Intelligent detection method and system for library intrusion based on deep learning

Abstract

The invention relates to the technical field of intelligent security and protection monitoring, and discloses a library intrusion intelligent detection method and system based on deep learning, and the method comprises the steps: collecting library video data, carrying out the preprocessing, extracting spatial-temporal features, and carrying out the recognition of the spatial-temporal features; a deep convolutional neural network and a long-short-term memory network are combined to construct a double-flow fusion model, action recognition and abnormal behavior detection are realized, an intrusion event is determined through a multi-level threshold judgment mechanism, and real-time alarm is realized. According to the method, the accuracy and the real-time performance of warehouse intrusion detection are improved, false alarms and missing alarms are effectively reduced, and meanwhile, the method has autonomous learning and environment adaptability.

Description

Technical Field

[0001] The present invention relates to the field of intelligent security monitoring technology, and more specifically, to a deep learning-based intelligent detection method and system for warehouse intrusion. Background Art

[0002] Storerooms and warehouses often store large quantities of materials and assets, making security crucial. Traditional storeroom security systems rely primarily on hardware devices like infrared detectors and magnetic switches for simple intrusion detection, or employ video surveillance systems based on traditional computer vision algorithms like background subtraction and optical flow. These methods have numerous limitations in practical applications.

[0003] Traditional security systems commonly suffer from the following problems: First, hardware-sensor-based alarm systems are susceptible to interference from environmental factors, such as temperature fluctuations and airflow fluctuations, which can lead to false alarms. Second, simple video surveillance systems typically use fixed threshold judgment mechanisms, which cannot adapt to complex situations such as changes in ambient lighting and weather conditions, resulting in a high false alarm rate. Third, traditional video analysis algorithms have difficulty distinguishing between normal and abnormal behavior and cannot identify complex intrusion scenarios, such as slow movement and disguised intrusions. Furthermore, existing systems are mostly passive monitoring and lack intelligent early warning capabilities. Alarms are often triggered only after an intrusion has already occurred, resulting in insufficient proactive security measures. Finally, traditional systems typically operate in an independent mode, with limited information sharing between subsystems, making it difficult to form a collaborative judgment mechanism, which reduces overall protection capabilities.

[0004] The development of deep learning technology has led to significant breakthroughs in computer vision, providing new technical solutions for addressing these issues. However, current approaches to applying deep learning to warehouse intrusion detection still suffer from issues such as insufficient real-time performance, weak model generalization, and high false alarm rates, failing to meet practical application requirements. Therefore, there is an urgent need for a deep learning-based intelligent warehouse intrusion detection method and system that can accurately identify complex intrusion behaviors, reduce false alarm rates, and improve system real-time performance and reliability, thereby comprehensively enhancing warehouse security. Summary of the Invention

[0005] The present invention provides a deep learning-based intelligent detection method and system for warehouse intrusion, which solves the technical problems of low warehouse intrusion detection accuracy, high false alarm rate, poor real-time performance and weak environmental adaptability in related technologies.

[0006] The present invention provides a deep learning-based intelligent detection method for storage room intrusion, comprising the following steps:

[0007] Acquire warehouse monitoring information through multi-source sensing equipment and pre-process the collected video data. The warehouse monitoring information includes video data and heat data;

[0008] Extract spatiotemporal features from preprocessed video data, including spatial features based on deep convolutional neural networks and temporal features based on optical flow methods;

[0009] Build a two-stream fusion model based on deep convolutional neural networks and long short-term memory networks to integrate spatial and temporal features;

[0010] Analyze the extracted spatiotemporal features based on deep learning models to identify intrusion behaviors in the warehouse, including behavior classification and anomaly score calculation;

[0011] Based on a multi-level threshold judgment mechanism, different levels of warnings are triggered, and trajectory analysis technology is combined to enhance decision-making accuracy;

[0012] The detection parameters are adaptively adjusted according to environmental changes, and abnormal behavior detection and early warning can be stably achieved under different conditions.

[0013] In a preferred embodiment, in the step of obtaining storage room monitoring information through multi-source sensing equipment, multiple cameras and auxiliary infrared sensors are deployed in key areas of the storage room to collect video and heat information in real time;

[0014] Among them, cameras with full HD resolution or higher resolution are used, with a frame rate of no less than 25 frames per second, and infrared sensors are installed at key locations of the warehouse entrance and windows to detect changes in human body heat and provide auxiliary judgment information for video analysis.

