Method and system for positioning production personnel of coal preparation plant based on video analysis

Abstract

The invention discloses a coal preparation plant production personnel positioning method and system based on video analysis, and the method comprises the steps: processing a video stream through employing an improved OfficientDet model, and obtaining an initial target personnel detection frame; processing the initial target person detection frame by using an improved FairMOT model and a dynamic motion model to obtain a continuous tracking trajectory of the target person; processing the continuous tracking trajectory of the target person by using a multiple linear regression model, predicting the position of the next frame, and allocating weights by using an entropy weight method to obtain a predicted motion trajectory; and processing the continuous tracking trajectory of the target person by using a ResNet-50 model to obtain a depth feature, performing identity confirmation of the target person based on the feature, and associating the identity with the predicted motion trajectory to complete person positioning. According to the method, tracking errors caused by environmental changes are reduced, and effective application in different environments is ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal preparation plant safety, and particularly relates to a coal preparation plant production personnel positioning method and system based on video analysis. BACKGROUND

[0002] With the continuous improvement of automation and intelligence level of coal preparation plants, the importance of video monitoring systems in safety management is increasingly prominent. Most coal preparation plants have installed video monitoring equipment to monitor potential safety hazards during production. However, the current monitoring system still mainly relies on manual monitoring, and there are a series of shortcomings.

[0003] Firstly, due to the complex production environment of coal preparation plants, including large equipment, narrow spaces and multi-story buildings, often leading to poor video signals, affecting the monitoring effect. Secondly, the accuracy of traditional video monitoring systems in personnel detection and tracking is low, especially in conditions of insufficient light, noise interference, etc. Manual monitoring is not only inefficient, but also prone to false negatives and false positives, and cannot respond to potential safety risks in real time. In addition, the existing system is easily affected by occlusion in motion human tracking, causing recognition errors and ID jumps, further affecting the effectiveness of safety management.

[0004] Based on the limitations of existing technologies in personnel tracking and identification, not only the safety production of coal preparation plants is affected, but also the further development of intelligent monitoring technology is hindered. Therefore, a new type of automated video analysis method is needed to improve the real-time and accuracy of personnel tracking. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a coal preparation plant production personnel positioning method and system based on video analysis, which solves the problem of inaccurate personnel tracking and positioning caused by environmental complexity in the prior art, and improves the real-time and accuracy of personnel positioning.

[0006] In order to solve the above technical problems, the present application adopts the following technical solutions:

[0007] A coal preparation plant production personnel positioning method based on video analysis, comprising the following steps:

[0008] S1, using an improved EfficientDet (Efficient Object Detection) model to process the video stream to obtain an initial target personnel detection frame.

[0009] S2, using an improved FairMOT (Fair Multiple Object Tracking) model and a dynamic motion model to process the initial target personnel detection frame to obtain a continuous tracking trajectory of the target personnel.

[0010] S3, processing the continuous tracking trajectory of the target person by using a multiple linear regression model, predicting a next frame position, assigning weights by using an entropy weight method, and obtaining a predicted motion trajectory.

[0011] S4, processing the continuous tracking trajectory of the target person by using a ResNet-50 (Residual Network-50) model, obtaining a deep feature, performing identity confirmation of the target person based on the feature and associating to the predicted motion trajectory, and completing positioning of the person.

[0012] Further, in step S1, obtaining the initial target person detection frame includes the following contents:

[0013] The parameters of the improved EfficientDet model are jointly optimized by using a compound loss function of dynamic data enhancement and focal loss and distance intersection over union loss.

[0014] Based on the optimized improved EfficientDet model, the confidence of the detection frame belonging to the target person is calculated by a classification branch, and the detection frame coordinates (x, y, w, h) are generated by a regression branch.

[0015] Wherein, x represents the horizontal direction position, y represents the vertical direction position, w represents the length in the horizontal direction, and h represents the length in the vertical direction.

[0016] When the confidence is >0.5, the detection frame corresponding to the confidence is retained to the candidate frame set, otherwise it is directly discarded.

