An intelligent lamp group control system based on artificial intelligence

Abstract

This invention relates to the field of lighting control technology, and more particularly to an intelligent lighting control system based on artificial intelligence, comprising: a lighting fixture and personnel identification module, which identifies the position of the lighting fixture, the position of the personnel, and the status of the personnel by using an image of the lighting area acquired by a visual sensor and a YOLO-based multi-object detection model; a personnel depth identification module, which identifies the vertical distance between the personnel and the lighting fixtures by using a monocular depth estimation model based on an improved Unet network architecture; a lighting fixture selection module, which determines the selected lighting fixture closest to the personnel by using a lighting fixture selection model based on a K-means clustering algorithm, based on the planar position map of the lighting fixture and the vertical distance; and a lighting fixture brightness control module, which sets the desired brightness according to the personnel status and controls the brightness of the selected lighting fixture according to the desired brightness. This invention achieves intelligent indoor lighting control, improves the lighting comfort of indoor personnel, and reduces lighting energy consumption.

Description

Technical Field

[0001] This invention relates to the field of lighting control technology, and in particular to an intelligent lighting control system based on artificial intelligence. Background Technology

[0002] Unified, non-distributed control of lighting fixtures using non-invasive indoor personnel positioning technology is currently a research hotspot. For example, while infrared personnel positioning offers high accuracy, it suffers from poor penetration through obstacles, short effective transmission distance, and high installation and deployment costs. Ultrasonic personnel positioning is susceptible to the Doppler effect and temperature fluctuations. Therefore, researchers believe that with the popularization and advancement of artificial intelligence technology, personnel positioning methods based on video image recognition, with their advantages of low cost, wide applicability, and ease of operation, should become the mainstream development direction of the lighting industry.

[0003] For example, the invention patent with application number 201680066164.1 discloses image-based lighting control, but it does not use an image recognition method based on artificial intelligence technology for lamp control.

[0004] Therefore, how to locate personnel using artificial intelligence-based image recognition and control the brightness of lighting groups with multiple lights is a technical problem that needs to be solved. Summary of the Invention

[0005] Therefore, this invention provides an intelligent lighting control system based on artificial intelligence. It achieves fast and high-precision identification of the location of people in the lighting fixtures through the YOLO single-stage multi-target detection model. It avoids the high hardware cost of using depth cameras for depth detection by improving the monocular depth estimation model of the Unet network architecture. It realizes intelligent energy-saving control of lighting fixtures through the Kmeans clustering algorithm, thereby realizing intelligent indoor lighting control, improving the lighting comfort of people in the room and reducing lighting energy consumption.

[0006] To achieve the above objectives, this invention proposes an intelligent lighting control system based on artificial intelligence, in which multiple lamps and a vision sensor form an intelligent lighting group, comprising:

[0007] The lighting fixture personnel identification module is used to identify the position of the lighting fixture, the position of the personnel, and the status of the personnel by using the image of the lighting area captured by the visual sensor through a YOLO-based multi-object detection model, and to generate a lighting fixture personnel planar position map based on the position of the lighting fixture and the position of the personnel.

[0008] A personnel depth recognition module is connected to the personnel position recognition module of the lighting fixture. It is used to identify the vertical distance between the personnel and the lighting fixture by using a monocular depth estimation model based on an improved Unet network architecture in the image of the area illuminated by the lighting group, and to mark the vertical distance on the personnel planar position map of the lighting fixture.

[0009] The lighting fixture selection module is connected to the personnel depth recognition module and is used to determine the selected lighting fixture by using the personnel planar position map of the lighting fixture and the vertical distance through the lighting fixture selection model, wherein the lighting fixture selection model is constructed based on the Kmeans clustering algorithm;

[0010] The lighting brightness control module is connected to the lighting selection module and the lighting personnel position recognition module, and is used to set the desired brightness according to the personnel status, and to control the brightness of the selected lighting fixture according to the desired brightness.

[0011] Furthermore, the personnel location recognition module for lighting fixtures includes a pre-training unit, a transfer learning unit, and a recognition unit, and the multi-target detection model includes an initial detection model, a pre-trained detection model, and a transfer model;

[0012] The pre-training unit is used to pre-train the initial detection model using a first dataset to generate the pre-trained detection model, wherein the first dataset is an unobstructed people dataset and a lighting dataset.