[0015] In a preferred embodiment, the step of preprocessing the collected video data includes:

[0016] Video decoding and frame extraction, image denoising and enhancement, illumination normalization, and moving target detection; illumination normalization uses an adaptive illumination normalization algorithm, specifically:

[0017]

[0018] Among them, I(x,y) represents the pixel value of the original image at the coordinate (x,y), I norm (x, y) represents the normalized pixel value, μ represents the average brightness value of the original image, σ represents the standard deviation of the original image, and μ target and σ target are the target mean brightness and standard deviation, respectively.

[0019] In a preferred embodiment, moving target detection uses an improved mixed Gaussian background modeling algorithm and applies morphological operations, including dilation and erosion, to the detected moving targets to eliminate noise and fill holes; the dilation operation is defined as:

[0020]

[0021] The erosion operation is defined as:

[0022]

[0023] Among them, A represents the binary image, representing the foreground mask; B represents the structural element, defining the shape and size of the morphological operation; B z Represents the result of translating the structural element B to point z; represents the expansion operator; Represents the erosion operator.

[0024] In a preferred embodiment, in the step of extracting spatiotemporal features from the preprocessed video data:

[0025] Spatial feature extraction uses a pre-trained ResNet-50 network, removing the last fully connected layer and retaining the feature maps extracted by the convolutional layer;

[0026] For each input video frame, the size is adjusted to 224×224×3, and a 2048-dimensional feature vector is obtained;

[0027] Temporal feature extraction uses the optical flow method to calculate the motion vector between multiple consecutive frames of target area images, and uses the Farneback algorithm to calculate the dense optical flow field. The horizontal and vertical optical flow fields are calculated for each pair of consecutive frames, and the optical flow fields of 10 consecutive frames are stacked to form an optical flow volume with a size of 224×224×20.

[0028] In a preferred embodiment, the steps of constructing a dual-stream fusion model include:

[0029] Construct a spatial stream network, receive the processed spatial feature vector, and output a 512-dimensional spatial stream feature vector through a fully connected layer and activation function;

[0030] Construct a temporal flow network, use a 3D convolutional neural network and a long short-term memory network to process the temporal information in the optical flow volume, and output a 128-dimensional temporal flow feature vector;

[0031] A feature fusion network is constructed to concatenate 512-dimensional spatial features and 128-dimensional temporal features into a 640-dimensional vector. Through the self-attention layer and the fully connected layer, a 256-dimensional fused feature vector and the probability distribution of behavior categories are output.

[0032] In a preferred embodiment, in the step of identifying intrusion behaviors in the storage room, behaviors in the storage room are classified into the following categories: normal walking, normal operation, wandering behavior, climbing behavior, abnormal operation, intrusion behavior, object handling, and other abnormal behaviors; and anomaly scores are calculated based on the probability distribution of the model output. The comprehensive anomaly score calculation formula is:

[0033] Sanomaly = α·S category + β·S duration + γ·S region + δ·S time ;

[0034] where α, β, γ, and δ are the weight coefficients of behavior category, behavior duration, occurrence area, and time sensitivity respectively, satisfying α + β + γ + δ = 1; S anomaly is the finally calculated comprehensive anomaly score.

[0035] In a preferred embodiment, the steps of triggering different levels of warnings based on a multi-level threshold determination mechanism are specifically as follows:

[0036] When the anomaly score S anomaly < T1, it is at the normal level, and the behavior is recorded in the system log but no alarm is triggered;

[0037] When T1 ≤ S anomaly < T2, it is at the attention level, and the system marks it as an event to be concerned about;

[0038] When T2 ≤ S anomaly < T3, it is at the warning level, and a warning notice is sent to the security personnel;

[0039] When S anomaly ≥ T3, it is at the emergency level, and a high-level alarm is triggered;

[0040] where T1, T2, and T3 are the first, second, and third thresholds of the anomaly score respectively.

[0041] In a preferred embodiment, the steps of enhancing decision accuracy by combining trajectory analysis technology include:

[0042] Using a Kalman filter to track the detected moving target;[[ID=四十七]] [[ID=四十八]]

[0043] [[ID=四十九]]Calculating the average speed, trajectory curvature, and dwell time characteristics of the trajectory;[[ID=五十]] [[ID=五十一]]

[0044] [[ID=五十二]]Calculating the trajectory anomaly score S[[ID=五十三]] trajectory [[ID=五十四]];[[ID=五十五]] [[ID=五十六]]

[0045] [[ID=五十七]]Combining the trajectory anomaly score with the behavior anomaly score to form the final comprehensive anomaly score:[[ID=五十八]] [[ID=五十九]]

