An integrated intelligent lamp pole monitoring device for personnel operation management of a farm greenhouse

Abstract

The present application relates to the technical field of intelligent lamp pole equipment, and particularly relates to an integrated intelligent lamp pole monitoring device for operation management of a farm greenhouse, which comprises an integrated intelligent lamp pole and a monitoring device installed on the integrated intelligent lamp pole, wherein the monitoring device comprises an intelligent monitoring module, an LED lamp management control module, a photoresistor module, a temperature monitoring module, an air conditioner module, a humidity monitoring module, a humidifier module, a carbon dioxide monitoring module, a soil pH value monitoring module, a solar panel module and a main control module; the intelligent monitoring module collects normal light image information and infrared image information in the greenhouse through a visual sensor, processes the collected normal light image information by using a MOL-POSE algorithm, and provides the distance of a detection target with the aid of an infrared image. The present application is used for assisting farmers in judging whether a person in the greenhouse is performing reasonable operation on the greenhouse, and maintaining the internal environment of the greenhouse in a state most suitable for the growth of internal plants, so as to realize the monitoring integration of the farm greenhouse.

Description

Technical Field

[0001] This invention relates to the field of intelligent light pole equipment technology, and in particular to an integrated intelligent light pole monitoring device for personnel operation and management in farm greenhouses. Background Technology

[0002] Agriculture is a vital pillar industry for the nation. As the world's largest agricultural country, my country's agricultural production plays a crucial role in its economic construction and social development. However, there are still many areas in my country's greenhouse environmental control system that need improvement. Research on intelligent greenhouse monitoring systems involves multiple technologies and disciplines, including machine vision, sensor technology, control technology, communication technology, biotechnology, and environmental science.

[0003] Therefore, this application proposes an integrated intelligent light pole monitoring device for the operation and management of farm greenhouse personnel to solve the above-mentioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an integrated intelligent light pole monitoring device for farm greenhouse operation and management. This device assists farmers in determining whether people inside the greenhouse are operating the greenhouse properly and maintains the internal environment of the greenhouse in the most suitable state for plant growth, thereby achieving integrated monitoring of farm greenhouses.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an integrated intelligent light pole monitoring device for personnel operation and management in farm greenhouses, comprising an integrated intelligent light pole and a monitoring device installed on the integrated intelligent light pole. The monitoring device includes an intelligent monitoring module, an LED light management and control module, a photoresistor module, a temperature monitoring module, an air conditioning module, a humidity monitoring module, a humidifier module, a carbon dioxide monitoring module, a soil pH monitoring module, a solar panel module, and a main control module. The intelligent monitoring module, LED light management and control module, air conditioning module, humidifier module, carbon dioxide monitoring module, soil pH monitoring module, and solar panel module are all electrically connected to the main control module.

[0006] The intelligent monitoring module collects normal light image information and infrared image information inside the greenhouse through a visual sensor. It processes the collected normal light image information using the MOL-POSE algorithm, while the infrared image information is used to assist in determining the distance to the detected target. First, the area containing the human body is selected, and the coordinates of 16 key points on the human body within this area are identified. Based on the positional relationship of the coordinates, a state transition model is used to derive the current posture information of the human body, thereby determining whether the detected person is performing a reasonable operation on the greenhouse.

[0007] The temperature monitoring module, air conditioning module, humidity monitoring module, humidifier module, carbon dioxide monitoring module, and soil pH monitoring module are used to detect and regulate the internal environmental conditions of the greenhouse.

[0008] Preferably, the MOL-POSE algorithm can simultaneously perform pose recognition and target detection tasks. The MOL-POSE algorithm includes a backbone network and an output network (Head structure). The backbone network uses the lightweight network DI-MobileNet to construct a scale pyramid to extract features at different scales of the input image; normal light image information and infrared image information are combined as input to MOL-POSE.

[0009] Preferably, the encoding layer shared by object detection and pose detection uses the lightweight network DI-MobileNet. After downsampling, it obtains low-level features at levels 1 / 16, 1 / 32, and 1 / 64. Since the low-level feature map at level 1 / 64 is small and faces limitations in width and height, a TSPP module is added to the low-level feature map at level 1 / 64 to enhance the extraction of multi-scale features from a single feature map. The multi-scale features are used as the three inputs to the Head structure, which includes three consecutive sets of upsampling and convolution operations. Upsampling uses the PIX-up module to effectively amplify small feature maps. At the same time, the results of upsampling and convolution operations are stacked with the low-level features from the backbone network. The stacked results are then convolved to obtain three outputs. Finally, after passing through the Head structure, the pose task directly regresses to obtain the normalized coordinates of 16 human keypoints, and the object detection task obtains the endpoint results of the prediction boxes at multiple scales.

