially in high - rise buildings or public places with large passenger flows, such as office buildings, shopping malls, hospitals, etc., the elevator scheduling strategy plays a crucial role in reducing waiting time and improving transportation capacity. Traditional elevator scheduling methods often rely on fixed schedules or simple floor request queues, and it is difficult to adapt to the dynamic changes of the actual passenger flow, resulting in low elevator operation efficiency and long passenger waiting times.

[0003] With the development of computer vision and intelligent control technologies, methods for dynamically adjusting elevator scheduling strategies by detecting the number of people on each floor through cameras have emerged. However, existing pedestrian detection methods still have some limitations. Existing detection methods related to machine vision include: R - CNN, Fast R - CNN, Faster R - CNN for two - step detection, which can obtain good accuracy through step - by - step detection, but the speed is relatively slow; YOLO series, SSD, CornerNet, etc. for one - step detection, which greatly increases the speed through end - to - end methods, but there is a partial loss of accuracy compared with two - step detection methods. In the elevator scheduling scenario, due to the influence of passenger occlusion factors, some features of passengers are missing, resulting in poor detection effects of pedestrian detection algorithms.

[0004] In summary, existing elevator scheduling systems have problems such as low target detection accuracy, poor detection effects for occluded pedestrians, and inflexible and non - dynamic scheduling strategies. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide an elevator group control system scheduling method based on improved YOLOv10. By improving the YOLOv10 pedestrian detection algorithm, enhancing its detection ability for complex scenarios and occluded pedestrian targets, and applying the improved algorithm to the elevator group control system, real - time and dynamic adjustment of the elevator scheduling strategy is realized. The present invention aims to improve the scheduling efficiency of elevators, reduce the waiting time of passengers, and enhance the overall operation performance of the elevator system and the riding experience of passengers.

[0006] To achieve the above - mentioned purpose, the present invention provides the following technical solutions:

[0007] An elevator group control system scheduling method based on improved YOLOv10, comprising the following steps:

[0008] S1: Collect pedestrian images in the indoor elevator corridor scene to construct a training set and a test set;

[0009] S2: Improve the YOLOv10 network;

[0010] S3: Use the training set to train the improved YOLOv10 network to obtain a pedestrian detection model, and use the test set for testing;

[0011] S4: Apply the trained pedestrian detection model to the actual pedestrian detection in the elevator floors;

[0012] S5: According to the detection results, count the number of pedestrians on each floor, and dynamically adjust the elevator dispatching strategy according to the number distribution of high and low floors.

[0013] Furthermore, step S1 specifically includes the following steps:

[0014] Collect floor images through surveillance cameras installed in the indoor elevator corridor;

[0015] Extract images frame by frame from the collected video and perform preprocessing;

[0016] Label the pedestrians in each image and label the category to which the corresponding area belongs. The categories include partially occluded pedestrians and unoccluded pedestrians; form a pedestrian detection data set with the labeled images and label files;

[0017] Divide the pedestrian detection data set into a training set and a test set.

[0018] Furthermore, step S1 also includes: statistically analyzing and screening the images and annotations in the data set, screening out extremely small or extremely large annotation boxes, and verifying the existing errors in the annotations.

[0019] Furthermore, step S2 specifically includes the following steps:

[0020] S21: Modify the C2f module and the C2fCIB module in the YOLOv10 network structure, and refer to and combine the receptive field attention convolution in RFAConv to construct a C2f_RFAConv feature extraction module to replace the C2f and C2fCIB modules in the backbone network of the original YOLOv10 model; use C2f_RFAConv to construct a new backbone network for extracting pedestrian features in the image;

[0021] S22: Add a feature pyramid detection layer and a detection channel to the original feature fusion structure of YOLOv10; the feature pyramid detection layer is obtained by splicing and fusing a shallow feature map containing detailed information and a deep feature map with rich context information after continuous upsampling;

[0022] S23: Introduce the MultiSEAM attention mechanism into the downsampling part of the YOLOv10 feature fusion structure to optimize the feature fusion structure of the network. The MultiSEAM attention mechanism enhances the learning of key features of pedestrians and suppresses unimportant occluded interference information through the calculation of attention weights;

[0023] After passing through the S23 MultiSEAM attention mechanism module, input the feature map into the YOLOv10 detection head; detect pedestrians in each frame of the video image to obtain pedestrian position information.