[0046] [[ID=六十]]S[[ID=六十一]] final [[ID=六十二]]= λ1S[[ID=六十三]] anomaly [[ID=六十四]]+ (1 - λ1)S[[ID=六十五]] trajectory [[ID=六十六]];[[ID=六十七]] [[ID=六十八]]

[0047] [[ID=六十九]]where λ1 is the trade-off coefficient, S[[ID=七十]] anomaly [[ID=七十一]]is the finally calculated comprehensive anomaly score, S[[ID=七十二]] trajectoryis the calculated trajectory anomaly score, S final is the final calculated comprehensive anomaly score.

[0048] In a preferred embodiment, a deep learning-based intelligent detection system for warehouse intrusion is used to perform the steps in the above-mentioned deep learning-based intelligent detection method for warehouse intrusion, including:

[0049] The multi-source sensing acquisition module is used to obtain warehouse monitoring information, including high-definition cameras and infrared sensors. The multi-source sensing acquisition module is responsible for deploying multiple full-HD cameras and infrared sensors in key areas of the warehouse to collect video streams and thermal data in real time.

[0050] The video preprocessing module is used to decode, denoise, enhance, normalize illumination, and detect moving targets on the collected video data. The video preprocessing module receives the original video stream and decodes it, applies a Gaussian filter algorithm to remove noise, implements adaptive illumination normalization, and performs moving target detection using an improved mixed Gaussian background modeling algorithm.

[0051] The feature extraction module is used to extract spatiotemporal features from preprocessed video data. The feature extraction module uses a pre-trained ResNet-50 network to extract spatial features and generate a 2048-dimensional feature vector. At the same time, it uses the Farneback algorithm to calculate the dense optical flow field to extract temporal features and form an optical flow volume, providing a complete spatiotemporal feature representation for subsequent deep learning models.

[0052] The deep learning model module is used to build a dual-stream fusion model and perform behavior recognition and anomaly score calculation. The deep learning model module consists of three sub-modules: spatial stream network, temporal stream network, and feature fusion network. It fuses spatial and temporal features through a self-attention mechanism.

[0053] The early warning decision module is used to trigger different levels of early warnings based on a multi-level threshold judgment mechanism. Based on the calculated anomaly score and combined with trajectory analysis technology, the early warning decision module classifies abnormal events into four levels: normal, caution, warning, and emergency through a multi-level threshold judgment mechanism, and triggers corresponding early warning measures to ensure the timeliness and accuracy of safety responses.

[0054] The environmental adaptation module is used to automatically adjust detection parameters according to environmental changes; the environmental adaptation module continuously monitors the environmental conditions of the storage room and automatically adjusts the detection threshold, feature weights and model parameters through a feedback learning mechanism to ensure that the system can maintain stable detection performance under various environmental conditions and improve the robustness and adaptability of the system.

[0055] The beneficial effects of the present invention are:

[0056] Deep learning technology improves the intelligence and accuracy of warehouse intrusion detection, reduces false alarm and missed alarm rates, and enhances the system's environmental adaptability. The system uses a deep learning dual-stream fusion model combined with spatiotemporal feature analysis to accurately identify intrusions in complex scenarios and is highly robust to lighting changes, occlusion, slow movement, and other issues.

[0057] This system enables early identification and early warning of abnormal behavior, significantly improving the proactiveness and foresight of security systems. A multi-level alert mechanism reduces unnecessary human intervention. The system possesses autonomous learning capabilities, continuously optimizing model performance based on new data and improving environmental adaptability. Furthermore, the invention incorporates a lightweight network structure and optimization algorithm, enabling real-time processing capabilities for edge devices, reducing system deployment and maintenance costs. It can also be seamlessly integrated into existing security systems, offering strong scalability and a wide range of application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a flow chart of a deep learning-based intelligent detection method for warehouse intrusion of the present invention. DETAILED DESCRIPTION

[0059] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0060] At least one embodiment of the present invention discloses a deep learning-based intelligent detection method for warehouse intrusion, such as Figure 1 As shown, the following steps are included:

[0061] S100, acquiring warehouse monitoring information through multi-source sensing equipment and preprocessing the acquired video data, the warehouse monitoring information including video data and heat data;

[0062] In one embodiment of the present invention, multiple cameras and auxiliary infrared sensors are deployed in key areas of the storage room to collect video and heat information in real time. For example, infrared sensors are installed at key locations such as the storage room entrance and windows to detect changes in human body heat and provide auxiliary judgment information for video analysis. The details are as follows:

[0063] Real-time collection of high-definition video data within the warehouse; using cameras with full HD resolution (1920×1080 pixels) or higher, with a frame rate of no less than 25 frames per second, covering the main areas of the warehouse to form a monitoring network with no blind spots;

[0064] The camera's installation position and angle have been optimized to maximize the monitoring field of view while reducing interference factors such as light reflection and occlusion. To enhance the system's performance in low-light conditions, some cameras are equipped with infrared night vision function;

[0065] Infrared sensors are used to assist in detecting changes in human body heat. These sensors are installed at key locations, such as entrances and windows, to detect changes in human body heat and provide information for video analysis. The collected raw data is transmitted in real time to the edge computing layer for subsequent processing. Output: A thermal data stream containing temperature changes and thermal imaging data within the detection area.

[0066] In one embodiment of the present invention, the collected video data is preprocessed, including video decoding, denoising and enhancement, illumination normalization, and moving target detection. The data preprocessing module receives the raw video data transmitted by the video capture device output in step 1 and performs a series of preprocessing operations on it, mainly including the following steps:

[0067] Video decoding and frame extraction: decode the received video stream and convert the compressed video into a processable raw frame sequence;

[0068] Extract key frames from the original video stream according to the sampling frequency set by the system (the default is 10 frames / second) to reduce the computational complexity of subsequent processing;

[0069] Output: decoded video frame sequence;

[0070] Image denoising and enhancement: Apply a Gaussian filter to the extracted video frames for denoising, with a filter kernel size of 5×5 and a standard deviation of 1.5;

[0071] The processing steps of the Gaussian filter are as follows:

[0072] Construct a 5×5 Gaussian kernel matrix with the largest value at the center and decreasing towards the surrounding areas;

[0073] Apply a Gaussian kernel to each pixel of the image and calculate the weighted average;

[0074] Replace the original pixel value with the weighted average value to achieve noise smoothing;

[0075] Perform histogram equalization on the denoised image to enhance image contrast and make image details clearer;

[0076] The processing steps of histogram equalization are as follows:

[0077] Calculate the image grayscale histogram and count the number of pixels at each grayscale level;

[0078] Calculate the cumulative distribution function and map the original grayscale value to the new grayscale space;

[0079] Redistribute pixel values ​​according to the mapping relationship and stretch the contrast;

[0080] Output: denoised and enhanced video frame sequence;

[0081] Lighting normalization: In order to eliminate the impact of lighting changes in different time periods, an adaptive lighting normalization algorithm is applied to the denoised and enhanced video frame sequence;

[0082] Calculate the average brightness value μ and standard deviation σ of the image, and then standardize the image:

[0083]

[0084] Among them, I(x,y) represents the pixel value of the original image at the coordinate (x,y), I norm (x, y) represents the normalized pixel value, μ represents the average brightness value of the original image, σ represents the standard deviation of the original image, and μ target and σ target are the target mean brightness and standard deviation, with default values ​​set to 128 and 64 respectively;

[0085] Output: video frame sequence after illumination normalization;

[0086] In one embodiment of the present invention, an improved Gaussian mixture modeling algorithm (GMM) is used to perform moving target detection on a video frame sequence after illumination normalization;

[0087] The mixed Gaussian background modeling algorithm is a background modeling method based on statistical learning. Its core idea is to use multiple Gaussian distributions to describe the possible state distribution of each pixel in the image. For each pixel in the video sequence, K Gaussian distributions (default K = 5) are used to represent its possible states, and each Gaussian distribution represents a background or foreground state.

[0088] The processing steps of the mixed Gaussian background modeling algorithm are as follows:

[0089] Initialization: establish K Gaussian distribution models for each pixel and set the initial mean, variance and weight;

[0090] Matching judgment: for the current pixel value, check whether it matches the existing Gaussian distribution;

[0091] Parameter update: update the weight, mean and variance of the Gaussian distribution according to the matching results;

[0092] Background judgment: sort the Gaussian distribution according to the ratio of weight to variance, and select the first few distributions whose cumulative weight exceeds the threshold as the background model;

[0093] Foreground detection: if the current pixel value does not match any background distribution, it is determined to be a foreground pixel;

[0094] Model maintenance: when no distribution matches the current pixel value, replace it with the distribution with the smallest weight;

[0095] Morphological operations, including dilation and erosion, are applied to the detected moving objects to remove noise and fill holes. The dilation operation is defined as:

[0096]

