A deep learning-based material scheduling process dynamic identification method

Abstract

The application belongs to the technical field of dispatch management, and particularly relates to a material dispatch process dynamic identification method based on deep learning. The method comprises the following steps: acquiring monitoring videos of each region of a dispatch work site, identifying targets and set regions in the monitoring videos by using a target detection model, and tracking the identified targets; setting a detection frame of the set region as a virtual portal, judging whether the coordinates of the tracked target move into the virtual portal, calculating distance changes between the target and the virtual portal when the target moves into the virtual portal, and judging that the target enters the set region if the distance changes between the target and the virtual portal do not exceed a set distance threshold in continuous set frames. Thus, the consumption of human cost is reduced, the operation of software by personnel is reduced, the work efficiency is improved, and additional positioning devices are not needed, so that the dynamic identification of material state information in the material dispatch process is realized at low cost.

Description

Technical Field

[0001] This invention belongs to the field of scheduling and management technology, specifically relating to a method for dynamic identification of material scheduling processes based on deep learning. Background Technology

[0002] Material dispatching systems typically involve multiple devices (such as handcarts and motorized vehicles) distributed over a wide area. When dispatching materials across a distributed area, it's necessary to collect operational progress information from each warehouse, including the job number and current progress. In existing technologies, once materials arrive at a specific area, personnel report the material status information (e.g., the area the materials have entered) to dispatchers via various communication methods. Dispatchers then manually update the corresponding material status information, and consequently, the operational progress information, using interactive software that collects operational progress information from each warehouse. This method requires manual reporting and software updates. For material dispatching systems with numerous warehouses and areas, the workload for personnel information interaction and software operation is substantial, resulting in low efficiency, high labor costs, and a certain lag in material status updates during the dispatching process.

[0003] Chinese patent application CN114140043A discloses an intelligent logistics monitoring system. This system involves installing a cargo terminal unit on a transport truck. The cargo terminal unit includes a positioning module for detecting the vehicle's location, monitoring the vehicle's driving status and cargo loading status. The monitoring data is transmitted wirelessly to a transportation management unit, and then wirelessly to a dispatch and control platform. While this method reduces manual workload by installing positioning modules to monitor the real-time location of vehicles carrying goods, it requires additional positioning modules to be installed on various devices within the material dispatch system. For material dispatch systems with numerous warehouses and areas, this significantly increases the cost of identifying material status information during the dispatch process, making it difficult to promote and implement widely. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based method for dynamic identification of material scheduling processes, in order to solve the problem of high cost in identifying material status information in existing material scheduling processes.

[0005] To address the aforementioned technical problems, this invention provides a deep learning-based dynamic identification method for material scheduling processes, comprising: acquiring monitoring videos of various areas at the scheduling operation site, wherein the monitoring videos include frames of images of vehicles and materials during the material scheduling process; using a target detection model to identify targets and designated areas in the monitoring videos, obtaining detection boxes for the targets and the designated areas, wherein the detection boxes include coordinates and detection box scales; wherein the targets include materials and / or vehicles; tracking the identified targets; setting the detection boxes of the designated areas as virtual portals, determining whether the coordinates of the tracked targets have moved into the virtual portals, calculating the distance change between the targets and the virtual portals when they move into the virtual portals, and determining that the target has entered the designated area if the distance change between the two does not exceed a set distance threshold in consecutive designated frames.

[0006] Furthermore, the target detection model is also used to detect target categories. The target detection box also includes target categories, which include material types and vehicle types. After identifying targets in the surveillance video, the number of each type of target in each area is counted.

[0007] Furthermore, the method also includes counting the number of various types of targets in each region and storing the statistical results in a database.

[0008] Furthermore, the target detection model is trained using a material scheduling target dataset, which includes images of empty vehicles, single material items, and vehicles loaded with materials.

[0009] Furthermore, the target detection model adopts a single-stage target model.

[0010] Furthermore, the target detection model adopts an improved YOLO model, which replaces the Bottleneck module in C2f of the YOLOv8 backbone network with the FasterBlock module of FasterNet.

[0011] Furthermore, the tracking of the identified target is achieved using the DeepSort algorithm.

[0012] Furthermore, the method also includes displaying the result of the target entering the designated area in the surveillance video.

[0013] Furthermore, before determining whether the coordinates of the tracked target have moved into the virtual portal, it is first determined whether the tracked target has moved. The method for determining whether the tracked target has moved is: calculate the distance between the target in the current frame and the set interval frame. If the distance between the target in the current frame and the set interval frame is greater than the set value, then it is determined that the target is moving.