[0024] Furthermore, constructing the new backbone network in step S21 includes the following steps:

[0025] Analyze the original backbone network structure of YOLOv10 and extract its main modules C2f and C2fCIB; the C2f module contains a convolutional layer, a batch normalization layer BN, and an activation function; the C2fCIB module incorporates depthwise separable convolution and an attention mechanism module on the basis of the C2f module to improve the feature extraction ability of the model and the attention to important information;

[0026] Modify the C2f module. By introducing the receptive field attention convolution RFAConv, replace the second convolutional layer on the Bottleneck convolution branch in the C2f feature extraction module with RFAConv. The modified network module is named C2f_RFAConv;

[0027] Use the C2f_RFAConv module to replace the C2f and C2fCIB modules in the backbone network of the YOLOv10 object detection network module by module to obtain different new backbone networks. According to the number of modules replaced in different stages, name it RFAConvNet i , i = 1, 2, 3, 4, where i = 1 represents that only one module is replaced, that is, only the first C2f module in the original backbone network of YOLOv10 is correspondingly replaced with the C2f_RFAConv module; i = 4 represents that all C2f and C2fCIB modules in YOLOv10 are replaced with the C2f_RFAConv module;

[0028] RFAConvNet i The overall architecture of the backbone network is the same as that of the original YOLOv10 backbone network CSPDNet, and there are outputs in four stages. The four stages contain features of different depths, and these four features act on the subsequent feature fusion structure.

[0029] Furthermore, step S22 specifically includes the following steps:

[0030] S221: The underlying feature map in the backbone network is processed by the SPPF feature pyramid and the PSA attention layer of YOLOv10 to output the processed feature map F;

[0031] S222: F is input into the YOLOv10 feature fusion structure. After upsampling, it is concatenated with the feature map of the corresponding scale output by the 3rd C2f_RFAConv module in the backbone network to obtain a new feature map F2;

[0032] S223: After F2 is upsampled, it is concatenated again with the feature map of the corresponding scale output by the second C2f_RFAConv in the backbone network to obtain a new feature map F3;

[0033] S224: Add an upsampling layer to the original YOLOv10 feature fusion structure. After F3 is upsampled by this layer, it is concatenated with the feature map of the corresponding scale output by the first C2f_RFAConv in the backbone network to obtain a new feature map F4;

[0034] S225: F4 is used as the feature map in the additionally added feature pyramid detection layer and is input into the subsequent detection head.

[0035] Further, step S23 specifically includes the following steps:

[0036] S231: Add a MultiSEAM attention block after each C2f and C2fCIB module that outputs detection results in the downsampling part of the YOLOv10 feature fusion structure;

[0037] S232: The result output by the previous MultiSEAM attention block is downsampled and then concatenated with the feature map of the corresponding scale in the upsampling part of the feature fusion to obtain a new feature map, which is used as the input of the next MultiSEAM attention block;

[0038] S233: Input the feature map output by each MultiSEAM attention block in S231 into the YOLOv10 detection head.

[0039] Further, step S3 includes the following steps:

[0040] S31: Input the training set into the improved YOLOv10 network for feature extraction and fusion;

[0041] S32: Input the extracted feature result map into the YOLOv10 detection head for classification and prediction;

[0042] S33: In the YOLOv10 detection head, predict and output the final result map, use logistic regression to screen and score the prior boxes on the final result map, and label the obtained prediction results to complete the training process of the pedestrian detection model;

[0043] S34: Use the test set to test the trained pedestrian detection model.