[0097] The erosion operation is defined as:

[0098]

[0099] Among them, A represents a binary image, representing the foreground mask; B represents a structural element, which defines the shape and size of the morphological operation. By default, a 3×3 rectangular structural element is used; B z Represents the result of translating the structural element B to point z; represents the expansion operator; represents the corrosion operator;

[0100] Output: Moving object detection results, including foreground mask (binary image) and a list of detected object bounding box coordinates;

[0101] S200, extracting spatiotemporal features from the preprocessed video data, including spatial features based on a deep convolutional neural network and temporal features based on an optical flow method;

[0102] This step receives the output pre-processed video data, extracts spatial features and temporal features respectively, and provides input for the deep learning model;

[0103] In one embodiment of the present invention, feature extraction is mainly divided into the following steps:

[0104] S201, spatial feature extraction; spatial feature extraction mainly uses deep convolutional neural network to extract spatial information from moving target detection results;

[0105] A pre-trained ResNet-50 network is used, but the last fully connected layer is removed, retaining the feature maps extracted by the convolutional layer;

[0106] The residual block structure of ResNet-50 effectively solves the gradient vanishing problem during deep network training. Its residual learning formula is:

[0107] H(x)=F(x)+x;

[0108] Here, H(x) is the desired underlying mapping, F(x) is the residual mapping, and x is the input feature.

[0109] Specifically, for each input video frame (resized to 224×224×3), a 2048-dimensional feature vector is obtained after passing through the ResNet-50 network, representing the spatial features of the frame;

[0110] The feature extraction process includes multi-layer convolution, batch normalization, activation function and pooling operations;

[0111] To improve computational efficiency, spatial features are extracted every 5 frames;

[0112] Output: 2048-dimensional spatial feature vector, representing the appearance characteristics of the target;

[0113] S202, temporal feature extraction; temporal feature extraction mainly captures motion information in the video sequence;

[0114] In one embodiment of the present invention, the temporal feature extraction uses an optical flow method to calculate the motion vector between multiple consecutive frames of detected target area images;

[0115] Optical flow refers to the visual apparent motion pattern of pixels in an image, reflecting the relative motion of objects, surfaces, and edges in a scene. The Farneback algorithm is used to calculate dense optical flow fields. This algorithm approximates the signal based on polynomial expansion and calculates the pixel-level optical flow between two frames.

[0116] The processing steps of the Farneback algorithm are as follows:

[0117] Polynomial approximation, which uses a polynomial function to approximate the local neighborhood of each pixel position between two consecutive frames;

[0118] Displacement estimation, assuming that there is a displacement relationship between the two frames, calculate the optimal displacement vector through the polynomial coefficients;

[0119] Multi-scale processing, using a pyramid strategy to first estimate large displacements at low resolution and then refine them at high resolution;

[0120] Iterative optimization, through multiple iterations of refinement, improves the accuracy of optical flow estimation;

[0121] Output: Generate horizontal and vertical optical flow fields;

[0122] In actual implementation, the system calculates the horizontal and vertical optical flow fields for each pair of consecutive frames to generate an optical flow image. The optical flow fields of 10 consecutive frames are stacked to form an optical flow volume as a representation of temporal features. The size of the optical flow volume is 224×224×20 (10 frames × 2 channels / frame), which captures the motion trend in short video clips.

[0123] Output: Optical flow volume with size of 224×224×20, representing the motion characteristics of the target;

[0124] S203, feature fusion preprocessing: preliminarily processing the extracted spatial features and temporal features to meet the input requirements of the subsequent deep learning model;

[0125] Apply global average pooling to the spatial features to compress the feature map into a fixed-dimensional vector;

[0126] Apply 3D convolutional layers to the temporal features to extract spatiotemporal correlations; the preliminarily processed features will serve as input to the two-stream fusion model;

[0127] Output: processed spatial feature vector (2048 dimensions) and temporal feature tensor (feature map after 3D convolution).