[0013] The transfer learning unit is used to fine-tune the network layers and output layer neurons of the pre-trained detection model using a second dataset to generate the transfer model, wherein the second dataset includes a person occlusion dataset, a person action dataset, and a person brightness dataset.

[0014] The recognition unit is used to identify the position of the lamps, the position of the personnel, and the status of the personnel in the image of the area illuminated by the lamp group through the migration model, and to generate a planar position map of the lamps and personnel based on the position of the lamps and the position of the personnel.

[0015] Furthermore, the identification unit includes an improved cross-union ratio calculation subunit;

[0016] The improved cross-union ratio (CUP) calculation subunit is used to train and test the pre-trained detection model using the improved CUP to generate the transfer model, wherein the improved CUP is generated by scaling the original CUP and calculating the average value based on a set upper limit and a set lower limit.

[0017] Furthermore, the improved intersection-union ratio (CUP) calculation subunit includes a scaled CUP node, an average value node, and a loss term iteration node;

[0018] The scaling crossover ratio node is used to divide the difference between the original crossover ratio and the median value of the set limit by the difference between the set upper limit value and the set lower limit value to obtain the scaling crossover ratio; the average value node is used to calculate the average value of the scaling crossover ratio and the original crossover ratio, and use the average value as the adjustment loss term; the loss term iteration node is used to iterate the output loss term according to the adjustment loss term.

[0019] In the above scheme, by utilizing the YOLO-based multi-object detection model based on transfer learning, the required amount of dataset is reduced, the training speed is accelerated, and the accuracy and generalization ability are improved. By improving the intersection-union ratio, a more accurate and constrained model penalty is provided, which enables the predicted boxes to better regress to the ground truth boxes. This, in turn, better guides the multi-object detection model to learn more accurate predicted boxes and bounding boxes.

[0020] Furthermore, the monocular depth estimation model includes an encoder, a decoder, and an improved residual network, and the personnel depth recognition module includes a residual connection generation unit;

[0021] The residual connection generation unit is used to connect each level of the encoder to each level of the decoder through the improved residual network, and to reduce the number of connections in the improved residual network through pruning operations to generate the monocular depth estimation model for identifying the vertical distance.

[0022] Furthermore, the encoder is provided with a first-level coding convolutional layer to a fourth-level coding convolutional layer in sequence along the encoding order, and the decoder is provided with a first-level decoding convolutional layer to a fourth-level decoding convolutional layer in sequence along the decoding order;

[0023] The output of the first-level encoding convolutional layer is connected to the inputs of the second-level decoding convolutional layer, the third-level decoding convolutional layer, and the fourth-level decoding convolutional layer through the improved residual network;

[0024] The output of the second-level encoding convolutional layer is connected to the input of the second-level decoding convolutional layer and the input of the third-level decoding convolutional layer through the improved residual network.

[0025] The output of the third-level encoding convolutional layer is connected to the input of the first-level decoding convolutional layer and the input of the second-level decoding convolutional layer through the improved residual network.

[0026] The fourth-level coding convolutional layer is connected to the input of the first-level decoding convolutional layer through the improved residual network.

[0027] Furthermore, the improved residual network has three levels of residual networks connected in sequence, each level of residual network summing the features of the outputs of the large convolutional pooling layer and the outputs of the small convolutional layer.

[0028] In the above scheme, by improving the monocular depth estimation model of the Unet network architecture, the feature fusion effect of the model is improved, the number of parameters is reduced, and the model can make full use of feature information at different scales.

[0029] Furthermore, the lamp selection module includes a planar distance calculation unit, a planar clustering unit, and a vertical nearest lamp selection unit;

[0030] The planar distance calculation unit is used to calculate a set of planar distances reflecting the distance between personnel and lighting fixtures based on the personnel planar position map;

[0031] The planar clustering unit is used to perform K-means clustering operation on the planar distance set to generate a cluster set reflecting the planar classification of people and lighting fixtures;

[0032] The vertical nearest lamp selection unit is used to perform calculations on the cluster set to select the lamp that is vertically closest to the person and determine it as the selected lamp.

[0033] In the above scheme, the method of selecting lamps through two Kmeans clustering operations achieves a fast, easy-to-implement lamp selection that meets the lighting comfort requirements, thereby realizing intelligent energy-saving lamp operation.

[0034] Furthermore, the personnel states include sleeping, looking down, standing, and discussing, with the corresponding desired brightness increasing sequentially.

[0035] The lamp brightness control module is used to perform fuzzy control on the selected lamp based on the desired brightness and the actual brightness collected by the vision sensor.