[0017] The candidate frame set is sorted according to the confidence, the confidence attenuation of the overlapping frame is performed by using a Soft-NMS (Soft Non-Maximum Suppression) mechanism, and the detection frame with a confidence >0.2 is output as the initial target person detection frame.

[0018] Further, in step S2, obtaining the continuous tracking trajectory of the target person includes the following contents:

[0019] The detection branch and the ReID branch are synchronously optimized by using a single-stage joint training framework, and a light invariance module is added to the optimized ReID branch to obtain an improved FairMOT model.

[0020] The initial target personnel detection frame is input into the improved FairMOT model to generate a ReID appearance feature, the feature is weighted and fused with an additional color histogram, LBP (Local Binary Pattern) texture feature and HOG (Histogram of Oriented Gradients) shape feature to generate a comprehensive feature vector, a Mahalanobis distance and cosine similarity are used to obtain a comprehensive motion trajectory matching cost matrix, and an improved Hungarian algorithm is used to associate the initial target personnel detection frame with the matrix.

[0021] The dynamic motion model comprises long short-term memory network layers and gated recurrent unit layers connected in sequence.

[0022] The historical motion parameter sequence is extracted from the associated result, the sequence comprising speed, direction position and acceleration, the sequence is denoised by using Kalman filtering and then input into the dynamic motion model, the long short-term memory network layer is used to capture long-term motion dependence, and the gated recurrent unit layer is used to output a next-frame motion state prediction value to obtain a continuous tracking trajectory of the target personnel.

[0023] Further, in step S3, the predicted motion trajectory comprises the following contents:

[0024] The historical motion parameter sequence of the last 10 frames in the continuous tracking trajectory of the target personnel is extracted, and a training data set is constructed after Z-Score standardization.

[0025] The training data set is used to train a multiple linear regression model to obtain a trained multiple linear regression model, and the model is used to predict a next-frame position, and the specific formula is as follows:

[0026]

[0027] wherein, x t+1 represents a horizontal direction position at the t+1 moment, x t represents a horizontal direction position at the t moment, represents a speed in the x direction at the t moment, and Δt represents a time interval, represents an acceleration in the x direction at the t moment.

[0028] When the prediction error of the trained multiple linear regression model is greater than 10 pixels for 3 consecutive frames, a SVM (Support Vector Machine) nonlinear model is used for correction, and the specific formula is as follows:

[0029]

[0030] wherein, represents the horizontal position at the t+1 time after correction, Δx env represents the environmental compensation term.

[0031] According to the predicted speed, direction position and acceleration, combined with the light intensity and obstacle proportion, the feature weight is distributed by using the entropy weight method, and the weighted prediction position is obtained.

[0032] When the light intensity is <50 or the obstacle proportion is >0.3, the weight is adjusted, and the weighted prediction position is updated; when the obstacle is detected, the linear mapping relationship between the light intensity L and the motion weight W motion is established, and the acceleration weight is improved, and the weighted prediction position is updated. motion =0.4+0.2·L / 255.

[0033] Based on the final weighted prediction position, the predicted motion trajectory is output.

[0034] Further, in step S4, the positioning of the personnel includes the following contents:

[0035] The static image of the target personnel is obtained, the image is data enhanced, including rotation ±15°, zoom ±20% and horizontal flip, the enhanced image is used to train the ResNet-50 model, and the trained ResNet-50 model is obtained.

[0036] The personnel image is extracted from the continuous tracking trajectory of the target personnel, and the image is input into the trained ResNet-50 model to obtain a 512-dimensional deep feature.

[0037] The weighted cosine similarity S detect between the deep feature F motion and the motion feature F cos in the feature library is calculated, and the specific formula is:

[0038]

[0039] Wherein, S hist represents the historical matching similarity mean of the same tracking trajectory.

[0040] When S cos ≥0.85, the identity of the target personnel is confirmed and associated to the predicted motion trajectory, and the inverse regular is marked as unknown personnel.