[0044] Further, step S4 includes the following steps:

[0045] S41: Load the trained pedestrian detection model into the improved YOLOv10 network;

[0046] S42: Apply the YOLOv10 network loaded with the pedestrian detection model to the actual pedestrian detection task in the elevator floor; by inputting the image or video data in the elevator floor, identify and label the positions of pedestrians, so as to achieve the goal of pedestrian detection.

[0047] Further, the specific steps of counting the number of pedestrians in the pedestrian detection result and adjusting the elevator dispatching strategy according to the number of people in step S5 are as follows:

[0048] S51: According to the position information of the pedestrian detection result, count the number of pedestrians on each floor;

[0049] S52: Dynamically adjust the elevator dispatching strategy according to the number of pedestrians on each floor:

[0050] When it is the up-peak period, the elevator executes the up-peak period dispatching strategy;

[0051] When it is the down-peak period, the elevator executes the down-peak period dispatching strategy.

[0052] The beneficial effects of the present invention are as follows: The present invention proposes an elevator dispatching method based on improved YOLOv10. This method is particularly suitable for high-rise buildings or public places with large traffic flow. It can efficiently detect the number of pedestrians on different floors and dynamically adjust the elevator dispatching strategy according to the real-time detection results, significantly improving the operation efficiency of the elevator and the riding experience of passengers.

[0053] The present invention improves the YOLOv10 object detection algorithm for the indoor crowded scene with complex background and pedestrian occlusion problems. By introducing receptive field convolution and using the attention mechanism module to enhance the attention to features, important information can be noticed even in complex backgrounds, enhancing the network's detection ability for occluded pedestrians, thereby improving the pedestrian detection accuracy.

[0054] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the examination of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. Brief Description of the Drawings

[0055] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0056] Figure 1 It is a flowchart of the elevator group control system scheduling method based on the improved YOLOv10 according to the present invention;

[0057] Figure 2 It is the overall network structure diagram of the improved YOLOv10;

[0058] Figure 3 They are the structure diagrams of C2f, C2fCIB, and C2f_RFAConv that make up the backbone network;

[0059] Figure 4 It is the structure diagram of the MultiSEAM module;

[0060] Figure 5 It is a flowchart of the dynamic elevator scheduling scheme executed according to the detected number of people. Detailed Embodiment

[0061] The following illustrates the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention schematically. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0062] It should be noted that the drawings provided in the following embodiments only illustrate the basic concept of the present invention schematically. Therefore, only the components related to the present invention are shown in the drawings, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and proportion of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0063] In the following description, numerous details are explored to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0064] As Figure 1 shown, the present invention proposes an elevator scheduling method based on improved YOLOv10, including the following steps:

[0065] S1: Collect pedestrian images in the indoor elevator corridor scene to construct a training set and a test set;

[0066] S2: Improve the YOLOv10 network;

[0067] S3: Use the training set to train the improved YOLOv10 network to obtain a pedestrian detection model, and use the test set for testing;

[0068] S4: Apply the trained pedestrian detection model to the actual pedestrian detection in the elevator floors;

[0069] S5: According to the detection results, count the number of pedestrians on each floor, and dynamically adjust the elevator scheduling strategy according to the number distribution of high and low floors.

[0070] In one embodiment, in step S1, collecting pedestrian images in the indoor elevator corridor scene to construct a training set and a test set specifically includes the following steps:

[0071] S11: Obtain the equipment required for images of different floors of the building, including: monitoring equipment, terminal computer.

[0072] Installation of the monitoring equipment: Under the condition of meeting the requirements of the monitoring target's field of view range, the installation height of the monitoring equipment is required to be not less than 2.5 m from the ground indoors;

[0073] Optionally, signal lines and power lines should be introduced separately, and the exposed parts need to be protected by hoses and do not affect rotation.

[0074] Installation of the terminal computer: All controls, displays, and records of the monitoring equipment need to be displayed on the terminal computer. The terminal computer should be installed stably and be easy to operate. The computer screen should avoid direct sunlight from the outside. When it is unavoidable, light-shielding measures should be taken;

[0075] Optionally, the equipment installed in the console and cabinet should have ventilation and heat dissipation measures, and the internal connectors should be firmly connected to the equipment;

[0076] After installation is completed, ensure that the terminal computer can receive the floor pedestrian images transmitted by the monitoring device.