[0128] S300, builds a two-stream fusion model based on deep convolutional neural networks and long short-term memory networks, integrating spatial and temporal features;

[0129] In one embodiment of the present invention, a deep learning-based dual-stream fusion model is constructed for identifying intrusion behaviors in storage rooms. The model mainly includes three parts: a spatial stream network, a temporal stream network, and a feature fusion network. The details are as follows:

[0130] S301, spatial stream network construction; the spatial stream network receives the processed spatial feature vector and maps it through a series of fully connected layers and activation functions; the specific structure is:

[0131] Fully connected layer 1: input dimension 2048, output dimension 1024, using ReLU activation function;

[0132] Dropout layer: dropout rate 0.5 to prevent overfitting;

[0133] Fully connected layer 2: input dimension 1024, output dimension 512, using ReLU activation function;

[0134] Batch normalization layer: stabilizes the training process;

[0135] The output of the spatial stream network is a 512-dimensional feature vector, which represents the static visual information in the video frame;

[0136] Output: 512-dimensional spatial stream feature vector;

[0137] S302, Temporal Stream Network Construction; The Temporal Stream Network uses a combination of 3D convolutional neural networks and long short-term memory networks (LSTMs) to process temporal information in the optical flow volume. 3D convolutional neural networks can capture features in both spatial and short-term ranges, while LSTMs specialize in processing temporal dependencies in long sequences. The specific structure is as follows:

[0138] 3D convolutional layer 1: 64 3×3×3 convolution kernels with a stride of (1, 1, 1) and a ReLU activation function.

[0139] 3D max pooling layer 1: pooling size 2×2×2, stride 2;

[0140] 3D convolutional layer 2: 128 3×3×3 convolution kernels with a stride of (1, 1, 1) and ReLU activation function;

[0141] 3D max pooling layer 2: pooling size 2×2×2, stride 2;

[0142] Flattening layer: converts 3D feature maps into 1D vectors;

[0143] LSTM layer: 128 hidden units, processing temporal relationships;

[0144] In one embodiment of the present invention, the LSTM layer receives the flattened 3D feature vector sequence and outputs a 128-dimensional feature vector representing the dynamic motion information in the video sequence. This vector encodes the temporal variation pattern in the video sequence, which is crucial for identifying different types of intrusion behaviors.

[0145] Output: 128-dimensional time stream feature vector.

[0146] S303, constructing a feature fusion network; the feature fusion network fuses the output features of the output spatial stream and the output temporal stream to form a unified representation;

[0147] Use the attention mechanism to perform feature fusion and enhance the weight of important features; the specific fusion method is:

[0148] Feature concatenation: concatenate 512-dimensional spatial features and 128-dimensional temporal features into a 640-dimensional vector;

[0149] Self-attention layer: calculates the attention weight of the features and enhances the representation of important features;

[0150] Fully connected layer: input dimension 640, output dimension 256, using ReLU activation function;

[0151] Output layer: Input dimension is 256, output dimension is equal to the number of behavior categories, and Softmax activation function is used;

[0152] Output: 256-dimensional fused feature vector and behavior category probability distribution (the dimension is equal to the number of behavior categories);

[0153] S304, model training; model training adopts a phased strategy:

[0154] First, the spatial stream network and the temporal stream network are pre-trained independently, and then the entire fusion network is fine-tuned end-to-end.

[0155] The loss function uses the cross entropy loss function to measure the difference between the predicted category and the true category:

[0156]

[0157] Where N is the number of samples, C is the number of categories, and y i,j is the true label of sample i belonging to category j (value is 0 or 1), p i,j is the probability that the model predicts that sample i belongs to category j, and L is the calculated cross entropy loss value;

[0158] Optimizer: Use Adam optimizer, set the initial learning rate to 0.001, and the weight decay coefficient to 0.0001;

[0159] Learning rate scheduling: A cosine annealing learning rate strategy is used, and the learning rate gradually decreases as training progresses;

[0160] Data enhancement: Apply multi-source data enhancement techniques, including random cropping, horizontal flipping, rotation, and brightness adjustment;

[0161] Training batch size: set to 32, and the number of training rounds is 100;

[0162] Early stopping strategy: Stop training when the validation set performance does not improve for 5 consecutive rounds to prevent overfitting;

[0163] Model saving: save the model parameters with the best performance on the validation set;

[0164] Output: trained deep learning model (dual-stream fusion model) and model parameters.