[0010] Preferably, the Head structure is based on 1 / 16, 1 / 32, and 1 / 64 level low-level features, and decodes different feature maps to obtain the prediction result, i.e., the coordinates of the upper left endpoint (C) of the predicted bounding box for the object detection task. x C y ) and the coordinates of the lower right endpoint (D) x D y Furthermore, based on the scaling ratio of the corresponding feature map relative to the original image, an inverse transformation is performed. Taking the coordinates of the lower right endpoint as an example, the formula is as follows:

[0011]

[0012] In the formula D' x D' y D represents the x and y coordinates after inverse transformation. x D y The coordinates are before the inverse transformation, and K is the scaling ratio. Here, 16, 32, and 64 correspond to the reciprocals of the scaling ratios of the 1 / 16, 1 / 32, and 1 / 64 level bottom feature maps, respectively.

[0013] Based on the point coordinates obtained from the inverse transformation and the original image size, three points with similar positions are grouped together, and the mean values ​​of their x and y coordinates are calculated. The mean point is then used as the endpoint of the final prediction box.

[0014] Preferably, the TSPP module uses depthwise separable convolution instead of ordinary convolutional blocks to expand the receptive field of the 1 / 64 level low-level feature map. The 1 / 64 level low-level feature map is used as input and processed through five channels: In the first channel, no depthwise separable convolution is added; a traditional 1x1 point convolution operation is used. In the second channel, a 3x3 convolution kernel with a depthwise separable convolution of 6 is used for image processing. In the third channel, a 3x3 convolution kernel with a depthwise separable convolution of 12 is used for image processing. In the fourth channel, a 3x3 convolution kernel with a depthwise separable convolution of 18 is used for image processing. In the fifth channel, pooling is performed directly on the original image, and an upsampling operation is completed. Finally, the results from all five channels are stacked together, and point convolution is used to reduce the depth before output.

[0015] Preferably, the PIX-up module includes a convolution operation with a 3x3 kernel and a stride of 1, a normalization operation, an activation operation, and a pixel rearrangement operation; the pixel rearrangement operation transforms the original feature map of size (h,w,r*r*C) into a feature map of size (h*r,w*r,C), where h and w are the height and width of the original feature map, r is the upsampling expansion factor, and C is the number of channels after transformation.

[0016] Preferably, the lightweight network DI-MobileNet is an improvement on MobileNet. First, the 12th, 13th, average pooling layers and fully connected layers of MobileNet are removed. Then, the stride of the 6th layer is changed to 1 / 2 of the original value. Finally, a dilated convolutional layer is added to the 7th layer.

[0017] In convolutional networks, the deeper the network, the larger its receptive field. Dilated convolution can increase the receptive field of the network without stacking convolutional layers. Let the kernel size of both dilated and regular convolution be k. s The number of holes in a dilated convolution is d. r The equivalent kernel size of dilated convolution is K, and the equivalent conversion formula is as follows:

[0018] K = k s +(k s -1)×(d r -1)

[0019] In the formula, K is the size of the ordinary convolution kernel equivalent to dilated convolution, and d r k is the number of holes in the dilated convolution. s Let be the kernel size of dilated convolution. The formula for calculating the receptive field of dilated convolution is as follows:

[0020]

[0021] In the formula T n T represents the size of the receptive field of the nth dilated convolution layer. n-1 This represents the size of the receptive field of the (n-1)th dilated convolution layer, and the size of the equivalent ordinary convolution kernel for the K-dilated convolution layer. This indicates the current step size.

[0022] Preferably, in the state transition model, the arm state at any given time is divided into 7 categories: "raised arm," "lowered arm," "flexed arm," "extended arm," "hooked hand," "raised hand," and "still." The leg state at any given time is divided into 7 categories: "raised leg," "lowered leg," "flexed leg," "extended leg," "raised foot," "lowered foot," and "still." The coordinates of the four key points of each arm and the four key points of each leg are then converted into angles and matched with the aforementioned 14 state categories. The four key points corresponding to each arm, from top to bottom, are shoulder (A), elbow (B), wrist (C), and hand (D). The four key points corresponding to each leg, from top to bottom, are hip (E), knee (F), ankle (G), and foot (H), corresponding to 8 sets of coordinates (x1, y1), (x2, y2), (x3, y3), (x4, y4), (x5, y5), (x6, y6), (x7, y7), and (x8, y8). The formula is then used:

[0023]

[0024] Get the distance between points;

[0025] Define d1, d2, d3, d4, d5, d6, and d7 as the distances from A to B, B to C, C to D, D to E, E to F, F to G, and G to H, respectively.