[0077] S12: Label and filter the collected images, which mainly include the following steps:

[0078] S121: Select a labeling tool;

[0079] Optionally, use labelimg to label the pictures, and select a rectangular box to label the target objects. Among them, the target objects at least include partially occluded pedestrians and unoccluded pedestrians.

[0080] Optionally, after all the images are labeled, save and generate a YOLO format label file. The generated YOLO format label file can be directly used for model training.

[0081] S122: Use the selected labeling tool to select human targets from each collected image, and label the corresponding categories to obtain an information dataset;

[0082] S123: Screen and view the information data, and the categories at least include information of partially occluded and unoccluded pedestrians.

[0083] In one embodiment, as Figure 2 shown, step S2 includes the following steps:

[0084] S21: Modify the C2f and C2fCIB modules in the YOLOv10 backbone network according to the receptive field attention convolution in RFAConv, and refer to and combine the Bottleneck structure of YOLOv4 to construct a new C2f_RFAConv structure for extracting pedestrian features of images;

[0085] S22: Add a feature pyramid detection layer and a detection channel to the original feature fusion structure of YOLOv10. This feature pyramid detection layer is obtained by splicing and fusing the shallow feature map output by the first C2f_RFAConv module in the backbone network and the deep feature map after continuous upsampling of the bottom layer of the backbone network. The shallow feature map contains more detailed information, and the deep feature map contains more semantic information;

[0086] S23: Add a MultiSEAM attention block after the C2f and C2fCIB modules of each output detection result in the downsampling part of the YOLOv10 feature fusion structure to optimize the feature fusion structure of the network. The MultiSEAM attention mechanism can effectively enhance the model's learning of pedestrian key features and suppress unimportant occlusion interference information through the calculation of attention weights.

[0087] S24: After passing through the S23 MultiSEAM attention mechanism module, the feature map is input into the YOLOv10 detection head. Detect pedestrians in each frame of the video to obtain pedestrian position information.

[0088] Optionally, as Figure 3 shown, step S21 constructs C2f_RFAConv, which specifically includes the following steps:

[0089] S211: Analyze the original backbone network structure of YOLOv10 and extract its main modules C2f and C2fCIB. The C2f module contains basic components such as convolutional layers, batch normalization layers (BN), and activation functions (such as Leaky ReLU). By stacking these basic components, the C2f module realizes effective feature extraction of the input image. As an improved version of the C2f module, the C2fCIB module further integrates depthwise separable convolution and attention mechanism modules on the basis of maintaining the original efficient feature extraction ability, thus enhancing the model's feature extraction ability and attention to important information;

[0090] S212: Replace the second convolutional layer (Conv) on the Bottleneck convolution branch in the C2f feature extraction module of the original YOLOv10 model with RFAConv to construct the Bottleneck_RFAConv structure;

[0091] S213: Use Bottleneck_RFAConv to replace the Bottleneck structure and CIB structure in the C2f and C2fCIB modules of the original YOLOv10 backbone network to construct the C2f_RFAConv feature extraction module;

[0092] S214: Use the C2f_RFAConv module to replace the C2f and C2fCIB modules in the backbone network of the YOLOv10 object detection network module by module to obtain different new backbone networks. According to the number of modules replaced in different stages, name them RFAConvNeti (i = 1, 2, 3, 4), where i = 1 means only one module is replaced, that is, only the first C2f module in the original YOLOv10 backbone network is correspondingly replaced with the C2f_RFAConv module; i = 4 means all C2f and C2f_CIB modules in YOLOv10 are replaced with the C2f_RFAConv module.