[0036] Furthermore, multiple lighting fixtures communicate with one of the aforementioned visual sensors via NBIoT, using the OneNET cloud platform as the main node in an IoT network. The OneNET cloud platform stores and displays the planar position map of the lighting fixtures and the vertical distance.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. The YOLO single-stage multi-target detection model enables fast and high-precision identification of personnel positions in lighting fixtures. The improved Unet network architecture monocular depth estimation model avoids the high hardware costs associated with depth camera depth detection. The Kmeans clustering algorithm enables intelligent energy-saving control of lighting fixtures, thereby achieving intelligent indoor lighting control, improving the lighting comfort of indoor personnel, and reducing lighting energy consumption.

[0039] 2. By utilizing transfer learning in a YOLO-based multi-object detection model, the required dataset size is reduced, the training speed is accelerated, and the accuracy and generalization ability are improved. By improving the intersection-union ratio, a more accurate and constrained model penalty is provided, enabling the predicted boxes to better regress to the ground truth boxes. This, in turn, better guides the multi-object detection model to learn more accurate predicted boxes and bounding boxes.

[0040] 3. By improving the monocular depth estimation model of the Unet network architecture, the feature fusion effect of the model is improved, the number of parameters is reduced, and the model can make full use of feature information at different scales.

[0041] 4. By using two K-means clustering operations to select luminaires, a fast, easy-to-implement luminaire selection that meets lighting comfort requirements is achieved, thereby realizing intelligent energy-saving luminaire switching. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the structure of an artificial intelligence-based intelligent lighting control system according to an embodiment of the present invention;

[0043] Figure 2 This is a flowchart illustrating an artificial intelligence-based intelligent lighting control system according to an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the structure of a monocular depth estimation model of the improved Unet network architecture of the intelligent lighting control system based on artificial intelligence according to an embodiment of the present invention.

[0045] Figure 4 This is a schematic diagram of the improved residual network structure of the monocular depth estimation model of the improved Unet network architecture of the intelligent lighting control system based on artificial intelligence according to an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0047] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0048] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0049] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0050] like Figures 1 to 4 As shown, this invention provides an intelligent lighting control system based on artificial intelligence. It achieves fast and high-precision identification of the location of people in the lighting fixtures through the YOLO single-stage multi-target detection model. It avoids the high hardware cost of using depth cameras for depth detection by improving the monocular depth estimation model of the Unet network architecture. It realizes intelligent energy-saving control of lighting fixtures through the Kmeans clustering algorithm, thereby realizing intelligent indoor lighting control, improving the lighting comfort of people in the room and reducing lighting energy consumption.

[0051] like Figures 1 to 4 As shown, this embodiment proposes an intelligent lighting control system based on artificial intelligence, in which multiple lamps and a vision sensor form an intelligent lighting group, including:

[0052] The lighting fixture personnel identification module is used to identify the position of the lighting fixture, the position of the personnel, and the status of the personnel by using the image of the lighting area captured by the visual sensor through a YOLO-based multi-object detection model, and to generate a lighting fixture personnel planar position map based on the position of the lighting fixture and the position of the personnel.

[0053] A personnel depth recognition module, connected to the lighting personnel recognition module, is used to identify the vertical distance between personnel and lighting fixtures in the image of the area illuminated by the lighting group through a monocular depth estimation model based on an improved Unet network architecture, and to mark the vertical distance on the personnel planar position map of the lighting fixture;

[0054] The lighting fixture selection module is connected to the personnel depth recognition module and is used to determine the selected lighting fixture by using the personnel planar position map of the lighting fixture and the vertical distance through the lighting fixture selection model, wherein the lighting fixture selection model is constructed based on the Kmeans clustering algorithm;

[0055] The lighting brightness control module is connected to the lighting selection module and the lighting personnel identification module, and is used to set the desired brightness according to the personnel status, and to control the brightness of the selected lighting fixture according to the desired brightness.