[0026] Then, using the formula:

[0027]

[0028] The angles at six points—shoulder (A), elbow (B), wrist (C), hip (E), knee (F), and ankle (G)—are defined as θ1, θ2, θ3, θ4, θ5, and θ6, respectively.

[0029] A person's posture at this moment is comprehensively assessed based on the changes in the angle obtained.

[0030] Preferably, in the state transition model, the matching method between the obtained angle and the above 14 types of states is as follows:

[0031] (1) First, determine θ1. If θ1 increases and the range of motion is 0 to 145 degrees, it is defined as "raising the arm". If θ1 decreases and the range of motion is 145 to 0 degrees, it is defined as "lowering the arm". If θ1 remains unchanged, it is defined as "not moving". Record the determination results.

[0032] (2) Then determine θ2. If θ2 decreases and the range of motion is 165 to 0 degrees, it is defined as "arm flexion". If θ2 increases and the range of motion is 0 to 165 degrees, it is defined as "arm extension". If θ2 remains unchanged, it is defined as "immobility". Record the determination results.

[0033] (3) Then determine θ3. If θ3 increases and the range of motion is 0 to +90 degrees, it is defined as "hooking hand". If θ3 decreases and the range of motion is 0 to -90 degrees, it is defined as "raising hand". If θ3 remains unchanged, it is defined as "not moving". Record the determination result.

[0034] (4) Then, θ4 is judged. If θ4 increases and the range of motion is 0 to 120 degrees, it is defined as "raising the leg". If θ4 decreases and the range of motion is 120 to 0 degrees, it is defined as "lowering the leg". If θ4 remains unchanged, it is defined as "not moving". Record the judgment result.

[0035] (5) Then determine θ5. If θ5 decreases and the range of motion is 135 to 0 degrees, it is defined as "leg flexion". If θ5 increases and the range of motion is 0 to 135 degrees, it is defined as "leg extension". If θ5 remains unchanged, it is defined as "immobility". Record the determination results.

[0036] (6) Then determine θ6. If θ6 decreases and the range of motion is 0 to -60 degrees, it is defined as "lifting the foot". If θ6 increases and the range of motion is 0 to +60 degrees, it is defined as "putting the foot down". If θ6 remains unchanged, it is defined as "not moving". Record the determination result.

[0037] Finally, the six judgment results are combined for judgment. Since the states of the arms and legs corresponding to each action are different, the meaning expressed by the action can be deduced.

[0038] The present invention has the following beneficial effects:

[0039] 1. This invention combines normal images with infrared images as input to the detection algorithm, effectively solving the problem of nighttime detection in greenhouses.

[0040] 2. This invention adopts the method of first selecting the human body region and then locating the key points of the human body, which greatly improves the accuracy of locating the key points of the human body, and thus increases the algorithm's ability to distinguish human body movements with high similarity.

[0041] 3. This invention proposes an efficient dual-task algorithm that shares the same encoder when performing target detection and pose detection tasks, which can reduce the amount of computation and improve the detection speed.

[0042] 4. This invention proposes a new algorithm for detecting key points on the human body, which can be applied to more human-computer interaction scenarios through human body movements.

[0043] 5. This invention integrates temperature, humidity, carbon dioxide concentration and soil pH monitoring functions, enabling comprehensive monitoring and control of abnormal conditions in greenhouses, and reducing crop losses.

[0044] 6. The photosensitive module and LED light management and control module used in this invention can monitor and regulate the light environment inside the greenhouse in real time, thereby improving the efficiency of crop photosynthesis and increasing crop yield.

[0045] 7. The intelligent light pole proposed in this invention reduces the need for manpower in greenhouses to a certain extent. It can accurately and efficiently acquire real-time data inside the greenhouse through sensors, which greatly promotes the development of smart agriculture in my country. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the structure of the first type of smart light pole of the present invention;

[0047] Figure 2 This is a schematic diagram of the structure of the second type of smart light pole of the present invention;

[0048] Figure 3 This is a schematic diagram of the structure of the MOL-POSE algorithm in this invention;

[0049] Figure 4 This is a schematic diagram of 16 key human body regions detected by the model proposed in this invention.