[0093] Optionally, step S22 specifically includes the following steps:

[0094] S221: The underlying feature map (dimension 20*20*1024) in the backbone network is processed by the SPPF feature pyramid and the PSA attention layer of YOLOv10, and then the processed feature map F (dimension 20*20*1024) is output;

[0095] S222: F is input into the YOLOv10 feature fusion structure. After upsampling, it is concatenated with the feature map of the corresponding scale output by the 3rd C2f_RFAConv module in the backbone network to obtain a new feature map F2 (dimension 40*40*512);

[0096] S223: After upsampling, F2 is concatenated again with the feature map of the corresponding scale output by the second C2f_RFAConv in the backbone network to obtain a new feature map F3 (dimension 80*80*256);

[0097] S224: An upsampling layer is added to the original YOLOv10 feature fusion structure. After F3 passes through this upsampling, it is concatenated with the feature map of the corresponding scale output by the first C2f_RFAConv in the backbone network to obtain a new feature map F4 (dimension 160*160*128);

[0098] S225: F4 is used as the feature map in the additionally added feature pyramid detection layer and is input into the subsequent detection head.

[0099] Optionally, step S23 specifically includes the following steps:

[0100] S231: A MultiSEAM attention block is added after each C2f and C2fCIB module that outputs detection results in the downsampling part of the YOLOv10 feature fusion structure;

[0101] S232: The result output by the previous MultiSEAM attention block is downsampled and then concatenated with the feature map of the corresponding scale in the upsampling part of the feature fusion to obtain a new feature map, which is used as the input of the next MultiSEAM attention block;

[0102] S233: The feature map output by each MultiSEAM attention block in S231 is input into the YOLOv10 detection head.

[0103] Optionally, as Figure 4 shown, the feature fusion steps after the feature map is input into the MultiSEAM attention block in step S23 are as follows:

[0104] The first stage: The feature map Fx input to the MultiSEAM attention block is input to three different patch deep convolutions (DCovN). DCovN first performs preliminary feature extraction through a standard two-dimensional convolutional layer (Conv). Subsequently, batch normalization (BatchNorm) and the SiLU activation function are applied to increase non-linearity;

[0105] The second stage: Next, the DCovN module consists of a combination of multiple depthwise separable convolutions and residual connections. Each combination implements depthwise separable convolution through a grouped convolution, and then the SiLU activation function and batch normalization are applied. The residual connection adds the output of the grouped convolution to the previous feature map to maintain the flow of information;

[0106] The third stage: The output feature maps of the DCovN with 3 different Patch sizes and the input feature map of the input MultiSEAM are respectively passed through a global average pooling layer (Avgpool), then concatenated and combined and divided by 4 to obtain the fused feature map Y;

[0107] The fourth stage: The fused feature map Y is input into two linearly connected fully connected layers to generate attention weights. The attention weights are multiplied by the input feature map after channel expansion (exp) to obtain the output feature map Fs of the final MultiSEAM.

[0108] Optionally, the result output by the previous MultiSEAM attention block, after downsampling, is concatenated with the corresponding scale feature map of the upsampling part of the feature fusion to obtain a new feature map, and the new feature map is used as the input of the next MultiSEAM attention block. The output of each MultiSEAM is used as a detection feature map and is sequentially input into the YOLOv10 detection head.

[0109] Optionally, the feature map is classified and regressed through the YOLOv10 detection head to complete the construction of the improved YOLOv10 network for pedestrian detection. The improved YOLOv10 network is uniformly named RM-YOLO.

[0110] In one embodiment, step S3 includes the following steps:

[0111] S31: Use the public dataset for training and the self-annotated dataset for testing. Statistically screen and check the images and annotations of pedestrians in the corridor in the dataset, filter out extremely small or large annotation boxes, verify the existing errors in the annotations and modify and improve them;

[0112] S32: For the convenience of test consistency, batch modify the annotation category of the public dataset to "person";

[0113] S33: Set the parameters required for training: number of training epochs: epoch = 100, input batch size each time: batch_size = 4, size of training images: train_img_size = 608, momentum: momentum = 0.9, IoU threshold required in the loss function: Iou_threshold_loss = 0.5, weight decay rate: weight_decay = 0.0005, confidence threshold: conf_thresh = 0.005, multi-scale validation: multi_scale_val = True;

[0114] Optionally, it also includes the setting of classes and anchors. The class setting {"num": 1, "classes": ['person']}, and the anchors obtained by clustering are: [[(11, 13), (15, 18), (18, 21),], [(23, 27), (31, 36), (45, 51),], [(69, 80), (117, 138), (258, 296)]], strides: [8, 16, 32], output scale: anchors_pre_scale = 4.