[0056] Understandably, YOLO (You Only Look Once) is a fast and highly accurate single-stage image recognition model. Its core idea is to transform the object detection task into a regression problem, thereby directly predicting bounding boxes and their class probabilities at multiple locations in an image. Therefore, YOLO is applied to multi-object detection of people and lighting fixtures. Monocular depth estimation involves predicting the depth value of each pixel from a detection point in a single RGB image. When the vision sensor and lighting fixture are integrated on the same horizontal plane or above the horizontal plane of the lighting fixture, this depth value reflects the vertical distance. The UNet network architecture, with its multi-scale feature fusion, end-to-end training, and strong adaptability, is very suitable for monocular depth estimation, thus demonstrating powerful performance and high generalization ability. KMeans is a clustering algorithm based on iterative optimization. Its core idea is to divide the dataset into multiple clusters, ensuring that each data point belongs to the cluster corresponding to its nearest cluster center, thereby minimizing the sum of squared errors within the cluster. This achieves simple and efficient clustering calculations, making the activation of lighting fixtures more intelligent. YOLO and UNet network architectures, as deep learning algorithms, and KMeans, as an unsupervised machine learning algorithm, both belong to the field of artificial intelligence.

[0057] Furthermore, such as Figure 2 As shown, the lighting personnel recognition module includes a pre-training unit, a transfer learning unit, and a recognition unit; the multi-target detection model includes an initial detection model, a pre-trained detection model, and a transfer model.

[0058] The pre-training unit is used to pre-train the initial detection model using a first dataset to generate the pre-trained detection model, wherein the first dataset is an unobstructed people dataset and a lighting dataset.

[0059] The transfer learning unit is used to fine-tune the network layers and output layer neurons of the pre-trained detection model using a second dataset to generate the transfer model, wherein the second dataset includes a person occlusion dataset, a person action dataset, and a person brightness dataset.

[0060] The recognition unit is used to identify the position of the lamps, the position of the personnel, and the status of the personnel in the image of the area illuminated by the lamp group through the migration model, and to generate a planar position map of the lamps and personnel based on the position of the lamps and the position of the personnel.

[0061] Understandably, the training of object detection algorithms heavily relies on large amounts of data, and the collection and production of large-scale datasets are expensive. Transfer learning can use an existing dataset to fine-tune the model on a related new task, thereby reducing the size of the dataset and ensuring model performance.

[0062] Specifically, the first dataset is the MS COCO dataset, a large object detection dataset released by Microsoft. It includes images of people and lighting fixtures, but the images are relatively simple and conventional, making them unsuitable for identifying people and lighting fixtures under varying brightness and environmental conditions. The second dataset is a dataset of images collected from actual lighting environments. For example, in a classroom setting, images of people in various postures (sleeping, standing, discussing, looking down, etc.) and with various occlusions, as well as images of lighting fixtures of different types, angles, and sizes, are collected. The personnel and lighting fixture images are labeled using the LabelImg tool, generating a corresponding XML configuration file for the second dataset. The 3600 labeled images in the second dataset are divided into two subsets: 70% is used to train the multi-object detection model, and 30% is used to test the training effect of the multi-object detection model.

[0063] Specifically, the vision sensor uses a distortion-free industrial camera CMOS sensor with a resolution of 1920×1080.

[0064] Furthermore, the identification unit includes an improved cross-union ratio (CUP) calculation subunit; the improved CUP calculation subunit is used to train and test the pre-trained detection model using the improved CUP to generate the transfer model, wherein the improved CUP is generated by scaling the original CUP and calculating the average value based on a set upper limit and a set lower limit.

[0065] Furthermore, the improved intersection-union ratio (CUP) calculation subunit includes a scaled CUP node, an average value node, and a loss term iteration node;

[0066] The scaling crossover ratio node is used to divide the difference between the original crossover ratio and the median value of the set limit by the difference between the set upper limit value and the set lower limit value to obtain the scaling crossover ratio; the average value node is used to calculate the average value of the scaling crossover ratio and the original crossover ratio, and use the average value as the adjustment loss term; the loss term iteration node is used to iterate the output loss term according to the adjustment loss term.

[0067] Specifically, the intermediate value of the set limit is one-half of the difference between the set upper limit and the set lower limit.

[0068] Specifically, the calculation process for the improved crossover-union ratio is as follows:

[0069]

[0070] L′ iou =1, IOU>b

[0071]

[0072] loss IOUnew =L iou +loss IOUold

[0073] In the formula, IOU is the original intersection-union ratio, S1 and S2 are the areas of the intersection and union of the ground truth box and the prediction box, respectively, and L′ is the sum of the areas of the intersection and union of the ground truth box and the prediction box, respectively. iou The scaled cross-union ratio is obtained by scaling the original cross-union ratio. a and b are the upper and lower limits, respectively, preferably 0.05 and 0.97. iou The adjustment loss term is a linear adjustment term that is used to calculate the average of the scaled intersection-union ratio and the original intersection-union ratio. IOUnew loss IOUold These are the current output loss term and the previous prediction box output loss term, respectively.