[0050] Figure 5 This is a diagram showing the detection results of an embodiment of the present invention;

[0051] Figure 6 This is a block diagram of the monitoring device in this invention;

[0052] Figure 7 This is a diagram of the TSPP module used in the attitude detection network of this invention;

[0053] Figure 8 This is a diagram of the PIX-up module proposed in the attitude detection network of this invention. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0055] Reference Figures 1-8 An integrated intelligent light pole monitoring device for personnel operation and management in farm greenhouses includes an integrated intelligent light pole and a monitoring device installed on the integrated intelligent light pole. The monitoring device includes an intelligent monitoring module, an LED light management and control module, a photoresistor module, a temperature monitoring module, an air conditioning module, a humidity monitoring module, a humidifier module, a carbon dioxide monitoring module, a soil pH monitoring module, a solar panel module, and a main control module. The intelligent monitoring module, LED light management and control module, air conditioning module, humidifier module, carbon dioxide monitoring module, soil pH monitoring module, and solar panel module are all electrically connected to the main control module.

[0056] The intelligent monitoring module collects normal light and infrared images from inside the greenhouse using visual sensors. It processes the collected normal light images using the MOL-POSE algorithm, while the infrared images assist in determining the distance to the detected target. First, it selects the area containing the human body and identifies the coordinates of 16 key points within that area. Based on the positional relationships of these coordinates, it uses a state transition model to determine the current posture of the human body, thereby judging whether the detected person is performing appropriate operations within the greenhouse.

[0057] Among them, the temperature monitoring module, air conditioning module, humidity monitoring module, humidifier module, carbon dioxide monitoring module, and soil pH monitoring module are used to detect and regulate the internal environmental conditions of the greenhouse.

[0058] Specifically, the MOL-POSE algorithm can simultaneously perform pose recognition and target detection tasks. The MOL-POSE algorithm consists of a backbone network and an output network (Head). The backbone network uses the lightweight DI-MobileNet to construct a scale pyramid to extract features at different scales from the input image. Normal light image information and infrared image information are combined as input to MOL-POSE to address the problem of nighttime detection inside greenhouses. The structure of the MOL-POSE algorithm is as follows: Figure 3 As shown.

[0059] Specifically, a shared encoder for object detection and pose detection uses the lightweight DI-MobileNet network. After downsampling, it obtains low-level features at 1 / 16, 1 / 32, and 1 / 64 levels. Due to the small size of the 1 / 64 level low-level feature map, which faces limitations in width and height, a TSPP module is added to the 1 / 64 level low-level features to enhance the extraction of multi-scale features from a single feature map. The multi-scale features are used as the three inputs to the Head model, which includes three consecutive sets of upsampling and convolution operations. Upsampling uses the PIX-up module to effectively amplify small feature maps. Simultaneously, the results of upsampling and convolution operations are stacked with features from the Backbone model, and the stacked results are then convolved to obtain three outputs. Finally, through the Head structure, the pose task directly regresses to obtain the normalized coordinates of 16 human keypoints, and the object detection task obtains the endpoint results of the predicted bounding boxes at multiple scales.

[0060] Specifically, the Head structure decodes different feature maps based on 1 / 16, 1 / 32, and 1 / 64 level low-level features to obtain the prediction result, i.e., the coordinates of the upper left endpoint of the predicted bounding box in the object detection task (C). x C y ) and the coordinates of the lower right endpoint (D) x D y Furthermore, based on the scaling ratio of the corresponding feature map relative to the original image, an inverse transformation is performed. Taking the coordinates of the lower right endpoint as an example, the formula is as follows:

[0061]

[0062] In the formula D' x D' y D represents the x and y coordinates after inverse transformation. x D y The coordinates are before the inverse transformation, and K is the scaling ratio. Here, 16, 32, and 64 correspond to the reciprocals of the scaling ratios of the 1 / 16, 1 / 32, and 1 / 64 level bottom feature maps, respectively.

[0063] Based on the point coordinates obtained from the inverse transformation and the original image size, three points with similar positions are grouped together, and the mean values ​​of their x and y coordinates are calculated. The mean point is then used as the endpoint of the final prediction box.

[0064] Specifically, such as Figure 7As shown, the TSPP module uses depthwise separable convolutions instead of ordinary convolutional blocks to expand the receptive field of the 1 / 64 level low-level feature map. The 1 / 64 level low-level feature map is used as input and processed through five channels: In the first channel, no depthwise separable convolution is added; instead, a traditional 1x1 point convolution operation is used. In the second channel, a 3x3 convolution kernel with a depthwise separable convolution of 6 is used for image processing. In the third channel, a 3x3 convolution kernel with a depthwise separable convolution of 12 is used for image processing. In the fourth channel, a 3x3 convolution kernel with a depthwise separable convolution of 18 is used for image processing. In the fifth channel, pooling is performed directly on the original image, and an upsampling operation is completed. Finally, the results from all five channels are stacked together, and point convolutions are used to reduce the depth before output.