[0115] S34: Input the data image normalized to the scale of 608×608×3 into a 3×3 convolution to convert the image channels from 3 to 32, and then enter the backbone network and feature fusion network in the RM-YOLO constructed in S2 for feature extraction and fusion;

[0116] S35: Input the feature map after fusion into the YOLOv10 detection head, predict and output the final result map, and use logistic regression to screen the prior boxes on the final result map, and complete the training of pedestrian detection through backpropagation.

[0117] In one embodiment, in step S4, the improved YOLOv10 network is used to apply the trained pedestrian detection model to actual pedestrian detection in the elevator floor. The specific steps are as follows:

[0118] S41: In actual detection, information such as the detection model, log path, gpu / cpu setting, and image path to be detected needs to be loaded, and multiple bounding boxes are predicted through the model;

[0119] S42: Further, screening is performed through a dual label assignment strategy to obtain the final bounding boxes. The screened bounding boxes and related class information are directly drawn on the image to obtain the detection result.

[0120] In one embodiment, the specific steps of step S5 are as follows:

[0121] S51: According to the position information of the pedestrian detection results, count the number of pedestrians on each floor, determine the floor attribution of the detected pedestrians, and count the number of pedestrians on each floor. When specifically implemented, a time window can be set, such as every minute or every five minutes, to count the pedestrian detection results within this time window to obtain the distribution of the number of pedestrians on each floor.

[0122] S52: During the upward peak period, passengers are mainly concentrated in the lower floors, and the number of people in the lower floors is much larger than that in the upper floors. At this time, the present invention executes the upward peak period scheduling strategy.

[0123] S53: During the downward peak period, passengers are mainly concentrated in the upper floors, and the number of people in the upper floors is much larger than that in the lower floors. At this time, the present invention executes the downward peak period scheduling strategy.

[0124] As Figure 5 shown, taking the elevator group control system of three elevators as an example according to the number distribution, the process of dynamically adjusting the elevator scheduling strategy is as follows:

[0125] Step 1: Detect the number of pedestrians on each floor through RM - YOLO;

[0126] Step 2: Determine whether the number of pedestrians on each floor is greater than the maximum capacity of a single elevator;

[0127] Step 3: If the number of people on the lower floors is more than twice that on the upper floors, at this time it is the upward peak period, and execute the upward peak period scheduling strategy. Prioritize scheduling the two elevators closer to the lower floors to go to the lower floor area to meet the needs of upward passengers. The remaining one elevator is used for downward demand or as a standby elevator to handle possible special situations;

[0128] If the number of people on the upper floors is more than twice that on the lower floors, at this time it is the downward peak period, and execute the downward peak period scheduling strategy. Prioritize scheduling the two elevators closer to the upper floors to go to the upper floor area to meet the needs of downward passengers. The remaining one elevator is used for upward demand or as a standby elevator to handle possible special situations.

[0129] In the above embodiments, the mention of "this embodiment" in the specification means that the specific features, structures or characteristics described in conjunction with the embodiments are included in at least some embodiments, but not necessarily all embodiments. The multiple occurrences of "this embodiment" do not necessarily all refer to the same embodiment.

[0130] In the above embodiments, although the present invention has been described in conjunction with specific embodiments of the present invention, many alternatives, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art based on the foregoing description. For example, other storage structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. Embodiments of the present invention are intended to cover all such alternatives, modifications, and variations that fall within the broad scope of the appended claims.

[0131] This embodiment also provides a computer-readable storage medium having stored thereon a computer program, which when executed by a processor implements any one of the methods in this embodiment.