[0074] Understandably, according to the original Cross-Union Ratio (CUI) calculation formula, the original CUI is between 0 and 1. Scaling the CUI increases its value by scaling it to the region of the middle value of the set limit, thereby enhancing the loss term. This results in a larger output loss term and a stronger penalty effect when the difference between the ground truth bounding box and the predicted bounding box is large. This allows for better constraint on the positional relationship between the model's predicted bounding box and the ground truth bounding box, avoiding overfitting or underfitting. By linearly adjusting the loss term, it helps to improve the accuracy of the object detection model and reduce missed detections.

[0075] Specifically, considering the characteristics of small target size and dense distribution of personnel images collected by the light group, and the problems of insufficient resolution, weak detail extraction ability and poor prediction effect of dense targets in the YOLOv4 algorithm, this embodiment preferably adopts a multi-target detection model based on YOLOv5.

[0076] In the above scheme, by utilizing the YOLO-based multi-object detection model based on transfer learning, the required amount of dataset is reduced, the training speed is accelerated, and the accuracy and generalization ability are improved. By improving the intersection-union ratio, a more accurate and constrained model penalty is provided, which enables the predicted boxes to better regress to the ground truth boxes. This, in turn, better guides the multi-object detection model to learn more accurate predicted boxes and bounding boxes.

[0077] Furthermore, such as Figure 3 and 4As shown, the monocular depth estimation model includes an encoder, a decoder, and an improved residual network. The personnel depth recognition module includes a residual connection generation unit. The residual connection generation unit is used to connect each level of the encoder to each level of the decoder through the improved residual network, and to reduce the number of connections in the improved residual network through pruning operations to generate the monocular depth estimation model for recognizing the vertical distance.

[0078] Furthermore, such as Figure 3 As shown, the encoder has a first-level coding convolutional layer to a fourth-level coding convolutional layer in sequence along the encoding order, and the decoder has a first-level decoding convolutional layer to a fourth-level decoding convolutional layer in sequence along the decoding order;

[0079] The output of the first-level encoding convolutional layer is connected to the inputs of the second-level decoding convolutional layer, the third-level decoding convolutional layer, and the fourth-level decoding convolutional layer through the improved residual network;

[0080] The output of the second-level encoding convolutional layer is connected to the input of the second-level decoding convolutional layer and the input of the third-level decoding convolutional layer through the improved residual network.

[0081] The output of the third-level encoding convolutional layer is connected to the input of the first-level decoding convolutional layer and the input of the second-level decoding convolutional layer through the improved residual network.

[0082] The fourth-level coding convolutional layer is connected to the input of the first-level decoding convolutional layer through the improved residual network.

[0083] Understandably, the UNet network architecture only performs multi-scale feature extraction at the encoder end, and then gradually upsamples at the decoder end to restore the original image size. Therefore, by redesigning the network connections, the model can fully utilize feature information at different scales. This is achieved by having the decoder learn which layers of features from the encoder are useful, and by using pruning operations to obtain... Figure 3 The network structure shown.

[0084] Specifically, from the first to the fourth level of the encoder's encoding convolutional layers, the first convolutional layer of each level uses a 3×3 kernel with a stride of 1 and 64 output channels. The activation function is LeakyReLU, which adds non-linearity and allows for some negative value propagation. The second convolutional layer of each level has 128 output channels and uses the same 3×3 kernel and LeakyReLU activation. Therefore, through two convolutions, low-level features of the image are extracted layer by layer, including edges, textures, and shapes, ensuring rich feature information. The LeakyReLU activation function allows the network to capture more complex feature patterns, avoids neuron death problems, and improves the network's learning ability.

[0085] Next, the Multilayer Perceptron (MLP) generates dynamic convolutional kernels based on the input feature map, using 1×1 convolutions to generate multiple kernels. These generated kernels are then used to convolve the input feature map, adapting to different input features and thus enhancing the network's adaptability to features. The max-pooling layer uses a 2×2 pooling window and a stride of 2 to reduce the spatial resolution of the feature map. The maximum value within the pooling window preserves the most significant feature information, enhancing the representativeness of the features.

[0086] The first to fourth levels of encoding convolutional layers aim to further enhance the model's learning capabilities. As the number of parameters increases, the convolutional layers can learn more abstract and complex features. These features may involve a wider range of visual patterns, or even be understood as a high-level description of the entire object or scene, preserving the most salient feature information and laying the foundation for feature enhancement paths.

