Internet of Things video monitoring scene adaptive adjustment method combining big data mining and deep learning

Abstract

The invention discloses an Internet of Things video monitoring scene adaptive adjustment method combining big data mining and deep learning, and belongs to the technical field of Internet of Things video monitoring. The method comprises the following steps: firstly, acquiring multi-source data such as a video stream and a pan-tilt attitude, and performing association fusion and preprocessing to form an effective data set; a typical scene is obtained through clustering of an improved D-CURE algorithm, scene classification is achieved in combination with a random forest, and initial monitoring parameters are output; then, parameters are dynamically corrected by using a WLSTM model, and video coding and transmission collaborative optimization are completed by relying on DHC-Net; and finally, continuously optimizing the system through multi-dimensional evaluation and iterative feedback. The problems that traditional monitoring fixed parameters are poor in adaptability, data mining is insufficient and the like are solved, the monitoring image quality and transmission stability are improved, the operation and maintenance cost is reduced, and the method is suitable for a multi-scene intelligent monitoring system for urban security and traffic control and the like.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of Internet of Things video monitoring, and particularly relates to an Internet of Things video monitoring scene adaptive adjustment method combining big data mining and deep learning technology, which is suitable for an intelligent monitoring system that needs to dynamically adapt to complex environments and multi-source data, can realize monitoring parameter optimization, video coding efficiency improvement and transmission performance collaborative optimization, and is widely applied to multi-scene Internet of Things video monitoring deployment scenes such as city security, traffic control, park management and industrial plant monitoring. BACKGROUND

[0002] The current Internet of Things video monitoring system generally faces the problem of insufficient scene adaptation capability. Traditional monitoring schemes mostly use fixed parameter configurations, which are difficult to cope with complex and variable environments and scene differences. In terms of data processing, the multi-source data generated by Internet of Things monitoring, such as video stream, pan-tilt attitude, environmental perception and network transmission, has the characteristics of strong heterogeneity and high redundancy. The existing technology lacks an efficient multi-source data correlation fusion and cleaning mechanism, resulting in insufficient data value mining. In the scene analysis link, traditional clustering and classification algorithms are difficult to accurately depict the dynamic changes and multi-dimensional characteristics of the monitoring scene, and are prone to scene recognition deviation, which further leads to mismatch between monitoring parameter configuration and actual scene demand. In terms of parameter adjustment and coding transmission, the existing schemes mostly use static parameter settings and single coding transmission strategies, which cannot be dynamically adjusted according to real-time changes such as scene complexity and network bandwidth fluctuation, often resulting in problems such as video quality blurring, high transmission delay and waste of bandwidth resources. At the same time, the existing system lacks an effective iterative feedback mechanism, making it difficult to continuously optimize the model and parameters according to long-term running performance, resulting in a gradual decline in the adaptation capability of the system in long-term use, and failing to meet the needs of Internet of Things monitoring for intelligence, adaptability and efficiency.

[0003] Therefore, the present application proposes an Internet of Things video monitoring scene adaptive adjustment method and system combining big data mining and deep learning to solve the above problems. SUMMARY

[0004] This section aims to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification of the present application to avoid obscuring the purpose of this section, abstract and title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0005] In view of the above and / or problems existing in the prior art, the present application is proposed.

[0006] Therefore, the purpose of the present application is to overcome the deficiencies in the prior art and provide an Internet of Things video monitoring scene adaptive adjustment method combining big data mining and deep learning, characterized by comprising the following steps:

[0007] Step S1: Multi-source monitoring data acquisition and preprocessing. Collect video streams, PTZ attitude, environmental perception and network transmission data in IoT video surveillance scenarios. Through the association and fusion of "device identifier - timestamp", and after improved low-pass filtering cleaning, coordinate transformation and feature enhancement, preprocessed data is formed.

[0008] Step S2: Scene clustering analysis based on the improved D-CURE algorithm. Redundant features are removed from the preprocessed data in Step S1 using the Pearson coefficient. A weighted multimodal similarity distance is constructed. Abnormal scenes are removed by DBSCAN pre-clustering, CURE clustering optimization and Davidson-Dinburg index are used to determine the optimal number of clusters, and typical monitoring scenes are output.