[0065] Specifically, such as Figure 8 As shown, the PIX-up module includes a convolution operation with a 3x3 kernel and a stride of 1, a normalization operation, an activation operation, and a pixel rearrangement operation. The pixel rearrangement operation transforms the original feature map of size (h,w,r*r*C) into a feature map of size (h*r,w*r,C), where h and w are the height and width of the original feature map, r is the upsampling factor, and C is the number of channels after transformation.

[0066] Specifically, the lightweight network DI-MobileNet is an improvement on MobileNet. First, it removes MobileNet's layers 12, 13, average pooling, and fully connected layers. Then, it changes the stride of layer 6 to half of its original value. Finally, it adds a dilated convolutional layer to layer 7.

[0067] In convolutional networks, the deeper the network, the larger its receptive field. Dilated convolution can increase the receptive field of the network without stacking convolutional layers. Let the kernel size of both dilated and regular convolution be k. s The number of holes in a dilated convolution is d. r The equivalent kernel size of dilated convolution is K, and the equivalent conversion formula is as follows:

[0068] K = k s +(k s -1)×(d r -1)

[0069] In the formula, K is the size of the ordinary convolution kernel equivalent to dilated convolution, and d r k is the number of holes in the dilated convolution. s Let be the kernel size of dilated convolution. The formula for calculating the receptive field of dilated convolution is as follows:

[0070]

[0071] In the formula T nT represents the size of the receptive field of the nth dilated convolution layer. n-1 This represents the size of the receptive field of the (n-1)th dilated convolution layer, where K is the kernel size of the ordinary convolution. This indicates the current step size.

[0072] Among them, such as Figure 4 The model identifies 16 key points on the human body, which are then connected to form corresponding vectors. Table 1 shows the 16 key human body regions detected by the model proposed in this embodiment.

[0073] Table 1. Corresponding serial numbers to key parts of the human body

[0074]

[0075]

[0076] like Figure 5 As shown, Figure 5 This is a diagram showing the actual detection results of the present invention, which can accurately locate key information of the human body.

[0077] Specifically, the state transition model is used to derive the current human posture information, which is a method of transforming a set of coordinates of a limb into multiple states. The specific operation is as follows:

[0078] The arm positions at any given time are categorized into 7 types: "raised arm," "lowered arm," "flexed arm," "extended arm," "hooked hand," "raised hand," and "still." The leg positions at any given time are categorized into 7 types: "raised leg," "lowered leg," "flexed leg," "extended leg," "raised foot," "lowered foot," and "still." The coordinates of the four key points for each arm and the four key points for each leg are then converted into angles and matched with the aforementioned 14 position categories. The four key points for each arm, from top to bottom, are shoulder (A), elbow (B), wrist (C), and hand (D). The four key points for each leg, from top to bottom, are hip (E), knee (F), ankle (G), and foot (H), corresponding to 8 sets of coordinates (x1, y1), (x2, y2), (x3, y3), (x4, y4), (x5, y5), (x6, y6), (x7, y7), and (x8, y8). The formula is then used to...

[0079]

[0080] Get the distance between points;

[0081] Define d1, d2, d3, d4, d5, d6, and d7 as the distances from A to B, B to C, C to D, D to E, E to F, F to G, and G to H, respectively.

[0082] Then, using the formula:

[0083]

[0084] The angles at six points—shoulder (A), elbow (B), wrist (C), hip (E), knee (F), and ankle (G)—are defined as θ1, θ2, θ3, θ4, θ5, and θ6, respectively.

[0085] A person's posture at this moment is comprehensively assessed based on the changes in the angle obtained.

[0086] Specifically, in the state transition model, the matching method between the obtained angle and the above 14 types of states is as follows:

[0087] (1) First, determine θ1. If θ1 increases and the range of motion is 0 to 145 degrees, it is defined as "raising the arm". If θ1 decreases and the range of motion is 145 to 0 degrees, it is defined as "lowering the arm". If θ1 remains unchanged, it is defined as "not moving". Record the determination results.

[0088] (2) Then determine θ2. If θ2 decreases and the range of motion is 165 to 0 degrees, it is defined as "arm flexion". If θ2 increases and the range of motion is 0 to 165 degrees, it is defined as "arm extension". If θ2 remains unchanged, it is defined as "immobility". Record the determination results.

[0089] (3) Then determine θ3. If θ3 increases and the range of motion is 0 to +90 degrees, it is defined as "hooking hand". If θ3 decreases and the range of motion is 0 to -90 degrees, it is defined as "raising hand". If θ3 remains unchanged, it is defined as "not moving". Record the determination result.

[0090] (4) Then, θ4 is judged. If θ4 increases and the range of motion is 0 to 120 degrees, it is defined as "raising the leg". If θ4 decreases and the range of motion is 120 to 0 degrees, it is defined as "lowering the leg". If θ4 remains unchanged, it is defined as "not moving". Record the judgment result.