[0087] The feature maps obtained after passing through the first to fourth levels of encoding convolutional layers are enhanced with an improved residual network to increase the network's focus on key features. By improving the combination of the residual network and the decoder, the network can better understand the key information in the input image, which is crucial for improving the model's accuracy and quality.

[0088] Upsampling is typically used to restore or increase the spatial resolution of feature maps, enabling better feature fusion and object detection. During downsampling, the spatial size of the feature map shrinks, resulting in the loss of some detailed information. Upsampling helps recover these details, thus improving the spatial resolution of the feature map.

[0089] Furthermore, such as Figure 4 As shown, the improved residual network has three levels of improved residual networks connected in sequence. Each level of the improved residual network adds the features of the outputs of the large convolutional pooling layer and the outputs of the small convolutional layer.

[0090] Specifically, such as Figure 4 As shown, the large convolutional pooling layer uses a 3x3 convolutional kernel and the activation function is ReLU, while the small convolutional layer uses a 1x1 convolutional kernel.

[0091] Understandably, since the conventional UNet network architecture uses only one convolutional layer to implement residual connections, it cannot make good use of spatial and semantic information. Therefore, the residual connection method is redesigned. This improvement to the residual network can not only improve the feature fusion effect, but also further reduce the number of parameters.

[0092] In the above scheme, by improving the monocular depth estimation model of the Unet network architecture, the feature fusion effect of the model is improved, the number of parameters is reduced, and the model can make full use of feature information at different scales.

[0093] Furthermore, the lamp selection module includes a planar distance calculation unit, a planar clustering unit, and a vertical nearest lamp selection unit;

[0094] The planar distance calculation unit is used to calculate a set of planar distances reflecting the distance between personnel and lighting fixtures based on the personnel planar position map;

[0095] The planar clustering unit is used to perform K-means clustering operation on the planar distance set to generate a cluster set reflecting the planar classification of people and lighting fixtures;

[0096] The vertical nearest lamp selection unit is used to perform calculations on the cluster set to select the lamp that is vertically closest to the person and determine it as the selected lamp.

[0097] Specifically, the planar distance calculation unit calculates the distance from a person to each lamp using a conventional distance calculation formula to generate the planar distance set. The set of lamps, the set of people, and the planar distance set are input into a K-means clustering algorithm. The number of clusters is set to 5, the maximum number of iterations to 2, and the iteration termination threshold to 10. The K-means clustering algorithm groups the people and lamps into 5 categories based on the planar distance set. Then, in each element of the cluster set, the lamp that is vertically closest to the person is found, designated as the selected lamp, and output.

[0098] More specifically, the loss function of the K-means clustering algorithm is set as follows:

[0099]

[0100] In the formula, J is the iterative output of the K-means clustering algorithm, the summation times of i are 5, the number of clusters is N, and N is the summation number of i. i d represents the number of samples contained in the i-th K-means clustering algorithm cluster. j Let u be the planar distance between the j-th person and the light fixture. j Let J be the cluster center of the j-th cluster in the K-means clustering algorithm. It can be understood that when the difference between the output J of two iterations is less than the iteration termination threshold of 10, the algorithm stops iterating and uses the cluster at this point as the final clustering result.

[0101] The final clustering results are used to determine the selected luminaires using the following formula:

[0102] C = min i∈ { 1,2,3......k}dist(d j,u j )

[0103] In the formula, C represents the selected lighting fixture, k is the number of clusters in the K-means clustering algorithm, and d j Let u be the planar distance between the j-th person and the light fixture. j Let `dist()` be the cluster center of the j-th cluster in the K-means clustering algorithm, and `dist()` is a function that calculates the planar distance and the vertical distance corresponding to the cluster center. Therefore, we can determine the closest light fixture to the person in the same cluster, and then select that closest light fixture as the chosen light fixture.

[0104] Furthermore, the personnel states include sleeping, looking down, standing, and discussing, with the corresponding desired brightness increasing sequentially; the lighting brightness control module is used to perform fuzzy control on the selected lighting fixture based on the desired brightness and the actual brightness collected by the visual sensor.