[0009] Step S3: Scene classification prediction based on random forest. Using the clustering results of step S2 as labels, construct a feature set and filter key features through out-of-bag error. Train the RF model of CART base learner, introduce similar sample weighting to correct and optimize the prediction results, and output scene classification and initial monitoring parameters.

[0010] Step S4: Dynamic correction of monitoring parameters based on WLSTM. Based on the results of step S3, a time series dataset of monitoring parameters is constructed. After improved normalization preprocessing, the LSTM hyperparameters are optimized using the whale optimization algorithm. The WLSTM model is trained and parameter optimization is achieved through rolling prediction + dynamic correction.

[0011] Step S5: Optimize video coding and transmission based on dynamic hierarchical segmentation network. Based on the parameter correction in step S4, evaluate the video coding complexity and classify it. Accelerate coding through DHC-Net dynamic weight segmentation and achieve collaborative optimization by combining adaptive bitrate control and dynamic transmission protocol.

[0012] Step S6: System Iteration Update and Feedback Optimization. Based on the results of step S5, a multi-dimensional performance evaluation model is constructed to locate the source of anomalies. Transmission parameters, incremental training models, and coding strategies are updated in a targeted manner to achieve continuous optimization of system performance.

[0013] Preferably, step S1 includes: acquiring data covering video streams with resolutions from 1080P to 4K and frame rates from 25 to 60fps, and the horizontal rotation angle of the PTZ. Vertical corner ,illumination ,humidity Wind speed ,visibility ,bandwidth Delay Packet loss rate A 28-dimensional initial matrix is ​​formed by associating "device identifier - timestamp";

[0014] Data cleaning adopts improved low-pass filter: , (video gray level), (pantilt angle), (environmental illumination), (network bandwidth);

[0015] Coordinate conversion:

[0016]

[0017] wherein, is focal length, is image width and height, is field of view; feature enhancement adds optical flow mean , motion vector , improved SSIM: .

[0018] Preferably, the step S2 comprises: feature screening adopts Pearson coefficient: , Eliminate redundancy, retain 15-dimensional core features; similarity measure: , (video dynamic), (environment), (pantilt), (network); DBSCAN pre-clustering: , Core sample determination ; representative point contraction , inter-cluster distance ; optimal cluster number is determined by DBI: , , .

[0019] Preferably, the step S3 comprises: constructing an 18-dimensional initial feature set, screening key features by out-of-bag error: , , retaining 12-dimensional features; RF model uses CART as base learner, Gini index: ; hyperparameters: , , , Bootstrap sampling and random feature selection are adopted; prediction correction: .

[0020] Preferably, the step S4 comprises: constructing a time series data set, time step 10 seconds, monitoring parameters , etc., improved normalization: ​; the input layer 240 neurons of the WLSTM model, the WOA optimized hyperparameters , objective function: ;

[0021] Preferably, the step S5 comprises: encoding complexity evaluation: , high (H), medium (M) and low (L) levels; the DHC-Net adopts 3D separable convolution: depth convolution , point-by-point convolution formula: , dynamic weight screening formula ; high complexity , medium complexity , low complexity ; the collaborative optimization adopts ABR algorithm: ; high complexity , medium complexity , low complexity , packet loss rate , time delay jitter . The application is a kind of Internet of Things video monitoring scene self-adaptive adjustment method combining big data mining and deep learning, which has the following beneficial effects: 1. Through multi-source data association fusion and improved D-CURE clustering algorithm, the monitoring scene is accurately classified, and the random forest classification prediction and WLSTM dynamic parameter correction are combined, so that the monitoring parameters can match the scene changes in real time, effectively solve the problem of disconnection between traditional fixed parameter configuration and scene demand, and significantly improve the monitoring picture quality and scene coverage effectiveness.

[0022] 2. Based on the encoding strategy and adaptive code rate control of dynamic hierarchical segmentation network (DHC-Net), dynamic transmission protocol collaborative optimization, the encoding and transmission scheme can be flexibly adjusted according to the video encoding complexity and network state, which reduces the encoding time and network bandwidth occupation, reduces the transmission delay and packet loss rate, and improves the overall transmission stability under the premise of ensuring the video quality.