[0091] (5) Then determine θ5. If θ5 decreases and the range of motion is 135 to 0 degrees, it is defined as "leg flexion". If θ5 increases and the range of motion is 0 to 135 degrees, it is defined as "leg extension". If θ5 remains unchanged, it is defined as "immobility". Record the determination results.

[0092] (6) Then determine θ6. If θ6 decreases and the range of motion is 0 to -60 degrees, it is defined as "lifting the foot". If θ6 increases and the range of motion is 0 to +60 degrees, it is defined as "putting the foot down". If θ6 remains unchanged, it is defined as "not moving". Record the determination result.

[0093] Finally, the six judgment results are combined for judgment. Since the states of the arms and legs corresponding to each action are different, the meaning expressed by the action can be deduced.

[0094] Among them, such as Figure 6 As shown, Figure 6 This diagram illustrates the relationships between the various modules of the invention. The photoresistor module collects ambient light intensity information. The LED light management and control module receives this information and uses it to control the LED lights' on / off states, light intensity, and illumination time. Additionally, the external lighting in the greenhouse is also controlled by this module. The temperature monitoring module, humidity monitoring module, air conditioning module, and humidifier module work together to monitor and regulate the temperature and humidity inside the greenhouse. When the temperature and humidity sensors detect that the temperature and humidity levels inside the greenhouse exceed thresholds, they send an alarm to the user via an application and automatically adjust the temperature and humidity levels inside the greenhouse using the air conditioner and humidifier. The solar energy module powers the system if the solar panels have sufficient power; otherwise, it automatically switches to electric power. The alarm function promptly notifies the user of any solar panel malfunctions, preventing the user from being unaware of the solar panels' operational status. The main control module, with its built-in processor, automatically adjusts the system based on various environmental parameters and communicates with the user via Bluetooth, prioritizing the execution of user commands.

[0095] Specifically, the smart light pole is designed with two structures. The first is the normal structure, such as... Figure 1 As shown; the second type is a simple structure that is easy to install, such as... Figure 2 As shown. The first normal structure consists of two symmetrically arranged lamp posts 101 as the main support. The main body of the lamp post has a circular cross-section and a large bottom contact area, which is beneficial to the stability of the lamp post. An air conditioning vent 102 is provided on the inner side of the lamp post, and a first monocular camera 103 with its lens angled downwards is located at the top of the lamp post. A temperature sensor 104, a humidity sensor 105, and a carbon dioxide concentration sensor 106 are installed from top to bottom on the lower part of the lamp post. An outdoor street light 107 is installed on the upper outer side of the lamp post. An external main control module device 108 is installed in the middle of one of the lamp posts. The main control module device has heat dissipation vents on the front and sides to provide a good working environment for its built-in processor. The tops of the two lamp posts are connected by a crossbeam. A second monocular camera 110 and a humidifier 111 are installed on the lower side of the crossbeam. A solar panel 109 is installed on the upper surface of the crossbeam, which, together with the power supply, powers the smart lamp post. Several soil pH sensors 114 are installed at the ground contact point of the frame 112, which connects to both sides of the light pole. LED tubes are connected between the arc-shaped ribs 113 on both sides of the frame to provide lighting for the greenhouse's intelligent operation. Figure 1As shown in the diagram; during installation, after measuring the distance, first install the two lamp posts symmetrically, then install the top crossbeam, and then install the solar module. Measure the spacing of the brackets (traditional greenhouse mounting brackets used to support the keel structure, not shown in the diagram) again, and install the brackets. After the brackets are installed, install several keel structures of the greenhouse (traditional greenhouse keel structures used to support the greenhouse surface, not shown in the diagram). After all are installed, install both ends of the LED tubes to the keel structures on both sides as shown in the diagram, connecting the LED light management and control module to the lamp post body. Secure the lamp tubes in the middle to prevent them from falling or breaking. Then install the soil pH sensor, and finally install the external streetlights on the outside of the lamp post and the external main control module device in the middle.