[0014] Furthermore, the target dataset for material dispatching is obtained by cleaning, labeling, and data augmenting the images of vehicles and materials collected from the material dispatching system.

[0015] The beneficial effects of the above technical solution are as follows: This invention is an improved invention. It acquires monitoring videos of various areas at the dispatching operation site, uses a target detection model to identify targets and designated areas in the acquired monitoring videos, and then tracks the targets. The detection boxes of the identified designated areas are set as virtual portals. When the tracked target coordinates move to the virtual portal, the distance changes between the two in multiple frames are used to determine whether the target has entered the designated area. Thus, based on the target detection and tracking results, the target status during the dispatching process is judged, enabling real-time automatic monitoring of the areas where materials arrive during the material dispatching process. This invention utilizes the monitoring videos of various areas automatically collected by the existing monitoring equipment at the dispatching operation site, combined with target detection and tracking technology, to judge the material dispatching process based on tracking and positioning. This greatly reduces the consumption of labor costs, reduces personnel's operation of software, improves work efficiency, and eliminates the need for additional positioning devices, achieving dynamic identification of material status information during the material dispatching process at low cost.

[0016] To address the aforementioned technical problems, this invention also provides a deep learning-based dynamic identification device for material scheduling processes, comprising a processor. The processor executes steps in a deep learning-based dynamic identification method for material scheduling processes. This method includes: acquiring monitoring videos of various areas at the scheduling operation site, the monitoring videos including frames of vehicle and material images during the material scheduling process; using a target detection model to identify targets and designated areas in the monitoring videos, obtaining detection boxes for the targets and designated areas, the detection boxes including coordinates and detection box scales; wherein the targets include materials and / or vehicles; tracking the identified targets; setting the detection boxes of the designated areas as virtual portals; determining whether the coordinates of the tracked target have moved into the virtual portal; calculating the distance change between the target and the virtual portal when the target moves into the virtual portal; if the distance change between the two does not exceed a set distance threshold in consecutive designated frames, then the target is determined to have entered the designated area.

[0017] Furthermore, the target detection model is also used to detect target categories. The target detection box also includes target categories, which include material types and vehicle types. After identifying targets in the surveillance video, the number of each type of target in each area is counted.

[0018] Furthermore, the method also includes counting the number of various types of targets in each region and storing the statistical results in a database.

[0019] Furthermore, the target detection model is trained using a material scheduling target dataset, which includes images of empty vehicles, single material items, and vehicles loaded with materials.

[0020] Furthermore, the target detection model adopts a single-stage target model.

[0021] Furthermore, the target detection model adopts an improved YOLO model, which replaces the Bottleneck module in C2f of the YOLOv8 backbone network with the FasterBlock module of FasterNet.

[0022] Furthermore, the tracking of the identified target is achieved using the DeepSort algorithm.

[0023] Furthermore, the method also includes displaying the result of the target entering the designated area in the surveillance video.

[0024] Furthermore, before determining whether the coordinates of the tracked target have moved into the virtual portal, it is first determined whether the tracked target has moved. The method for determining whether the tracked target has moved is: calculate the distance between the target in the current frame and the set interval frame. If the distance between the target in the current frame and the set interval frame is greater than the set value, then it is determined that the target is moving.

[0025] The beneficial effects of the above technical solution are as follows: This invention is an improved invention, providing a device capable of automatically identifying the status of the logistics scheduling process. This device acquires monitoring videos of various areas at the scheduling operation site, uses a target detection model to identify targets and designated areas in the acquired monitoring videos, and then tracks the targets. The detection frame of the identified designated area is set as a virtual portal. When the tracked target coordinates move to the virtual portal, the device determines whether the target has entered the designated area based on the distance change between the two in multiple frames. Therefore, based on the target detection and tracking results, the device judges the target status during the scheduling process, enabling real-time automatic monitoring of the areas where materials arrive during material scheduling. This invention utilizes existing monitoring equipment at the scheduling operation site to automatically collect monitoring videos of various areas, combined with target detection and tracking technology. Based on tracking and positioning, it judges the material scheduling process, greatly reducing labor costs, minimizing personnel's interaction with the software, improving work efficiency, and eliminating the need for additional positioning devices. This achieves dynamic identification of material status information during material scheduling at low cost. Attached Figure Description

[0026] Figure 1 This is a flowchart of the dynamic identification process of material scheduling based on deep learning, as described in this invention. Figure 2This is a diagram of the improved YOLO model architecture of the present invention; Figure 3 This is the C2f_Faster architecture diagram of the present invention; Figure 4 This is a diagram of the FasterBlock module architecture of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0028] This invention performs deep learning on target detection and tracking processing of monitoring videos of various areas during the material dispatching process, and identifies the area where the target is located during the material dispatching process based on the detection and tracking results.