[0023] 3. Through multi-dimensional performance evaluation and abnormal diagnosis mechanism, the problem root in system running can be accurately located, combined with incremental training, parameter dynamic updating and full iteration optimization, so that the system model and strategy can be continuously improved with scene changes and running data accumulation, ensuring the adaptation ability and performance stability of long-term system running, and reducing the maintenance cost in later period.

[0024] BRIEF DESCRIPTION OF DRAWINGS

[0025] BRIEF DESCRIPTION OF DRAWINGS

[0026] ​​To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0027] Figure 1 This diagram illustrates an adaptive adjustment method for IoT video surveillance scenarios that combines big data mining and deep learning.

[0028] Figure 2 This is a flowchart of step S1, multi-source monitoring data acquisition and preprocessing, in this invention;

[0029] Figure 3 This is a flowchart of step S2, scene clustering analysis based on the improved D-CURE algorithm, in this invention;

[0030] Figure 4 This is a flowchart of step S3, scene classification prediction based on random forest (RF), in this invention;

[0031] Figure 5 This is a flowchart of step S4, dynamic correction of monitoring parameters based on WLSTM, in this invention;

[0032] Figure 6 This is a flowchart of step S5 in the present invention, which involves optimizing video encoding and transmission based on a Dynamic Hierarchical Segmentation Network (DHC-Net).

[0033] Figure 7 This is a flowchart of step S6, system iterative update and feedback optimization in this invention;

[0034] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0035] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings of the embodiments of the present invention. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0036] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0037] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0038] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the examples in the specification.

[0039] This invention proposes an adaptive adjustment method for IoT video surveillance scenarios that combines big data mining and deep learning, characterized by the following steps:

[0040] Step S1: Multi-source monitoring data acquisition and preprocessing;

[0041] Step S2: Scene clustering analysis based on the improved D-CURE algorithm;

[0042] Step S3: Scene classification prediction based on random forest (RF);

[0043] Step S4: Dynamic correction of monitoring parameters based on WLSTM;

[0044] Step S5: Optimize video coding and transmission based on Dynamic Hierarchical Segmentation Network (DHC-Net);

[0045] Step S6: System iterative update and feedback optimization;

[0046] In a preferred embodiment of the present invention, in step S1, multi-source monitoring data acquisition and preprocessing, S11, multi-dimensional heterogeneous data acquisition and correlation fusion are performed first, followed by S12, data cleaning and feature enhancement processing.

[0047] In step S11, multi-dimensional heterogeneous data acquisition and correlation fusion, multi-source core data from IoT video surveillance scenarios are collected, including video stream data output by network cameras and real-time attitude data of PTZ devices, such as horizontal rotation angle. Vertical corner angular velocity of rotation The environmental sensing module collects multi-parameter environmental data, such as light intensity. Atmospheric humidity Wind speed ,visibility and network transmission link data, such as real-time bandwidth. Transmission delay Packet loss rate Based on the association logic between flight number and timestamp, a dual-key field of "device identifier - timestamp" is constructed to accurately associate the frame sequence number in video stream data, the angle information in gimbal attitude data, the sampling time in environmental data, and the transmission timestamp in network data, forming a multi-source data set under a unified time dimension. For video stream data, the RGB three-channel pixel matrix and grayscale value distribution of each frame are extracted; for gimbal attitude data, the angle change between adjacent sampling times is recorded; for environmental data, a sliding time window is used to calculate the mean and variance to smooth short-term fluctuations; for network data, the bandwidth fluctuation amplitude per unit time is statistically analyzed. With latency jitter This ultimately forms an initial data matrix containing 28-dimensional features.

[0048] In step S12, data cleaning and feature enhancement processing, an improved low-pass filtering algorithm is first used to remove outliers from the initial data matrix. Combining the smoothing concept of flight track data with the characteristics of monitoring data, the smoothing factor is adjusted to adapt to different types of data. The formula is as follows: ,in, For the first Filtered data from each sampling point This is the original sampled data. To address the smoothing factors for different data types, differentiated smoothing factor settings are used to reduce noise interference in different data types while preserving data trends.