[0041] When the occlusion rate of the detection box is >40% or the feature norm decreases by >30%, it is determined to be occluded; for the occlusion case, the distance D between the current detection box and the historical trajectory is calculated, and if D≤15 pixels, it is still associated to the predicted motion trajectory.

[0042] Further, the method further comprises a step S5 of monitoring the positioning result in real time, collecting the obtained relevant data, feeding the data to the management platform, and dynamically adjusting and monitoring.

[0043] Further, the application further provides a coal preparation plant production personnel positioning system based on video analysis, comprising:

[0044] The detection frame acquisition module is configured to process the video stream by using the improved EfficientDet model to obtain an initial target personnel detection frame.

[0045] The continuous tracking trajectory acquisition module is configured to process the initial target personnel detection frame by using the improved FairMOT model and the dynamic motion model to obtain a continuous tracking trajectory of the target personnel.

[0046] The motion trajectory acquisition module is configured to process the continuous tracking trajectory of the target personnel by using the multivariate linear regression model to predict a next frame position, and assign weights by using the entropy weight method to obtain a predicted motion trajectory.

[0047] The personnel positioning module is configured to process the continuous tracking trajectory of the target personnel by using the ResNet-50 model to obtain a deep feature, confirm the identity of the target personnel based on the feature, and associate the target personnel to the predicted motion trajectory, thereby completing the positioning of the personnel.

[0048] Further, the application further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the coal preparation plant production personnel positioning method based on video analysis.

[0049] Further, the application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is executable on a processor to execute the coal preparation plant production personnel positioning method based on video analysis.

[0050] Compared with the prior art, the application has the following technical effects:

[0051] (1) Improved detection and tracking accuracy: by using EfficientDet and a new redundant frame removal algorithm, the personnel detection rate in complex environments is significantly improved, the false detection rate is reduced, and the tracking accuracy is effectively guaranteed.

[0052] (2) Enhanced robustness: by combining the dynamic motion model and the multi-feature fusion strategy, the adaptability to fast movement and complex background is significantly improved, ensuring the stability and reliability of the system.

[0053] (3) Personalized motion trajectory prediction: The present application can more accurately predict the motion trajectory of the target by establishing a personalized motion model, reducing tracking errors caused by environmental changes.

[0054] (4) Identity recognition accuracy improvement: The present application establishes a comprehensive identity feature library and occlusion correction algorithm, enabling high identity recognition accuracy in occlusion or recognition error conditions, improving overall practicality.

[0055] (5) Adaptive optimization capability: Real-time data feedback and optimization mechanism enable the present application to dynamically adjust according to actual operation conditions, maintain algorithm adaptability and continuous performance improvement, and ensure effective application in different environments. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is the overall implementation flowchart of the present application. DETAILED DESCRIPTION

[0057] The present application will be further described below in conjunction with the drawings. The following examples are only used to more clearly illustrate the technical solutions of the present application, and cannot limit the protection scope of the present application.

[0058] To achieve the above purpose, the present application proposes a coal preparation plant production personnel positioning method based on video analysis, as shown in Figure 1 The specific steps are as follows:

[0059] S1, use the improved EfficientDet (Efficient Object Detection, high-efficiency target detection) model to process the video stream to obtain the initial target personnel detection frame. Specifically:

[0060] The improved EfficientDet model parameters are jointly optimized by using dynamic data enhancement and the composite loss function of focal loss and distance intersection over union loss.

[0061] Based on the optimized improved EfficientDet model, the confidence (0-1) of the detection frame belonging to the target personnel is calculated through the classification branch, and the detection frame coordinates (x, y, w, h) are generated through the regression branch.

[0062] Wherein, x represents the horizontal direction position, y represents the vertical direction position, w represents the horizontal direction length, and h represents the vertical direction length.

[0063] When the confidence is >0.5, the detection frame corresponding to the confidence is retained to the candidate frame set, otherwise it is directly discarded.