[0029] Method Implementation This invention discloses a deep learning-based dynamic identification method for material dispatching. Based on target detection algorithms and combined with target tracking algorithms, this method processes monitoring videos automatically collected by existing monitoring equipment at the dispatching site. It tracks support vehicles and materials during the dispatching process and automatically determines the arrival location of vehicles and materials. This reduces the need for personnel to perform software operations and communication around the arrival area after materials reach a specific location, thus improving the situational awareness capability of the material dispatching system during transportation. Figure 1 As shown, the method includes the following steps: 1. Obtain monitoring videos of each area at the dispatching operation site.

[0030] The surveillance videos for each area are automatically collected by the existing monitoring equipment at the dispatching site. These videos include frames of images of vehicles and materials during the material dispatching process, including images of empty vehicles, single-material vehicles, and vehicles loaded with materials. This invention utilizes existing monitoring equipment to automatically collect video data without incurring additional equipment costs.

[0031] 2. Target detection.

[0032] The target detection model is used to identify targets and defined areas in surveillance video, resulting in detection bounding boxes for both the targets and the defined areas. Each detection bounding box includes its coordinates and scale. Targets include materials and / or vehicles.

[0033] Preferably, the target detection model is also used to detect target categories and defined region types. In this case, the detection box includes coordinates, category, and detection box scale. That is, the target detection box includes target coordinates (center coordinates of the target's detection box), target category, and target detection box scale. The defined region detection box includes defined region coordinates (center coordinates of the defined region's detection box), defined region category, and defined region detection box scale. Target categories include material types and vehicle types.

[0034] 1) Target detection model.

[0035] Existing two-stage target detection models and single-stage target models can be used. Since the single-stage target model has a smaller volume than the two-stage target detection model, the single-stage target model can be preferred.

[0036] In the monitoring of material scheduling processes, the size and computational load of single-stage target models remain redundant. Therefore, as a preferred implementation, the target detection model adopts an improved YOLO model. The improved YOLO model replaces the Bottleneck module in the C2f module of the YOLOv8 backbone network with the FasterBlock module of FasterNet, resulting in C2f_Faster. The architecture of the improved YOLO model is as follows: Figure 2 As shown.

[0037] The architecture of C2f_Faster is as follows Figure 3 As shown, the n Bottleneck modules in C2f are replaced with n FasterBlock modules from FasterNet.

[0038] FasterBlock module architecture as follows Figure 4 As shown, the FasterBlock module replaces the standard 3×3 convolution module with partial convolution (PConv3×3). The number of channels in partial convolution is 1 / 4 of that in standard convolution. Since the computational complexity of convolution operations is calculated using the formula h×w×k... 2 ×c 2 In this context, h, w, and c represent the length, width, and number of channels of the feature map, respectively, and k is the size of the convolution kernel (the kernel size is 3 and the stride is 1 in partial convolution). Therefore, the computational cost of partial convolution is 1 / 16 of that of standard convolution. Furthermore, partial convolution reduces memory access due to the reduced number of convolution operations.

[0039] Replacing the original C2f module in YOLOv8 with C2f_Faster reduces the model size and computational cost. Experimental comparisons show that the improved YOLOv8 model has a 24.3% smaller size, a 33% lower computational cost, and a 24.6% fewer parameters, while maintaining essentially the same accuracy. Data before and after the improvement are shown in Table 1.

[0040] Table 1

[0041] 2) Training of the object detection model.

[0042] The target detection model was trained using a material scheduling target dataset. The material scheduling target dataset includes images of empty vehicles, images of single materials, and images of vehicles loaded with materials.

[0043] The construction process of the target dataset for material dispatching is as follows: Image information of vehicles and materials in the material dispatching system is collected through cameras and webcams. The collected images are diverse, including images of single vehicles (i.e., empty vehicles), single materials, and vehicles combined with materials. The obtained images are cleaned, removing duplicate and blurry images, and labels are assigned to vehicles, various types of materials, and regions. Then, the LabelImg image annotation tool is used to annotate the dataset according to the YOLO format, resulting in txt text labels. Finally, the dataset is divided into training, validation, and test sets in a 7:1:2 ratio.