[0049] The video frame data undergoes coordinate transformation, converting the camera's imaging coordinate system to the world coordinate system to facilitate subsequent scene spatial analysis. The formula is as follows:

[0050]

[0051] in, For the camera lens focal length, These are the image width and height, respectively. These are the horizontal and vertical field of view of the lens, respectively. This is the image center offset. This refers to the horizontal and vertical rotation angles of the gimbal, which is used to achieve the correspondence between the image pixel position and the real spatial position.

[0052] Feature enhancements were performed to improve data discriminability, including the addition of inter-frame dynamic features and scene complexity features. Inter-frame dynamic features include the average optical flow field intensity of adjacent frames. With motion vector amplitude The formula is as follows:

[0053]

[0054]

[0055] in, Image pixel size Each pixel Horizontal and vertical optical flow components, For adjacent frame pixels The coordinate change. Scene complexity features are derived by combining an improved structural similarity (SSIM) and color histogram distance, as shown in the following formula:

[0056]

[0057] in, These represent the covariance and variance of the three RGB channels, respectively. The histograms of the RGB three channels of adjacent frames are shown respectively. Through multi-dimensional feature enhancement, the dynamic changes and complexity differences of the monitoring scene are effectively depicted.

[0058] In a preferred embodiment of the present invention, in step S2, scene clustering analysis based on the improved D-CURE algorithm, firstly, S21, scene similarity measurement model with multi-feature fusion is constructed, and then S22, scene clustering is implemented using the improved D-CURE algorithm.

[0059] In the construction of the scene similarity measurement model involving S21 and multi-feature fusion, based on the 28-dimensional feature data matrix obtained after preprocessing in S1, the correlation between features is first analyzed using the Pearson coefficient, and the remainder feature is divided to reduce computational complexity. The correlation calculation formula is as follows:

[0060]

[0061] in, Represents any two types of features, The number of scene samples. For each scene sample, the feature dimension, when Features that are determined to be strongly correlated are removed, and features with high information redundancy are finally selected to obtain 15 core features for similarity calculation.

[0062] Taking into account the heterogeneity of multiple features in monitoring scenarios, a weighted multimodal similarity distance is constructed, as shown in the following formula:

[0063]

[0064] in, Two scene samples, For the first The weights of class features are determined through random forest feature importance evaluation, and the weights of video dynamic features are also determined. Environmental feature weights Gimbal attitude feature weights Network feature weights , For the first The number of sub-dimensions of a class feature. For the scene No. The first class feature Individual sub-dimensional data, and This is the mean of the sub-dimension across all samples. This formula achieves a standardized measurement of different types of features, while introducing feature weights to reflect the differences in the contribution of each dimension to scene similarity. Compared to traditional Euclidean distance, it is more in line with the characteristics of multi-source data in monitoring scenarios.

[0065] In the scene clustering implementation of the improved D-CURE algorithm described in S22, DBSCAN pre-clustering is first performed to remove abnormal scene samples. A scene similarity matrix is ​​then constructed based on the WMSD calculated in S201, and a neighborhood radius is set. With minimum sample size : pass Determined by distance method The scene sample density is set to 8. The core scene sample determination formula is as follows: .like ,but The core samples are merged to form an initial cluster based on density reachability. Non-core samples that are not density reachable are marked as abnormal scenarios. The proportion of abnormal samples is controlled within 3% to effectively reduce subsequent clustering interference.

[0066] Next, CURE clustering optimization was performed, employing a "stratified sampling + representative point shrinkage" strategy for the initial clusters: 12 representative points were randomly selected from each initial cluster, and the shrinkage factor was used to optimize the clusters. Adjust the position of the representative point to align with the cluster center trend; the shrinkage formula is as follows: ,in, As the original representative point, As the cluster center, When calculating inter-cluster distances, a modified average distance formula is used, which combines the distance between representative points and the distance between cluster centers. The formula is as follows: ,in, Clusters The representative set of points, The number of representative points is used. Clusters with the smallest distance are iteratively merged until the optimal number of clusters is achieved. The optimal number of clusters is determined by the Davidson-Dinburg index, as shown in the following formula: ,in, For clusters The mean of the inner sample , Between cluster centers .

[0067] In a preferred embodiment of the present invention, in step S3, scene clustering analysis based on the improved D-CURE algorithm, the following steps are performed: S31, scene classification feature set construction and optimization; S32, random forest classification model training; and finally S33, scene classification prediction and result optimization.