[0096] The second simple structure, such as Figure 2 As shown, two lampposts 101 are still used as the support structure. The area of ​​the air conditioning vent 102 is reduced and moved to the side. A monocular camera 103 with its lens angled downwards is installed at the top of the lamppost 101. From top to bottom, a temperature sensor 104, a humidity sensor 105, and a carbon dioxide concentration sensor 106 are installed at the lower end of the lamppost 101. Several soil pH sensors 109 are installed at the bottom of the lamppost 101. An outdoor street light 107 is installed on the upper outer side of the lamppost 101. An external main control module device 108 is installed in the middle of one of the lampposts. The main control module device has heat dissipation vents on the front and sides to provide a good working environment for its built-in processor. The tops of the two lampposts are still connected by a crossbeam 110, and the device on the crossbeam is the same as that in the first normal structure. Figure 2 As shown; during installation, after measuring the distance, first install the two lamp posts symmetrically, then install the top crossbeam, and then install the solar modules. Measure the spacing of the brackets (traditional greenhouse mounting brackets used to support the keel structure, not shown in the diagram) again, and install the brackets. After the brackets are installed, install several keel structures of the greenhouse (traditional greenhouse keel structures used to support the greenhouse surface, not shown in the diagram). After all are installed, finally install the external streetlights on the outside of the lamp posts and the external main control module device in the middle. The first normal structure achieves the best effect in monitoring and controlling the greenhouse environment, while the second simplified structure greatly improves installation efficiency.

[0097] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An integrated intelligent light pole monitoring device for personnel operation and management in farm greenhouses, characterized in that, The system includes an integrated smart light pole and a monitoring device installed on the integrated smart light pole. The monitoring device includes a smart monitoring module, an LED light management and control module, a photoresistor module, a temperature monitoring module, an air conditioning module, a humidity monitoring module, a humidifier module, a carbon dioxide monitoring module, a soil pH monitoring module, a solar panel module, and a main control module. The smart monitoring module, LED light management and control module, air conditioning module, humidifier module, carbon dioxide monitoring module, soil pH monitoring module, and solar panel module are all electrically connected to the main control module. The intelligent monitoring module collects normal light image information and infrared image information inside the greenhouse through a visual sensor. It processes the collected normal light image information using the MOL-POSE algorithm, and uses the infrared image information to help provide the distance of the detected target. First, it selects the area of ​​the human body and identifies the coordinates of 16 key points of the human body in the area. Based on the positional relationship of the coordinates, it uses a state transition model to obtain the current posture information of the human body, and then determines whether the detected person is performing reasonable operations on the greenhouse. The temperature monitoring module, air conditioning module, humidity monitoring module, humidifier module, carbon dioxide monitoring module, and soil pH monitoring module are used to detect and regulate the internal environmental conditions of the greenhouse. The MOL-POSE algorithm is used to simultaneously complete pose recognition and target detection tasks. The MOL-POSE algorithm includes a backbone network and an output network (Head structure). The backbone network uses the lightweight network DI-MobileNet to construct a scale pyramid to extract features of different scales from the input image. Normal light image information and infrared image information are combined as input to the MOL-POSE algorithm; The lightweight network DI-MobileNet, after downsampling, obtains , as well as Low-level features; The The TSPP module is added to the low-level features to enhance the extraction of multi-scale features from a single feature map. The multi-scale features are used as the three inputs to the Head structure, which includes three consecutive sets of upsampling and convolution operations. The upsampling uses the PIX-up module to effectively amplify small feature maps. At the same time, the results of the upsampling and convolution operations are stacked with the low-level features from the backbone network. The stacked results are then convolved to obtain three outputs. Finally, after passing through the Head structure, the pose task directly regresses to obtain the normalized coordinates of 16 human keypoints. The object detection task obtains the endpoint results of the prediction boxes at multiple scales. The Head structure is based on , as well as The lowest-level features are decoded to obtain the prediction result, i.e., the coordinates of the upper left endpoint of the predicted bounding box in the object detection task. and the coordinates of the lower right endpoint Furthermore, based on the scaling ratio of the corresponding feature map relative to the original image, an inverse transformation is performed. Taking the coordinates of the lower right endpoint as an example, the formula is as follows: ； In the formula , The x and y coordinates are after inverse transformation. , The coordinates before the inverse transformation, To reduce the ratio, we take... , as well as Corresponding to , as well as The reciprocal of the scaling ratio corresponding to the lowest level feature map; Based on the point coordinates obtained from the inverse transformation and the original image size, three points with similar positions are grouped together, and the mean of the x and y coordinates is calculated. The mean point is used as the endpoint of the final prediction box. The TSPP module uses depthwise separable convolutions instead of ordinary convolutional blocks, expanding... The receptive field of the low-level feature map; The low-level feature map is used as input and passed through five channels: In the first channel, no depthwise separable convolution is added, and traditional methods are used. Point convolution operation; the second channel uses The convolution kernel has a depthwise separable convolution with a separation degree of 6 for image processing; the third channel uses... The convolution kernel simultaneously uses a depthwise separable convolution with a separation degree of 12 for image processing; the fourth channel uses... The convolution kernel has a depthwise separable convolution with a separation degree of 18 for image processing; the fifth channel directly performs pooling on the original image and completes an upsampling operation; finally, the results of all five channels are stacked together, and point convolution is used to reduce the depth before output; The PIX-up module includes a convolutional kernel. Convolution operations with a stride of 1, normalization operations, activation operations, and pixel rearrangement operations; Pixel rearrangement operation will change the original size to Feature map conversion Size characteristics and The height and width of the original feature map. This is the upsampling expansion factor. This represents the number of channels after transformation.