[0132] This embodiment also provides an electronic terminal, comprising: a processor and a memory;

[0133] The memory is used for storing a computer program, and the processor is used for executing the computer program stored in the memory so that the terminal executes any one of the methods in this embodiment.

[0134] For the computer-readable storage medium in this embodiment, those of ordinary skill in the art can understand that all or part of the steps for implementing the above method embodiments can be completed by hardware related to a computer program. The foregoing computer program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: various media such as ROM, RAM, magnetic disk, or optical disc that can store program code.

[0135] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication therebetween. The memory is used for storing a computer program, and the communication interface is used for communication. The processor and the transceiver are used for running the computer program so that the electronic terminal executes each step of the above method.

[0136] In this embodiment, the memory may include a random access memory (Random Access Memory, abbreviated as RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0137] The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field-programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0138] The present invention can be used in numerous general-purpose or special-purpose computing system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multi-processor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and so on.

[0139] The present invention can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention can also be practiced in a distributed computing environment where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A dispatching method for an elevator group control system based on improved YOLOv10, characterized in that: The following steps are involved: S1: Collect pedestrian images in indoor elevator corridor scenes and construct training and test sets; S2: Improve the YOLOv10 network; S3: Use the training set to train the improved YOLOv10 network to obtain a pedestrian detection model, and use the test set for testing; S4: Use the trained pedestrian detection model for actual pedestrian detection in elevator floors; S5: According to the detection results, the number of pedestrians on each floor is counted, and the elevator dispatching strategy is dynamically adjusted according to the distribution of people on high and low floors.

2. The elevator group control system scheduling method based on improved YOLOv10 according to claim 1, characterized in that: Step S1 specifically includes the following steps: Collect floor images through surveillance cameras installed in indoor elevator corridors; Extract images frame by frame from the collected video and perform preprocessing; Label the pedestrians in each image and label the categories to which the corresponding areas belong, including partially obscured pedestrians and unobstructed pedestrians; the labeled images and label files form a pedestrian detection dataset; The pedestrian detection dataset is divided into training set and test set.

3. The elevator group control system scheduling method based on improved YOLOv10 according to claim 1, characterized in that: Step S1 also includes: performing statistics and screening on the images and annotations in the data set, screening out extremely small or extremely large annotation boxes, and checking for errors in the annotations.

4. The elevator group control system scheduling method based on improved YOLOv10 according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21: Modify the C2f module and C2fCIB module in the YOLOv10 network structure, and build the C2f_RFAConv feature extraction module with reference to and in combination with the receptive field attention convolution in RFAConv, replacing the C2f and C2fCIB modules in the original YOLOv10 model backbone network; use C2f_RFAConv to build a new backbone network for extracting pedestrian features in images; S22: adding a feature pyramid detection layer and a detection channel to the original feature fusion structure of YOLOv10; the feature pyramid detection layer is obtained by concatenating and fusing a shallow feature map containing detail information and a deep feature map having rich context information after continuous upsampling; S23: The MultiSEAM attention mechanism is introduced into the sampling part of the YOLOv10 feature fusion structure to optimize the feature fusion structure of the network. The MultiSEAM attention mechanism enhances the learning of key features of pedestrians and suppresses unimportant occlusion interference information through the calculation of attention weights. S24: After passing through the S23MultiSEAM attention mechanism module, the feature map is input into the YOLOv10 detection head; pedestrians in each frame of the video are detected to obtain the pedestrian position information.