[0068] In S31, the construction and optimization of the scene classification feature set, based on the nine typical monitoring scenes and corresponding sample data output by S2, multi-source features are first integrated to form an initial feature set. This feature set includes the 15 core features selected in S201, and adds scene temporal correlation features—the feature change rate of adjacent scene samples, such as the light intensity change rate. Gimbal rotation rate Bandwidth fluctuation rate Construct an initial input feature matrix containing 18-dimensional features.

[0069] To further eliminate redundant features and improve the model's generalization ability, out-of-bag error is used to calculate feature importance, referencing the feature importance evaluation logic in track classification. The formula is as follows: Where, For a certain feature to be evaluated, For the total number of decision trees, Features After the value is randomly shuffled, the first OOB error in decision trees Features When the value is not shuffled OOB error in decision trees.

[0070] In the training of the S32 and random forest classification models, the 9 scene categories obtained by S2 clustering are used as classification labels. The feature matrix optimized by S31 is divided into training set and test set in an 8:2 ratio, and the sample distribution ratio of each scene category in the training set and test set is kept consistent.

[0071] When constructing the random forest model, CART is used as the base learner, and the Gini index is used as the node splitting criterion, as shown in the following formula:

[0072]

[0073] in, For the current node's sample set, for The middle belongs to the first A subset of samples from the field. As a characteristic splitting attribute, Features The two subsets obtained after splitting. The model hyperparameter, number of decision trees, is optimized using a grid search method. Maximum depth of decision tree Minimum number of samples for node splitting Minimum number of samples for leaf nodes .

[0074] During training, the Bootstrap sampling method with replacement is used to generate training samples for each decision tree, and only 60% of the key features of each decision tree are randomly selected to participate in node splitting. At the same time, class weights are introduced to address the problem of insufficient samples in minority scenarios and improve the model's classification ability for niche scenarios.

[0075] In step S33, scene classification prediction and result optimization, the test set features are input into the trained random forest model, and the scene classification result is output through a "majority voting" mechanism: each decision tree predicts the category of the test sample, the frequency of all decision tree predictions is counted, and the category with the highest frequency is taken as the final predicted category. To further improve prediction accuracy, a "weighted correction for similar samples" strategy is introduced—for samples with a prediction confidence score below 0.85, the top 5 samples with the highest feature similarity to that sample in the training set are found. The formula is as follows: ,in, For test samples, For the samples in the training set, For the first Importance weights of key features The test sample and the training sample are respectively the first and second samples. 3D feature values. Based on the category distribution of the top 5 similar samples, the original prediction results are weighted and corrected using the following formula: ,in, The weights of the original prediction results Category in the original prediction The percentage of votes, For the first The similarity between each similar sample and the test sample. This is an indicator function.

[0076] In a preferred embodiment of the present invention, in step S4, dynamic correction of monitoring parameters based on WLSTM, the following steps are performed: S41, construction and preprocessing of time series dataset of monitoring parameters; S42, optimization of LSTM hyperparameters using the whale optimization algorithm; and finally S43, training of WLSTM model and dynamic correction of monitoring parameters.

[0077] In S41, the construction and preprocessing of the monitoring parameter time-series dataset, based on the 9-scene classification results and corresponding initial monitoring parameter configurations output by S3, the monitoring parameter time-series dataset is first constructed. For each scene category, monitoring parameters and scene status data are collected continuously for 24 hours, with a time step of 10 seconds. Each time-series sample contains a correspondence between "scene features of the previous 20 time steps - monitoring parameters of the current time step," where the scene features are the 12-dimensional key features selected by S31, and the monitoring parameters are 5-dimensional control targets. .

[0078] Considering the differences in magnitude between parameters of different dimensions, an improved normalization method is adopted to eliminate the influence of dimensions, as shown in the following formula: ,in, These are the original parameter values. The first The minimum and maximum values ​​of class parameters in the corresponding scenarios are mapped to the range [0.1, 0.9] by normalizing the data, so as to avoid the interference of extreme values ​​on model training, while preserving the trend of parameter changes.