[0064] The candidate box set is sorted by confidence, a Soft-NMS (Soft Non-Maximum Suppression) mechanism is used to attenuate the confidence of overlapping boxes, and a detection box with a confidence greater than 0.2 is output as an initial target personnel detection box.

[0065] S2, the initial target personnel detection box is processed by using the improved FairMOT (Fair Multiple Object Tracking) model and the dynamic motion model to obtain the continuous tracking trajectory of the target personnel. Specifically:

[0066] The detection branch and the ReID branch are simultaneously optimized by using a single-stage joint training framework, and a light invariance module is added to the optimized ReID branch to obtain the improved FairMOT model.

[0067] The initial target personnel detection box is input into the improved FairMOT model to generate a ReID appearance feature, the feature is weightedly fused with an additional color histogram (HSV space), an LBP (Local Binary Pattern) texture feature, and an HOG (Histogram of Oriented Gradients) shape feature to generate a comprehensive feature vector, wherein the weights are 0.3, 0.2, 0.2, and 0.3 respectively, a comprehensive motion trajectory matching cost matrix is obtained by using Mahalanobis distance and cosine similarity to improve the cross-frame matching accuracy, and the initial target personnel detection box is associated with the matrix by using an improved Hungarian algorithm.

[0068] The dynamic motion model includes a long short-term memory network layer and a gated recurrent unit layer connected in sequence.

[0069] A history motion parameter sequence including velocity, direction position, and acceleration is extracted from the associated processing result, the sequence is denoised by using Kalman filtering and then input into the dynamic motion model, a long-term motion dependence is captured by using the long short-term memory network layer, a next-frame motion state prediction value is output by using the gated recurrent unit layer, and a continuous tracking trajectory of the target personnel is obtained.

[0070] S3, the continuous tracking trajectory of the target personnel is processed by using a multivariate linear regression model to predict a next-frame position, and an entropy weight method is used to assign weights to obtain a predicted motion trajectory. Specifically:

[0071] A history motion parameter sequence of the last 10 frames in the continuous tracking trajectory of the target personnel is extracted, and a training data set is constructed after Z-Score standardization.

[0072] The multivariate linear regression model is trained using the training data set to obtain a trained multivariate linear regression model, and the model is used to predict the next frame position, and the specific formula is:

[0073]

[0074] Wherein, x t+1 represents the horizontal position at t+1 time, x t represents the horizontal position at t time, represents the velocity in x direction at t time, and Δt represents the time interval, represents the acceleration in x direction at t time.

[0075] When the prediction error of the trained multivariate linear regression model is greater than 10 pixels for 3 consecutive frames, the SVM nonlinear model is used for correction, and the specific formula is:

[0076]

[0077] Wherein, represents the corrected horizontal position at t+1 time, Δx env represents the environmental compensation term.

[0078] According to the predicted velocity, direction position and acceleration, combined with the light intensity and obstacle ratio, the entropy weight method is used to distribute the feature weight to obtain the weighted prediction position. Wherein, the weight W = [0.3, 0.2, 0.2, 0.15, 0.15].

[0079] Wherein, when the light intensity is less than 50 or the obstacle ratio is greater than 0.3, the weight is adjusted to W' = [0.4, 0.3, 0.1, 0.1, 0.1], and the weighted prediction position is updated; when the obstacle is detected, the linear mapping relationship between the light intensity L and the motion weight W motion is established, and the specific is W motion = 0.4 + 0.2·L / 255, and the acceleration weight is increased to 0.3, and the weighted prediction position is updated.

[0080] Based on the final weighted prediction position, the predicted high-precision motion trajectory with an error of less than 5 pixels is output.

[0081] S4, using ResNet-50 (Residual Network-50, residual network-50) model to process the continuous tracking trajectory of the target personnel, obtain the deep feature, based on the feature to confirm the identity of the target personnel and associate to the predicted motion trajectory, complete the positioning of the personnel. Specifically:

[0082] A static image of the target person is acquired, and data augmentation is performed on the image, including rotation ± 15°, scaling ± 20%, and horizontal flipping, to improve the generalization ability of the model. The enhanced image is used to train the ResNet-50 model, and a trained ResNet-50 model is obtained.