[0044] In real-world environments, varying lighting conditions can cause glare or blurring of targets such as vehicles and goods. Preferably, to enhance the robustness and anti-interference capabilities of the target detection model, data augmentation is performed on the obtained dataset. Data augmentation includes random flipping, color changes, translation, Mosaic effects, and Mixup.

[0045] During the training of the object detection model, the SGD gradient optimizer is used for optimization. The initial learning rate of the SGD gradient optimizer is 0.01, the momentum is 0.937, the decay factor is 0.0005, the batch size is 8, and the training period is 100 rounds. After training, the optimal weight file is obtained, and thus the optimal object detection model is obtained.

[0046] 3. Track the identified targets.

[0047] The identified target is tracked using a target tracking algorithm, such as the DeepSort algorithm.

[0048] This paper combines a target detection model with a target tracking algorithm. The target detection model detects targets in the surveillance video and identifies target bounding boxes in the video sequence. The target tracking algorithm predicts target bounding boxes in subsequent frames based on the detected bounding boxes and calculates the Interchange of Units (IOU) between the predicted and detected bounding boxes. If the IOU is greater than a set threshold, an ID is generated for the target. Targets with the same ID are considered the same target. This IOU is used for tracking support vehicles and supplies, as well as for subsequent area determination. The formula for calculating IOU is as follows:

[0049] Among them, B det It is the detection bounding box, B. track ∩ represents the intersection of two bounding boxes, and ∪ represents the union of two bounding boxes. A larger IOU indicates a higher degree of overlap between the two bounding boxes, and a greater probability that they come from the same target.

[0050] 4. Based on the identified and tracked targets, the material dispatching process is dynamically identified.

[0051] 1) Based on the identified and tracked targets, determine the location of the targets.

[0052] First, determine if the tracked target has moved: calculate the distance between the target in the current frame and the set interval frame. If the distance between the target in the current frame and the set interval frame is greater than the set value, then the target is considered to be moving. The distance between the target in the current frame and the set interval frame refers to the distance between the target's center point (the center coordinates of the detection box) in the current frame and the set interval frame.

[0053] Based on the characteristics of the material scheduling and transportation process, this invention does not use the calculation of the center point distance in every frame to determine whether the material has moved. Instead, it calculates the Euclidean distance of the center change after several frames. When the calculated distance is greater than a set threshold, it is determined that the material is moving.

[0054] Then, it is determined whether the target has entered the set area: the detection box of the set area identified by the target detection model in the video stream is set as a virtual portal. It is determined whether the coordinates of the tracked target have moved into the virtual portal. When the target moves into the virtual portal, the distance change between the target and the virtual portal is calculated. If the distance change between the two does not exceed the set distance threshold in consecutive set frames, it is determined that the target has entered the set area, so as to realize real-time monitoring of the area reached by the target during the material dispatching process.

[0055] The change in distance between the target and the virtual portal refers to the change in distance between the target's center point and the virtual portal. The result of the target entering the designated area is displayed in the surveillance video.

[0056] 2) Count the number of each type of target in each region and store the statistical results in the database.

[0057] Based on target detection, the number of various targets in the recognition area is counted. According to the different labeling numbers of different targets, the number of each number in the recognition video is counted, and the number is matched with the target category. The number of different targets identified is counted and written into the database.

[0058] Device Implementation The present invention provides a deep learning-based dynamic identification device for material scheduling processes, comprising a processor for executing steps in a deep learning-based dynamic identification method for material scheduling processes, the method comprising: 1. Obtain monitoring videos of each area at the dispatching operation site.

[0059] The surveillance videos for each area are automatically collected by the existing monitoring equipment at the dispatching site. These videos include frames of images of vehicles and materials during the material dispatching process, including images of empty vehicles, single-material vehicles, and vehicles loaded with materials. This invention utilizes existing monitoring equipment to automatically collect video data without incurring additional equipment costs.

[0060] 2. Target detection.

[0061] The target detection model is used to identify targets and defined areas in surveillance video, resulting in detection bounding boxes for both the targets and the defined areas. Each detection bounding box includes its coordinates and scale. Targets include materials and / or vehicles.

[0062] Preferably, the target detection model is also used to detect target categories and defined region types. In this case, the detection box includes coordinates, category, and detection box scale. That is, the target detection box includes target coordinates (center coordinates of the target's detection box), target category, and target detection box scale. The defined region detection box includes defined region coordinates (center coordinates of the defined region's detection box), defined region category, and defined region detection box scale. Target categories include material types and vehicle types.