[0079] In the S42 section, the Whale Optimization Algorithm (WOA) optimizes the hyperparameters of the LSTM network. Using the time-series data preprocessed in S401 as input, an LSTM network is constructed as the basic prediction model. The network structure includes: an input layer with 12-dimensional scene features and 20 time steps, totaling 240 neurons; an initial 3-layer hidden layer; and an output layer with 5-dimensional monitoring parameters. The activation function is Sigmoid to match the normalized data range. The Whale Optimization Algorithm (WOA) is introduced for hyperparameter optimization.

[0080] The objective function for WOA is set as minimizing the root mean square error (RMSE) of the predicted monitoring parameters on the validation set, as shown in the following formula: ,in, For the hyperparameters to be optimized, To determine the number of samples in the validation set, To monitor parameter dimensions, The first The first sample The true and predicted values ​​of the dimension parameters.

[0081] In the S43 and WLSTM model training and dynamic correction of monitoring parameters, the hyperparameters optimized by WOA are substituted into the LSTM network to construct a WLSTM monitoring parameter prediction model. The training process uses the Adam optimizer, and the loss function is the mean squared error (MSE), as shown in the formula: .

[0082] Meanwhile, a Dropout layer is introduced to prevent overfitting, and the number of training iterations is set to 1000. An early stopping mechanism is triggered when the validation set loss does not decrease for 10 consecutive iterations.

[0083] In a preferred embodiment of the present invention, in step S5, video coding and transmission optimization based on Dynamic Hierarchical Segmentation Network (DHC-Net), the following steps are performed: S51, dynamic evaluation and hierarchical classification of video coding complexity; S52, dynamic weight segmentation and coding acceleration using DHC-Net; and finally, S53, collaborative optimization of coding rate and transmission protocol.

[0084] In S51, the dynamic evaluation and grading of video coding complexity is based on the dynamically corrected monitoring parameters output by S4. First, a video coding complexity evaluation model is constructed to grade the complexity of real-time video streams. Then, video frame data corresponding to the corrected parameters are collected, and intra-frame texture complexity and inter-frame motion complexity are extracted to construct a 4-dimensional complexity feature vector.

[0085] An improved method combining structural similarity and color distribution differences is used to calculate the overall complexity score of video frames. The formula is: ,in, adjacent frames and Structural similarity, The Manhattan distance is the RGB three-channel histogram of adjacent frames. These are weighting coefficients. Combined with the bitrate output by S4. With bandwidth The matching relationship divides the video encoding complexity into three levels: high, medium, and low, providing a basis for subsequent dynamic encoding configuration.

[0086] In the S52 and DHC-Net dynamic weight segmentation and encoding acceleration, a dynamic hierarchical segmentation network is constructed based on the complexity ranking results of S51 to achieve adaptive adjustment of encoding weights. The encoding layer of DHC-Net adopts a 3D separable convolutional structure, decomposing the standard 3D convolution into depthwise convolution and pointwise convolution, as shown in the formula:

[0087]

[0088] in, These are the spatial convolution kernel size and the temporal convolution kernel size, respectively. These represent the number of input and output channels, and this decomposition reduces the computational cost compared to standard 3D convolution. This improves coding efficiency.

[0089] A dynamic weighting strategy is adopted for different complexity levels: 100% channel weight is used for high complexity scenarios, 75% channel weight is used for medium complexity scenarios, and 50% channel weight is used for low complexity scenarios. The weight selection formula is as follows: ,in, For all convolutional kernel weights, The weighting ratio is preserved. Simultaneously, switchable batch normalization (S-BN) is introduced to adapt to the statistical characteristics of different channel weights, as shown in the following formula. ,in, The channel mean and variance are for the corresponding complexity level. These are learnable parameters to ensure the stability of encoded features under different weighting modes.

[0090] In the S53 co-optimization of encoding rate and transmission protocol, a co-optimization mechanism for encoding rate and transmission protocol is constructed based on the encoded data output from S52 and the network bandwidth B updated in real time by S4. First, an adaptive rate control algorithm is used to adjust the encoding rate. Referring to the network transmission and rate matching logic, a rate adjustment model is established, as shown in the following formula: ,in, These are the current bitrate and bandwidth, respectively. This is the difference between the current transmission delay and the reference delay. For the time delay threshold, To adjust the coefficients and achieve real-time tracking of bitrate fluctuations in response to bandwidth variations.