[0105] Specifically, the indoor illuminance is generally between 50 and 300 lux (Lux). Therefore, in this embodiment, the basic domain of illuminance is taken as [70, 350]. When the person's state is sleeping, looking down, standing, or discussing, the expected brightness is 70, 100, 280, and 350 lux (Lux) respectively. The basic domain of the error e between the actual brightness and the expected brightness can be obtained as [-230, 300], the basic domain of the error change rate is [-30, 30], and the basic domain of the output control quantity is [10%, 100%]. By fuzzifying the error and error change rate on the corresponding basic domains, corresponding fuzzy linguistic variables are obtained. The fuzzy linguistic variables and the output control quantity are then subjected to fuzzy control inference through fuzzy control rules to obtain the fuzzy control quantity. The fuzzy control quantity is then defuzzified to obtain the output control quantity, thus realizing different control precision and control speed of the lamp brightness according to different states of the person.

[0106] Understandably, the YOLO-based multi-object detection model outputs an image of the illuminated area of ​​the lighting group, labeled with multiple bounding boxes. Each bounding box labels a single person and a single light fixture. The single person is categorized into one of the following states: sleeping, looking down, standing, or discussing. The brightness of a single light fixture is determined using a brightness conversion formula based on its RGB image, and this brightness is then used to control the light fixture. Therefore, the lighting control system in this embodiment only requires a single vision sensor to achieve intelligent control of lighting on / off and brightness for multiple people simultaneously, reducing its implementation cost.

[0107] Furthermore, multiple lighting fixtures communicate with one of the aforementioned visual sensors via NBIoT, using the OneNET cloud platform as the main node in an IoT network. The OneNET cloud platform stores and displays the planar position map of the lighting fixtures and the vertical distance.

[0108] Specifically, the lighting fixtures and vision sensors, as terminal devices, exchange data with the cloud platform through NB-IoT wireless communication technology. In this design, NB-IoT adopts the CoAP transparent transmission mode to establish a connection with the OneNET cloud platform that supports the CoAP protocol, thus achieving a stable network connection.

[0109] It is understandable that the single-stage multi-object detection model using YOLO described in this embodiment achieves fast and high-precision identification of people and lighting fixtures. The improved monocular depth estimation model using the Unet network architecture avoids the high hardware costs associated with depth camera-based depth detection. Intelligent energy-saving lighting control is achieved through K-means clustering, thereby realizing intelligent indoor lighting control, improving lighting comfort for indoor occupants, and reducing lighting energy consumption. By utilizing a YOLO-based multi-object detection model based on transfer learning, the required dataset size is reduced, training speed is accelerated, and accuracy and generalization ability are improved. An improved intersection-over-union ratio (IoU) provides a more accurate, constrained model penalty, allowing predicted boxes to better regress to ground truth boxes, thus better guiding the multi-object detection model to learn more accurate predicted and bounding boxes. The improved monocular depth estimation model using the Unet network architecture enhances feature fusion, reduces the number of parameters, and allows the model to fully utilize feature information at different scales. The method of selecting lighting fixtures through two K-means clustering operations achieves fast, easy-to-implement lighting selection that meets lighting comfort requirements, thereby realizing intelligent energy-saving lighting.

[0110] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent lighting control system based on artificial intelligence, characterized in that, A smart lighting system is composed of multiple lamps and vision sensors, including: The lighting fixture personnel identification module is used to identify the position of the lighting fixture, the position of the personnel, and the status of the personnel by using the image of the lighting area captured by the visual sensor through a YOLO-based multi-object detection model, and to generate a lighting fixture personnel planar position map based on the position of the lighting fixture and the position of the personnel. A personnel depth recognition module, connected to the lighting personnel recognition module, is used to identify the vertical distance between personnel and lighting fixtures in the image of the area illuminated by the lighting group through a monocular depth estimation model based on an improved Unet network architecture, and to mark the vertical distance on the personnel planar position map of the lighting fixture; The lighting fixture selection module is connected to the personnel depth recognition module and is used to determine the selected lighting fixture by using the personnel planar position map of the lighting fixture and the vertical distance through the lighting fixture selection model, wherein the lighting fixture selection model is constructed based on the Kmeans clustering algorithm; A lighting brightness control module is connected to the lighting selection module and the lighting personnel recognition module, and is used to set the desired brightness according to the personnel status, and to control the brightness of the selected lighting fixture according to the desired brightness. The monocular depth estimation model includes an encoder, a decoder, and an improved residual network; the personnel depth recognition module includes a residual connection generation unit. The residual connection generation unit is used to connect each level of the encoder to each level of the decoder through the improved residual network. The improved residual network has three levels of improved residual networks connected in sequence. Each level of the improved residual network adds the features of the output of the large convolutional pooling layer and the output of the small convolutional layer respectively. The encoder is provided with a first-level coding convolutional layer to a fourth-level coding convolutional layer in the coding order, and the decoder is provided with a first-level decoding convolutional layer to a fourth-level decoding convolutional layer in the decoding order. The output of the first-level encoding convolutional layer is connected to the inputs of the second-level decoding convolutional layer, the third-level decoding convolutional layer, and the fourth-level decoding convolutional layer through the improved residual network; The output of the second-level encoding convolutional layer is connected to the input of the second-level decoding convolutional layer and the input of the third-level decoding convolutional layer through the improved residual network. The output of the third-level encoding convolutional layer is connected to the input of the first-level decoding convolutional layer and the input of the second-level decoding convolutional layer through the improved residual network. The fourth-level coding convolutional layer is connected to the input of the first-level decoding convolutional layer through the improved residual network.