[0165] S400 analyzes the extracted spatiotemporal features based on a deep learning model to identify intrusion behaviors in the storage room, including behavior classification and anomaly score calculation;

[0166] The behavior recognition component analyzes video data based on a trained deep learning model (i.e., a dual-stream fusion model) to identify potential intrusion behaviors. The specific steps are as follows:

[0167] S401, behavior classification; the deep learning model performs real-time analysis on spatial feature vectors and optical flow volumes, and outputs a probability distribution of behavior categories;

[0168] In one embodiment of the present invention, behaviors in the warehouse are divided into the following categories: normal walking; normal operation; wandering behavior; climbing behavior; abnormal operation; intrusion behavior; object handling; other abnormal behaviors (including crouching and hiding, standing still for a long time, repeatedly trying to open the door, blocking the camera, using tools to pry open doors and windows, staying in non-working areas, carrying suspicious items, multiple people collaborating in suspicious activities, etc.);

[0169] The model outputs the probability value of each category, indicating the possibility that the current video clip belongs to each behavior category;

[0170] Output: Behavior category probability distribution, dimension is 8 (corresponding to 8 behavior categories);

[0171] S402, Abnormal Behavior Scoring: Calculate the anomaly score (AnomalyScore) based on the output probability distribution. The calculation of the anomaly score comprehensively considers the following factors:

[0172] Behavior category probability. The higher the probability value of the abnormal behavior category, the higher the anomaly score:

[0173]

[0174] Where M is the number of abnormal behavior categories, p i is the probability of the i-th abnormal category, w i is the weight coefficient of this category; the specific weight coefficients are set as follows: wandering behavior 0.6, climbing behavior 0.8, abnormal operation 0.7, intrusion 0.9, object handling 0.7, other abnormal behavior 0.8; S category is the calculated anomaly score based on the behavior category;

[0175] Duration of behavior. The longer the abnormal behavior lasts, the more obvious the abnormality is:

[0176]

[0177] Among them, T current is the duration of the current behavior (unit: seconds), T threshold is the threshold time (unit: seconds); for different types of abnormal behaviors, set different threshold times: 60 seconds for wandering behavior, 10 seconds for climbing behavior, 30 seconds for abnormal operation, 5 seconds for intrusion, 20 seconds for object handling, and 15 seconds for other abnormal behaviors; S duration is the calculated duration-based anomaly score, whose value is limited to between 0 and 1;

[0178] Different areas in the room have different sensitivities:

[0179] S region=R(x,y);

[0180] Where R(x,y) is the regional sensitivity coefficient at the coordinate (x,y); the storage area is divided into high-sensitivity area (sensitivity coefficient 1.0), medium-sensitivity area (sensitivity coefficient 0.7) and low-sensitivity area (sensitivity coefficient 0.4), according to the importance and security level of the stored items; S region is the calculated region-based anomaly score;

[0181] Time sensitivity: different time periods have different sensitivities (e.g., more sensitive at night):

[0182] S time =T(h);

[0183] Where T(h) is the time sensitivity coefficient of the current hour h. The system divides 24 hours a day into three time periods: working time (8:00-18:00, sensitivity coefficient 0.5), non-working time (18:00-23:00 and 6:00-8:00, sensitivity coefficient 0.8) and night time (23:00-6:00, sensitivity coefficient 1.0); S time is the calculated time-based anomaly score;

[0184] The formula for calculating the comprehensive anomaly score is:

[0185] S anomaly =α·S category +β·S duration +γ·S region +δ·S time ;

[0186] Among them, α, β, γ, and δ are the weight coefficients of behavior category, behavior duration, occurrence area, and time sensitivity, respectively, satisfying α+β+γ+δ=1; the default values ​​are 0.5, 0.2, 0.2, and 0.1 respectively; S anomaly is the final calculated comprehensive anomaly score;

[0187] Output: Anomaly score S anomaly , the value range is 0-1.

[0188] S500, based on a multi-level threshold judgment mechanism to trigger different levels of warning, and combined with trajectory analysis technology to enhance decision-making accuracy;

[0189] In one embodiment of the present invention, a multi-level threshold determination mechanism is adopted to trigger different levels of response based on the output anomaly score. The specific settings are as follows:

[0190] S501, multi-level threshold determination; normal level (green): S anomaly<T1; The behavior is recorded in the system log but no alarm is triggered;

[0191] Attention level (yellow): T1 ≤ S anomaly <T2; The system marks it as an event to be concerned about; Increase the monitoring frequency; Notify the on-duty personnel to pay attention;

[0192] Warning level (orange): T2 ≤ S anomaly <T3; Send a warning notice to the security personnel; Increase the acquisition frequency of surrounding cameras; Save the video clips related to the event;

[0193] Emergency level (red): S anomaly ≥T3; Trigger a high-level alarm; Send a real-time alarm to all security terminals; Initiate emergency response measures; Save high-definition video evidence;

[0194] Among them, T1, T2, and T are the first, second, and third thresholds of the anomaly score respectively, and the default values are 0.3, 0.6, and 0.8 respectively;

[0195] Output: Early warning level (normal, attention, warning or emergency) and corresponding response measures;