[0063] 1) Target detection model.

[0064] Existing two-stage target detection models and single-stage target models can be used. Since the single-stage target model has a smaller volume than the two-stage target detection model, the single-stage target model can be preferred.

[0065] In the monitoring of material scheduling process, the size and computational load of the single-stage target model are still redundant. Therefore, as a preferred implementation method, the target detection model adopts the improved YOLO model. The improved YOLO model is to replace the Bottleneck module in C2f of the YOLOv8 model backbone network with the FasterBlock module of FasterNet, resulting in C2f_Faster.

[0066] C2f_Faster is a FasterBlock module that replaces the n Bottleneck modules in C2f with n FasterNet modules.

[0067] The FasterBlock module replaces the standard 3×3 convolutional module with partial convolution (PConv3×3). The number of channels in partial convolution is 1 / 4 of that in standard convolution. Since the computational cost in convolution operations is calculated using the formula h×w×k... 2 ×c 2 In this context, h, w, and c represent the length, width, and number of channels of the feature map, respectively, and k is the size of the convolution kernel (the kernel size is 3 and the stride is 1 in partial convolution). Therefore, the computational cost of partial convolution is 1 / 16 of that of standard convolution. Furthermore, partial convolution reduces memory access due to the reduced number of convolution operations.

[0068] Replacing the original C2f module in YOLOv8 with C2f_Faster reduces the model size and computational cost. Experimental results show that the improved YOLOv8 model has a 24.3% smaller size, a 33% lower computational cost, and a 24.6% fewer parameters, while maintaining essentially the same accuracy.

[0069] 2) Training of the object detection model.

[0070] The target detection model was trained using a material scheduling target dataset. The material scheduling target dataset includes images of empty vehicles, images of single materials, and images of vehicles loaded with materials.

[0071] The construction process of the target dataset for material dispatching is as follows: Image information of vehicles and materials in the material dispatching system is collected through cameras and webcams. The collected images are diverse, including images of single vehicles (i.e., empty vehicles), single materials, and vehicles combined with materials. The obtained images are cleaned, removing duplicate and blurry images, and labels are assigned to vehicles, various types of materials, and regions. Then, the LabelImg image annotation tool is used to annotate the dataset according to the YOLO format, resulting in txt text labels. Finally, the dataset is divided into training, validation, and test sets in a 7:1:2 ratio.

[0072] In real-world environments, varying lighting conditions can cause glare or blurring of targets such as vehicles and goods. Preferably, to enhance the robustness and anti-interference capabilities of the target detection model, data augmentation is performed on the obtained dataset. Data augmentation includes random flipping, color changes, translation, Mosaic effects, and Mixup.

[0073] During the training of the object detection model, the SGD gradient optimizer is used for optimization. The initial learning rate of the SGD gradient optimizer is 0.01, the momentum is 0.937, the decay factor is 0.0005, the batch size is 8, and the training period is 100 rounds. After training, the optimal weight file is obtained, and thus the optimal object detection model is obtained.

[0074] 3. Track the identified targets.

[0075] The identified target is tracked using a target tracking algorithm, such as the DeepSort algorithm.

[0076] This paper combines a target detection model with a target tracking algorithm. The target detection model detects targets in the surveillance video and identifies target bounding boxes in the video sequence. The target tracking algorithm predicts target bounding boxes in subsequent frames based on the detected bounding boxes and calculates the Interchange of Units (IOU) between the predicted and detected bounding boxes. If the IOU is greater than a set threshold, an ID is generated for the target. Targets with the same ID are considered the same target. This IOU is used for tracking support vehicles and supplies, as well as for subsequent area determination. The formula for calculating IOU is as follows:

[0077] Among them, B det It is the detection bounding box, B. track ∩ represents the intersection of two bounding boxes, and ∪ represents the union of two bounding boxes. A larger IOU indicates a higher degree of overlap between the two bounding boxes, and a greater probability that they come from the same target.

[0078] 4. Based on the identified and tracked targets, the material dispatching process is dynamically identified.

[0079] 1) Based on the identified and tracked targets, determine the location of the targets.

[0080] First, determine if the tracked target has moved: calculate the distance between the target in the current frame and the set interval frame. If the distance between the target in the current frame and the set interval frame is greater than the set value, then the target is considered to be moving. The distance between the target in the current frame and the set interval frame refers to the distance between the target's center point in the current frame and the set interval frame.