[0091] Dynamically select the transmission protocol based on different complexity levels and bitrate ranges: for high-complexity scenarios, adopt... The protocol ensures real-time performance through a timestamp mechanism and employs a selective retransmission strategy for packet loss retransmission; in medium-complexity scenarios, the TCP protocol is used, with a sliding window mechanism to ensure reliability, and the window size is set to a specific value; in low-complexity scenarios, a different protocol is used. The protocol reduces transmission overhead through multiplexing.

[0092] Meanwhile, a transmission quality feedback mechanism is introduced to calculate the packet loss rate of video transmission every 30 seconds. latency jitter ,when or When the time is right, the weight splitting ratio of DHC-Net in S52 is adjusted, and the feedback is sent to S4 to adjust the monitoring parameters, forming a closed-loop optimization of "encoding-transmission-parameter correction".

[0093] In a preferred embodiment of the present invention, in step S6, system iterative update and feedback optimization, S61, multi-dimensional performance evaluation and anomaly diagnosis are performed first, followed by S62, adaptive iterative update of the model and parameters.

[0094] In S61, the multi-dimensional performance evaluation and anomaly diagnosis, a multi-dimensional performance evaluation model is constructed based on the video encoding quality data, transmission performance data, and monitoring parameter configuration data output by S5, to achieve quantitative diagnosis of the system's operating status. First, the packet loss rate from the transmission quality feedback in S5 is integrated. latency jitter Together with video coding quality indicators and monitoring parameter deviations, a 6-dimensional performance evaluation vector is formed.

[0095] Construct the weighted performance evaluation function E, as shown in the following formula:

[0096] E=W'M.

[0097] in, Videos The theoretical maximum and minimum values, For latency jitter threshold, To monitor the maximum permissible deviation of the parameters, This is the threshold for encoding time. These are the weighting coefficients. , A smaller value indicates better system performance.

[0098] when When a system is found to have performance anomalies, the source of the anomaly is further located through contribution analysis. The contribution calculation formula is as follows: Among them, indicators For the first The normalized value of the performance index, through The magnitude of the anomaly determines the dominant anomaly factor, providing a target direction for subsequent iterative updates.

[0099] In the adaptive iterative update of the model and parameters described in S62, the system iterative update mechanism is initiated for the anomaly source located in S601 to achieve dynamic optimization of model parameters and monitoring strategies. If the anomaly source is packet loss or latency jitter, it is fed back to the transmission protocol and rate control module in S503 to adjust the adjustment coefficients in the ABR algorithm. The updated formula is as follows:

[0100]

[0101] in, These are the target packet loss rate and the target latency jitter, respectively. To update the step size, ensure It adaptively adjusts within the range of [0.1, 0.5] to optimize the matching accuracy between bit rate and bandwidth.

[0102] If the anomaly source is a deviation in monitoring parameters or insufficient encoding quality, then incremental training of the WOA-LSTM model in S402 and updates to the DHC-Net weight splitting strategy in S502 are triggered. During incremental training, monitoring data from the period of performance anomalies in S601 are selected, fused with the historical training set, and the WLSTM model is retrained, the model weights are updated, and the optimization objective function of WOA is adjusted, introducing a constraint term in the performance evaluation function E, as shown in the following formula: ,in, These are constraint coefficients to ensure the newly optimized hyperparameters. This not only reduces parameter prediction errors but also improves overall system performance. For DHCNet, the weight retention ratio is adjusted based on anomaly diagnosis results. In high-complexity scenarios, such as when the encoding quality is insufficient. From version 1.0 to 1.05, when packet loss is severe, the performance rating drops from 0.75 and 0.5 to 0.7 and 0.45 in low-to-medium complexity scenarios.

[0103] A full iteration update is performed every 24 hours, fusing the daily performance data with historical data, and retraining the RF classification model of S3, the WLSTM model of S4, and the DHC-Net of S5 to ensure that the models adapt to changes in the scenario.