5. The elevator group control system scheduling method based on improved YOLOv10 according to claim 4 is characterized in that: The construction of a new backbone network in step S21 includes the following steps: Analyze the original backbone network structure of YOLOv10 and extract its main modules C2f and C2fCIB; the C2f module includes a convolutional layer, a batch normalization layer BN, and an activation function; the C2fCIB module integrates the deep separable convolution and attention mechanism modules on the basis of the C2f module to improve the model's feature extraction capability and attention to important information; The C2f module is modified by introducing the receptive field attention convolution RFAConv, replacing the second convolution layer on the BottleNeck convolution branch in the C2f feature extraction module with RFAConv. The modified network module is named C2f_RFAConv. Use the C2f_RFAConv module to replace the C2f and C2fCIB modules in the backbone network of the YOLOv10 target detection network module by module to obtain different new backbone networks. According to the number of modules replaced at different stages, they are named RFAConvNet. i , i=1,2,3,4, where i=1 means that only one module is replaced, that is, only the first C2f module in the original YOLOv10 backbone network is replaced with the C2f_RFAConv module; i=4 means that all C2f and C2fCIB modules in YOLOv10 are replaced with the C2f_RFAConv module; RFAConvNet i The overall architecture of the backbone network is the same as that of the original YOLOv10 backbone network CSPDNet. There are outputs in four stages, and the four stages contain features of different depths. These four features act on the subsequent feature fusion structure.

6. The elevator group control system scheduling method based on improved YOLOv10 according to claim 4 is characterized in that: Step S22 specifically includes the following steps: S221: The underlying feature map in the backbone network is processed by the SPPF feature pyramid and the PSA attention layer of YOLOv10, and then the processed feature map F is output; S222: F is input into the YOLOv10 feature fusion structure, and after upsampling, it is concatenated with the feature map of the corresponding scale output by the third C2f_RFAConv module in the backbone network to obtain a new feature map F2; S223: After upsampling, F2 is concatenated with the feature map of the corresponding scale output by the second C2f_RFAConv in the backbone network to obtain a new feature map F3; S224: add an upsampling layer to the original YOLOv10 feature fusion structure, and after the upsampling, F3 is concatenated with the feature map of the corresponding scale output by the first C2f_RFAConv in the backbone network to obtain a new feature map F4; S225: F4 is used as a feature map in an additional feature pyramid detection layer for input into subsequent detection heads.

7. The elevator group control system scheduling method based on improved YOLOv10 according to claim 4 is characterized in that: Step S23 specifically includes the following steps: S231: Add a MultiSEAM attention block after each C2f and C2fCIB module that outputs the detection result in the downsampling part of the YOLOv10 feature fusion structure; S232: The result output by the previous MultiSEAM attention block is downsampled and concatenated with the feature map of the corresponding scale of the upsampled part of the feature fusion to obtain a new feature map, which is used as the input of the next MultiSEAM attention block; S233: Input the feature map output by each MultiSEAM attention block in S231 into the YOLOv10 detection head.

8. The elevator group control system scheduling method based on improved YOLOv10 according to claim 1, characterized in that: Step S3 includes the following steps: S31: Input the training set into the improved YOLOv10 network for feature extraction and fusion; S32: Input the extracted feature result graph into the YOLOv10 detection head for classification and prediction; S33: In the YOLOv10 detection head, the final result image is predicted and output, and the prior boxes on the final result image are screened and scored using logistic regression, and the predicted results are annotated to complete the training process of the pedestrian detection model; S34: Use the test set to test the trained pedestrian detection model.

9. The elevator group control system scheduling method based on improved YOLOv10 according to claim 1, characterized in that: Step S4 includes the following steps: S41: Load the trained pedestrian detection model into the improved YOLOv10 network; S42: Apply the YOLOv10 network loaded with the pedestrian detection model to the actual task of pedestrian detection in the elevator floor; by inputting the image or video data in the elevator floor, identify and mark the location of the pedestrians, so as to achieve the goal of pedestrian detection.

10. The elevator group control system scheduling method based on improved YOLOv10 according to claim 1, characterized in that: In step S5, the specific steps of counting the number of pedestrians detected and adjusting the elevator dispatching strategy according to the number of pedestrians are as follows: S51: Counting the number of pedestrians on each floor according to the location information of the pedestrian detection result; S52: Dynamically adjust the elevator dispatching strategy based on the number of pedestrians on each floor: When the upward peak occurs, the elevator implements the upward peak scheduling strategy; When the down-bound traffic is at its peak, the elevator executes the down-bound peak dispatching strategy.

Images