[0083] A person image is extracted from the continuous tracking trajectory of the target person, and the image is input into the trained ResNet-50 model to obtain a 512-dimensional deep feature.

[0084] The deep feature F detect is calculated. motion The weighted cosine similarity S cos between the feature F detect and the motion feature F motion in the feature library is calculated, and the specific formula is:

[0085]

[0086] where S hist represents the historical matching similarity mean value of the same tracking trajectory.

[0087] When S cos ≥ 0.85, the identity of the target person is confirmed and associated with the predicted motion trajectory, and the inverse regular is marked as an unknown person.

[0088] When the occlusion rate of the detection box is > 40% or the feature norm decreases by > 30%, it is determined to be occluded; for the occlusion case, the distance D between the current detection box and the historical trajectory is calculated, and if D ≤ 15 pixels, it is still associated with the predicted motion trajectory.

[0089] S5, real-time monitoring of positioning results, collecting relevant data, feeding back the data to the management platform, so as to facilitate the manager to make dynamic adjustment and monitoring; regularly analyzing data, optimizing and adjusting the positioning process; using online learning or incremental learning method to ensure that the model can adapt to new environmental conditions and personnel changes, thereby improving the efficiency and accuracy of the overall system.

[0090] Embodiment:

[0091] A production workshop of a coal preparation plant with an area of about 2000m 2 deployed 20 high-definition cameras (1080P@30fps) to monitor the positions and identities of 50 workers in real time. The scene characteristics include:

[0092] High dynamicity: frequent movement of personnel (speed 0.5-2m / s), high-speed operation of equipment mechanical arm;

[0093] Complex occlusion: coal pile, conveyor belt, equipment occlusion rate > 40% area ratio 30%;

[0094] Light change: there is dust in some areas to block the light source, and the light intensity fluctuates in the range of 30-200 Lux.

[0095] After adopting the Soft-NMS mechanism to attenuate the confidence of the overlapping frame, the number of detection frames is reduced from an average of 25 per frame to 18, which reduces the false detection rate by 28%.

[0096] The ID uniqueness retention rate of the continuous tracking trajectory of the target personnel is 98.5%. In the shielding area of the mechanical arm, the prediction compensation of the long short-term memory network layer makes the trajectory interruption rate <2%.

[0097] The prediction error of the SVM nonlinear model in the shielding area is ≤8 pixels. The multi-view feature library matching success rate is 92%, and the recognition accuracy after shielding correction is improved to 88%.

[0098] The model is optimized every 30 minutes, and the detection accuracy in the dust environment is improved from 85% to 89%.

[0099] The positioning results obtained by the present application are compared with the positioning results obtained by the method of YOLOv4(You Only Look Once version 4, you only look once version 4)+DeepSORT(Deep Simple Online and Realtime Tracking, deep simple online and real-time tracking), and the corresponding data are shown in Table 1.

[0100] Table 1 Related data of positioning results obtained by different methods

[0101] Indicators The present invention YOLOv4+DeepSORT Detection accuracy (mAP) 92％ 78％ Trajectory continuity rate (1 hour) 98.5％ 82％ Shading scene recognition accuracy 88％ 65％ System delay (1080P video) ≤ 200 ms ≥ 500 ms

[0102] As can be seen from Table 1, the present application is higher than the existing method in terms of detection accuracy, trajectory continuity rate, shielding scene recognition accuracy and system delay.

[0103] The present application also proposes a coal preparation plant production personnel positioning system based on video analysis, which includes a detection frame acquisition module, a continuous tracking trajectory acquisition module, a motion trajectory acquisition module, a personnel positioning module and a computer program executable on a processor. It should be noted that each module in the above system corresponds to the specific steps of the method provided by the present application, has the corresponding function modules and beneficial effects of executing the method. Technical details not described in detail in the present embodiment can be referred to the method provided by the present application.