2. The integrated intelligent light pole monitoring device for personnel operation management in farm greenhouses according to claim 1, characterized in that, The lightweight network DI-MobileNet is an improvement on MobileNet. First, the 12th, 13th, average pooling layers and fully connected layers of MobileNet are removed. Then, the stride of the 6th layer is changed to 1 / 2 of the original. Finally, a dilated convolutional layer is added to the 7th layer. In convolutional networks, the deeper the network, the larger its receptive field. Dilated convolution can increase the receptive field of the network without stacking convolutional layers. Let the kernel size of dilated convolution and ordinary convolution both be 1. The number of holes in a dilated convolution is The equivalent size of a dilated convolution kernel is... The equivalent conversion formula is as follows: ； In the formula This is the size of the ordinary convolution kernel, which is equivalent to dilated convolution. The number of holes in the dilated convolution. Let be the kernel size of dilated convolution. The formula for calculating the receptive field of dilated convolution is as follows: ； In the formula Indicates the first The size of the receptive field of the layered void convolution. Indicates the first The size of the receptive field of the layered void convolution. The kernel size is the equivalent of a dilated convolution in a regular convolution. This indicates the current step size.

3. The integrated intelligent light pole monitoring device for personnel operation management in farm greenhouses according to claim 1, characterized in that, In the state transition model, the arm state at any given time is divided into 7 categories: "raise arm", "lower arm", "bent arm", "extended arm", "hook hand", "raise hand", and "still". The leg state at any given time is divided into 7 categories: "raise leg", "lower leg", "bent leg", "extended leg", "raise foot", "lower foot", and "still". The coordinates of the four key points of each arm and the four key points of each leg are then converted into angles and matched with the aforementioned 14 state categories. The four key points corresponding to each arm, from top to bottom, are shoulder (…). ),elbow( ),wrist( ),hand( Four points, each leg corresponds to four key points from top to bottom: hip ( ),knee( ),ankle( ),foot( ), corresponding to 8 sets of coordinates ( (), (), (), (), (), (), (), ), and use the formula: ， ； Get the distance between points; Define separately , , , , , , for arrive distance, arrive distance, arrive distance, arrive distance, arrive distance, arrive distance, arrive distance; Then, using the formula: ，（ ） The angles at six points—shoulder (A), elbow (B), wrist (C), hip (E), knee (F), and ankle (G)—are defined as follows: ; A person's posture at this moment is comprehensively assessed based on the changes in the angle obtained.

4. The integrated intelligent light pole monitoring device for personnel operation management in farm greenhouses according to claim 3, characterized in that, In the state transition model, the matching method between the obtained angle and the above 14 types of states is as follows: (1) First, Make a judgment, if Increase, range of activity ~ The degree is defined as "raising the arm", if Reduce, activity range ~ The degree is defined as "lowering arm". If it remains unchanged, it is defined as "unchanged," and the judgment result is recorded; (2) Again Make a judgment, if Reduce, activity range ~ Degree, then defined as "bent arm", if Increase, range of activity ~ The degree is defined as "arm extension". If it remains unchanged, it is defined as "unchanged," and the judgment result is recorded; (3) Again Make a judgment, if Increase, range of activity ~ The degree is defined as "hook hand". Reduce, activity range ~ The degree is defined as "raising the hand". If it remains unchanged, it is defined as "unchanged," and the judgment result is recorded; (4) After that Make a judgment, if Increase, range of activity ~ The degree is defined as "lifting the leg". Reduce, activity range ~ The degree is defined as "leg extension". If it remains unchanged, it is defined as "unchanged," and the judgment result is recorded; (5) Again Make a judgment, if Reduce, activity range ~ Degree, then defined as "bent leg", if Increase, range of activity ~ The degree is defined as "extending the leg". If it remains unchanged, it is defined as "unchanged," and the judgment result is recorded; (6) Again Make a judgment, if Reduce, activity range ~ The degree is defined as "lifting the foot". Increase, range of activity ~ The degree is defined as "landing point". If it remains unchanged, it is defined as "unchanged," and the judgment result is recorded; Finally, the six judgment results are combined for judgment. Since the states of the arms and legs corresponding to each action are different, the meaning expressed by the action can be deduced.

Images