2. The intelligent lighting control system based on artificial intelligence according to claim 1, characterized in that, The personnel recognition module for lighting fixtures includes a pre-training unit, a transfer learning unit, and a recognition unit; the multi-target detection model includes an initial detection model, a pre-trained detection model, and a transfer model. The pre-training unit is used to pre-train the initial detection model using a first dataset to generate the pre-trained detection model, wherein the first dataset is an unobstructed people dataset and a lighting dataset. The transfer learning unit is used to fine-tune the network layers and output layer neurons of the pre-trained detection model using a second dataset to generate the transfer model, wherein the second dataset includes a person occlusion dataset, a person action dataset, and a person brightness dataset. The recognition unit is used to identify the position of the lamps, the position of the personnel, and the status of the personnel in the image of the area illuminated by the lamp group through the migration model, and to generate a planar position map of the lamps and personnel based on the position of the lamps and the position of the personnel.

3. The intelligent lighting control system based on artificial intelligence according to claim 2, characterized in that, The identification unit includes an improved cross-union ratio calculation subunit; The improved cross-union ratio (CUP) calculation subunit is used to train and test the pre-trained detection model using the improved CUP to generate the transfer model, wherein the improved CUP is generated by scaling the original CUP and calculating the average value based on a set upper limit and a set lower limit.

4. The intelligent lighting control system based on artificial intelligence according to claim 3, characterized in that, The improved cross-union ratio (CUP) calculation subunit includes a scaled CUP node, an average value node, and a loss term iteration node; The scaled intersection-union ratio node is used to divide the difference between the original intersection-union ratio and the median value of the set limit by the difference between the set upper limit value and the set lower limit value to obtain the scaled intersection-union ratio. The average value node is used to calculate the average of the scaled cross-union ratio and the original cross-union ratio, and use the average value as the adjustment loss term; the loss term iteration node is used to iterate the output loss term according to the adjustment loss term.

5. The intelligent lighting control system based on artificial intelligence according to claim 1, characterized in that, The residual connection generation unit is used to reduce the number of connections in the improved residual network through pruning operations to generate the monocular depth estimation model for identifying the vertical distance.

6. The intelligent lighting control system based on artificial intelligence according to claim 1, characterized in that, The lighting fixture selection module includes a planar distance calculation unit, a planar clustering unit, and a vertical nearest lighting fixture selection unit; The planar distance calculation unit is used to calculate a set of planar distances reflecting the distance between personnel and lighting fixtures based on the personnel planar position map; The planar clustering unit is used to perform K-means clustering operation on the planar distance set to generate a cluster set reflecting the planar classification of people and lighting fixtures; The vertical nearest lamp selection unit is used to perform calculations on the cluster set to select the lamp that is vertically closest to the person and determine it as the selected lamp.

7. The intelligent lighting control system based on artificial intelligence according to any one of claims 1 to 6, characterized in that, The personnel states include sleeping, head down, standing, and discussing, with the corresponding desired brightness increasing sequentially. The lamp brightness control module is used to perform fuzzy control on the selected lamp based on the desired brightness and the actual brightness collected by the vision sensor.

8. The intelligent lighting control system based on artificial intelligence according to any one of claims 1 to 6, characterized in that, Multiple lighting fixtures communicate with a visual sensor via NBIoT, with the OneNET cloud platform as the main node. The OneNET cloud platform stores and displays the planar position map of the lighting fixtures and the vertical distance.

Images