[0104] The embodiment of the present application further provides an electronic device, comprising a memory, a processor and a computer program stored in the memory and capable of running on the processor. It should be noted that the processor executes the computer program to correspond to the specific steps of the method provided by the embodiment of the present application, has the function modules and beneficial effects corresponding to the execution method. The technical details not described in detail in the embodiment can be referred to the method provided by the embodiment of the present application.

[0105] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program. It should be noted that the computer program is run by the processor to correspond to the specific steps of the method provided by the embodiment of the present application, has the function modules and beneficial effects corresponding to the execution method. The technical details not described in detail in the embodiment can be referred to the method provided by the embodiment of the present application.

[0106] The above is only the preferred embodiment of the present application, it should be pointed out that, for the ordinary skilled in the art, without departing from the technical principles of the present application, can make a number of improvements and deformation, these improvements and deformation also should be considered as the protection scope of the present application.

Claims

1. A method for locating production personnel in a coal preparation plant based on video analytics, characterized in that, include: S1. The video stream is processed using the improved EfficientDet model to obtain the initial target person detection box; S2. The initial target personnel detection box is processed using the improved FairMOT model and dynamic motion model to obtain the continuous tracking trajectory of the target personnel. S3. The continuous tracking trajectory of the target person is processed using a multiple linear regression model to predict the position of the next frame. The weights are then assigned using the entropy weight method to obtain the predicted motion trajectory. S4. The ResNet-50 model is used to process the continuous tracking trajectory of the target person to obtain depth features. Based on these features, the identity of the target person is confirmed and associated with the predicted motion trajectory to complete the localization of the person.

2. The method for locating production personnel in a coal preparation plant based on video analysis according to claim 1, characterized in that, In step S1, the initial target personnel detection box includes the following: The parameters of the improved EfficientDet model are jointly optimized using dynamic data augmentation and a composite loss function consisting of focus loss and distance intersection-union ratio loss. Based on the optimized and improved EfficientDet model, the confidence score of the detection box belonging to the target person is calculated through the classification branch, and the coordinates (x,y,w,h) of the detection box are generated through the regression branch. Where x represents the horizontal position, y represents the vertical position, w represents the horizontal length, and h represents the vertical length; If the confidence level is greater than 0.5, the detection box corresponding to that confidence level is retained in the candidate box set; otherwise, it is discarded. The candidate box set is sorted by confidence level, and the confidence level of overlapping boxes is reduced using the Soft-NMS mechanism. The detection boxes with confidence level > 0.2 are output as the initial target personnel detection boxes.

3. The method for locating production personnel in a coal preparation plant based on video analysis according to claim 1, characterized in that, In step S2, the continuous tracking trajectory of the target person is obtained, including the following: The detection branch and ReID branch are optimized simultaneously using a single-stage joint training framework, and an illumination invariant module is added to the optimized ReID branch to obtain an improved FairMOT model. The initial target person detection box is input into the improved FairMOT model to generate ReID appearance features. These features are then weighted and fused with additional color histogram, LBP texture features, and HOG shape features to generate a comprehensive feature vector. The comprehensive motion trajectory matching cost matrix is ​​obtained using Mahalanobis distance and cosine similarity. The improved Hungarian algorithm is then used to associate the initial target person detection box with this matrix. The dynamic motion model consists of a long short-term memory network layer and a gated recurrent unit layer connected in sequence; Historical motion parameter sequences, including velocity, orientation, position, and acceleration, are extracted from the results of the correlation processing. After denoising the sequence using Kalman filtering, the sequence is input into the dynamic motion model. Long short-term memory network layers are used to capture long-term motion dependencies, and gated recurrent unit layers are used to output the predicted motion state value of the next frame, thus obtaining the continuous tracking trajectory of the target person.