[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive adjustment method for IoT video surveillance scenarios combining big data mining and deep learning, characterized in that, Includes the following steps: Step S1: Multi-source monitoring data acquisition and preprocessing. Collect video streams, PTZ attitude, environmental perception and network transmission data in IoT video surveillance scenarios. Through "device identifier - timestamp" association and fusion, and after improved low-pass filtering cleaning, coordinate transformation and feature enhancement, preprocessed data is formed. Step S2: Scene clustering analysis based on the improved D-CURE algorithm. Redundant features are removed from the preprocessed data in Step S1 using the Pearson coefficient. A weighted multimodal similarity distance is constructed. Abnormal scenes are removed by DBSCAN pre-clustering, CURE clustering optimization and Davidson-Dinburg index are used to determine the optimal number of clusters, and typical monitoring scenes are output. Step S3: Scene classification prediction based on random forest. Using the clustering results of step S2 as labels, construct a feature set and filter key features through out-of-bag error. Train the RF model of CART base learner, introduce similar sample weighting to correct and optimize the prediction results, and output scene classification and initial monitoring parameters. Step S4: Dynamic correction of monitoring parameters based on WLSTM. Based on the results of step S3, a time series dataset of monitoring parameters is constructed. After improved normalization preprocessing, the LSTM hyperparameters are optimized using the whale optimization algorithm. The WLSTM model is trained and parameter optimization is achieved through rolling prediction + dynamic correction. Step S5: Optimize video coding and transmission based on dynamic hierarchical segmentation network. Based on the parameter correction in step S4, evaluate the video coding complexity and classify it. Accelerate coding through DHC-Net dynamic weight segmentation and achieve collaborative optimization by combining adaptive bitrate control and dynamic transmission protocol. Step S6: System Iteration Update and Feedback Optimization. Based on the results of step S5, a multi-dimensional performance evaluation model is constructed to locate the source of anomalies. Transmission parameters, incremental training models, and coding strategies are updated in a targeted manner to achieve continuous optimization of system performance.

2. The IoT video surveillance scene adaptive adjustment method according to claim 1, wherein step S1 includes: The collected data covers video streams with resolutions ranging from 1080P to 4K and frame rates from 25 to 60fps, as well as the horizontal rotation angle of the pan-tilt unit. Vertical corner ,illumination ,humidity Wind speed ,visibility ,bandwidth Delay Packet loss rate A 28-dimensional initial matrix is ​​formed by associating "device identifier - timestamp"; Data cleaning employs an improved low-pass filter: , (Video grayscale) (Gimbal angle) (Ambient lighting) (Network bandwidth); Coordinate transformation: in, Focal length For the image width and height, For the field of view; feature enhancement adds the mean optical flow. Motion vector Improve SSIM: .

3. The IoT video surveillance scene adaptive adjustment method according to claim 1, wherein step S2 includes: Feature Filtering Using Pearson coefficients: , Redundancy was eliminated, and 15 core features were retained. Similarity measurement: , (Video animation) (environment), (Gimbal) (Network); DBSCAN pre-clustering: , Core sample determination ; represents point contraction Inter-cluster distance The optimal number of clusters is determined using DBI: , , .

4. The IoT video surveillance scene adaptive adjustment method according to claim 1, wherein step S3 includes: Construct an 18-dimensional initial feature set and filter key features using out-of-bag error: ( ), retaining 12-dimensional features; the RF model uses CART as the base learner, and the Gini index is: Hyperparameters: , , Bootstrap sampling and random feature selection are employed; prediction correction: ( ).

5. The IoT video surveillance scene adaptive adjustment method according to claim 1, wherein step S4 includes: Construct a time-series dataset with a time step of 10 seconds and monitor the parameters. Improved normalization: The WLSTM model has 240 neurons in its input layer, and hyperparameters are optimized using WOA. Objective function: .

6. The IoT video surveillance scene adaptive adjustment method according to claim 1, wherein step S5 includes: Coding complexity evaluation: , divided into three levels: high ( ), medium ( ), and low ( ); DHC-Net uses 3D separable convolution: depth convolution , and the formula for pointwise convolution is: , and the formula for dynamic weight screening is , high complexity , medium complexity , low complexity ; Collaborative optimization uses the ABR algorithm: , high complexity uses , medium uses , low uses , packet loss rate , delay jitter .

Images