[0081] Based on the characteristics of the material scheduling and transportation process, this invention does not use the calculation of the center point distance in every frame to determine whether the material has moved. Instead, it calculates the Euclidean distance of the center change after several frames. When the calculated distance is greater than a set threshold, it is determined that the material is moving.

[0082] Then, it is determined whether the target has entered the set area: the detection box of the set area identified by the target detection model in the video stream is set as a virtual portal. It is determined whether the coordinates of the tracked target have moved into the virtual portal. When the target moves into the virtual portal, the distance change between the target and the virtual portal is calculated. If the distance change between the two does not exceed the set distance threshold in consecutive set frames, it is determined that the target has entered the set area, so as to realize real-time monitoring of the area reached by the target during the material dispatching process.

[0083] The change in distance between the target and the virtual portal refers to the change in distance between the target's center point and the virtual portal. The result of the target entering the designated area is displayed in the surveillance video.

[0084] 2) Count the number of each type of target in each region and store the statistical results in the database.

[0085] Based on target detection, the number of various targets in the recognition area is counted. According to the different labeling numbers of different targets, the number of each number in the recognition video is counted, and the number is matched with the target category. The number of different targets identified is counted and written into the database.

[0086] This invention processes video data from the dispatching operation site using deep learning. Based on tracking and positioning, it judges the material dispatching process, identifying vehicles, material types, and whether they have entered specific areas. This reduces the need for personnel to perform software operations and communication when materials arrive at a specific area. It can also statistically analyze the quantity of materials identified within the monitored area and write it to a database for use by other software, reducing the workload of personnel information interaction and software operation. Furthermore, it can automatically collect information using monitoring equipment from various locations without adding extra equipment, reducing equipment and labor costs, improving work efficiency, meeting the dynamic identification needs of distributed regional material dispatching, and enhancing the operational awareness of the distributed regional material dispatching process.

Claims

1. A method for dynamic identification of material scheduling process based on deep learning, characterized in that, include: Acquire monitoring videos of various areas at the dispatching operation site, including frames of images of vehicles and materials during the material dispatching process; The target detection model is used to identify targets and defined areas in the surveillance video, and the detection bounding boxes of the targets and defined areas are obtained. The detection bounding boxes include coordinates and scale. Targets include materials and / or vehicles. Track the identified target; Set the detection box of the set area as a virtual portal, determine whether the coordinates of the tracked target have moved into the virtual portal, calculate the distance change between the target and the virtual portal when the target moves into the virtual portal, and if the distance change between the two does not exceed the set distance threshold in consecutive set frames, then it is determined that the target has entered the set area.

2. The method for dynamic identification of material scheduling process based on deep learning according to claim 1, characterized in that, The target detection model is also used to detect target categories. The target detection box also includes the target category, which includes material type and vehicle type. After identifying the targets in the surveillance video, the number of each type of target in each area is counted.

3. The method for dynamic identification of material scheduling process based on deep learning according to claim 2, characterized in that, The method also includes, After counting the number of various targets in each region, the statistical results are stored in the database.

4. The method for dynamic identification of material scheduling process based on deep learning according to claim 1, characterized in that, The target detection model is trained using a material scheduling target dataset, which includes images of empty vehicles, single material items, and vehicles loaded with materials.

5. The method for dynamic identification of material scheduling process based on deep learning according to claim 1 or 2, characterized in that, The target detection model adopts a single-stage target model.

6. The method for dynamic identification of material scheduling process based on deep learning according to claim 1 or 2, characterized in that, The target detection model adopts an improved YOLO model, which replaces the Bottleneck module in C2f of the YOLOv8 backbone network with the FasterBlock module of FasterNet.

7. The method for dynamic identification of material scheduling process based on deep learning according to claim 1, characterized in that, The tracking of the identified target is achieved using the DeepSort algorithm.

8. The method for dynamic identification of material scheduling process based on deep learning according to claim 1, characterized in that, The method also includes displaying the result of the target entering the designated area in the surveillance video.

9. The method for dynamic identification of material scheduling process based on deep learning according to claim 1, characterized in that, Before determining whether the coordinates of the tracked target have moved into the virtual portal, it is first determined whether the tracked target has moved. The method for determining whether the tracked target has moved is: calculate the distance between the target in the current frame and the set interval frame. If the distance between the target in the current frame and the set interval frame is greater than the set value, then it is determined that the target is moving.

10. The method for dynamic identification of material scheduling process based on deep learning according to claim 4, characterized in that, The target dataset for material dispatching is obtained by cleaning, labeling, and data augmenting images of vehicles and materials collected from the material dispatching system.

Images