4. The method for locating production personnel in a coal preparation plant based on video analysis according to claim 3, characterized in that, In step S3, the predicted motion trajectory includes the following: Extract the historical motion parameter sequence of the most recent 10 frames from the continuous tracking trajectory of the target person, and construct the training dataset after Z-Score normalization; The multiple linear regression model is trained using the training dataset to obtain the trained model. This model is then used to predict the position of the next frame. The specific formula is as follows: Where, x t+1 Let x represent the horizontal position at time t+1. t This represents the horizontal position at time t. Let Δt represent the velocity in the x-direction at time t, and Δt represent the time interval. This represents the acceleration in the x-direction at time t; When the prediction error of the trained multiple linear regression model is greater than 10 pixels for three consecutive frames, an SVM nonlinear model is used for correction. The specific formula is as follows: in, Δx represents the corrected horizontal position at time t+1. env Indicates environmental compensation items; Based on the predicted velocity, direction, position, and acceleration, and combined with the light intensity and the proportion of obstacles, the entropy weight method is used to assign feature weights to obtain the weighted predicted position. Specifically, when the light intensity is <50 or the obstacle ratio is >0.3, the weights are adjusted and the weighted predicted position is updated; when an obstacle is detected, a relationship is established between the light intensity L and the motion weight W. motion The linear mapping relationship, specifically W motion =0.4 + 0.2·L / 255, and increase the acceleration weight to update the weighted prediction position; Based on the final weighted predicted position, the predicted motion trajectory is output.

5. The method for locating production personnel in a coal preparation plant based on video analysis according to claim 1, characterized in that, In step S4, the location of personnel includes the following: Obtain static images of the target personnel, perform data augmentation on the images, and use the augmented images to train the ResNet-50 model to obtain the trained ResNet-50 model. The image of the person is extracted from the continuous tracking trajectory of the target person, and the image is input into the trained ResNet-50 model to obtain 512-dimensional deep features; Calculate depth features F detect Compared with motion features F in the feature library motion Weighted cosine similarity S cos The specific formula is as follows: Among them, S hist This represents the average historical matching similarity score for the same tracking trajectory. When S cos When the value is ≥0.85, the identity of the target person is confirmed and associated with the predicted movement trajectory; otherwise, it is marked as an unknown person. When the occlusion rate of the detection box is greater than 40% or the feature norm decreases by more than 30%, it is determined to be occlusion. For occlusion, the distance D between the current detection box and the historical trajectory is calculated. If D ≤ 15 pixels, it is still associated with the predicted motion trajectory.

6. The method for locating production personnel in a coal preparation plant based on video analysis according to claim 5, characterized in that, Data augmentation includes rotation ±15°, scaling ±20%, and horizontal flipping.

7. The method for locating production personnel in a coal preparation plant based on video analysis according to claim 1, characterized in that, It also includes step S5: real-time monitoring of positioning results, collection of relevant data, and feedback of the data to the management platform for dynamic adjustment and monitoring.

8. A system applied to the video analysis-based personnel positioning method for coal preparation plants as described in claim 1, characterized in that, include: The detection box acquisition module is used to process the video stream using the improved EfficientDet model to obtain the initial target person detection box; The continuous tracking trajectory acquisition module is used to process the initial target person detection box using the improved FairMOT model and dynamic motion model to obtain the continuous tracking trajectory of the target person. The motion trajectory acquisition module is used to process the continuous tracking trajectory of the target person using a multiple linear regression model, predict the position of the next frame, and allocate weights using the entropy weight method to obtain the predicted motion trajectory. The personnel localization module uses the ResNet-50 model to process the continuous tracking trajectory of the target personnel, obtain depth features, and then uses these features to confirm the identity of the target personnel and associate them with the predicted motion trajectory to complete the personnel localization.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the video analysis-based coal preparation plant production personnel positioning method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, The computer program, when executed by the processor, performs the video analysis-based method for locating production personnel in a coal preparation plant, as described in any one of claims 1 to 7.

Images