[0196] S502, trajectory analysis; In addition to single-frame behavior recognition, the system also analyzes the trajectory information of the detected moving targets (obtained from the moving target position information in the moving target detection result) to identify abnormal movement patterns. Trajectory analysis mainly includes the following steps:

[0197] Moving target tracking, using the Kalman filter to track the detected moving targets;

[0198] State representation, the system uses the state vector X k to represent the state of the target at time k, including the position (x, y) and speed (v<00,00050>, v y );

[0199] Prediction step, based on the state estimate value of the previous moment and the state transition model, predict the state of the current moment:

[0200] X k|k-1 = F k X k-1|k-1 + B k u k ;

[0201] Among them, X k|k-1 is the prior state estimate at time k, X k-1|k-1 is the posterior state estimate at time k-1, F k is the state transition matrix, B k [[ID=6,3]]is the control matrix, u kis the control vector;

[0202] Update step, combining observations to correct the state estimate:

[0203] X k|k =X k|k-1 +K k (Z k -H k X k|k-1 );

[0204] Among them, X k|k is the posterior state estimate at time k, K k is the Kalman gain, Z k is the observation value at time k, H k is the observation matrix;

[0205] Output: target trajectory data, including position sequence (x i ,y i ) and the corresponding timestamp t i ;

[0206] Trajectory feature extraction, calculating trajectory curvature, speed change, dwell time and other features;

[0207] Average speed, calculate the average moving speed of each point in the trajectory:

[0208]

[0209] Among them, (x i ,y i )、(x i+1 ,y i+1i ) are the coordinates of the i-th and i+1-th points on the trajectory, n1 is the number of trajectory points, Δt is the sampling time interval (unit: second), v avg is the calculated average speed;

[0210] The trajectory curvature reflects the curvature of the path:

[0211]

[0212] Where x′ i , y′ i is the first-order derivative (velocity) at trajectory point i, x″ i ,y″ i is the second-order derivative (acceleration) at trajectory point i, κ i is the calculated curvature at trajectory point i;

[0213] Dwell time, which calculates the time a target stays in a specific area:

[0214]

[0215] Among them, t i is the target’s i-th stay time in the area, m is the number of stays, T stay is the calculated total residence time;

[0216] Output: Trajectory feature set, including average speed v avg , trajectory curvature κ i and residence time T stay ;

[0217] Trajectory anomaly score, calculates the trajectory anomaly score based on the extracted features:

[0218] S trajectory =w1S speed +w2S curvature +w3S stay +w4S pattern ;

[0219] Among them, S speed is the speed anomaly score, S curvature is the curvature anomaly score, S stay is the resident anomaly score, S pattern is the pattern anomaly score, w1, w2, w3 and w4 are the weight coefficients of speed, curvature, residence and pattern anomaly respectively, and the default values ​​are 0.3, 0.2, 0.3 and 0.2 respectively. trajectory is the calculated trajectory anomaly score;

[0220] Output: Trajectory anomaly score S trajectory , the value range is 0-1;

[0221] The trajectory anomaly score is combined with the behavior anomaly score to form the final comprehensive anomaly score:

[0222] S final =λ1S anomaly +(1-λ1)S trajectory

[0223] Among them, λ1 is the trade-off coefficient, the default value is 0.7, S anomaly is the final calculated comprehensive anomaly score, S trajectory is the calculated trajectory anomaly score, S final is the final comprehensive anomaly score calculated;

[0224] Output: Final comprehensive anomaly score S final , the value range is 0-1, used for multi-level threshold determination;

[0225] S600, which adaptively adjusts detection parameters according to environmental changes, can stably detect and provide early warnings for abnormal behavior under different conditions;

[0226] In one embodiment of the present invention, the environment adaptation module automatically adjusts detection parameters according to environmental changes to improve the performance of the system under different conditions. The main functions include:

[0227] S601, Environmental Parameter Perception: The system monitors environmental parameters in real time, including:

[0228] Light intensity, by analyzing the average brightness and contrast of the video frames:

[0229]

[0230] Where W and H are the width and height of the image, I(x,y) is the brightness value of the pixel at the coordinate (x,y), and I light is the calculated average light intensity;

[0231] Background complexity, by calculating the texture complexity and edge density of the background area:

[0232]

[0233] in, is the gradient magnitude of the image at coordinate (x, y), C bg is the calculated background complexity;

[0234] Crowd density, the number and distribution of human bodies detected by statistics:

[0235]

[0236] Among them, N people is the number of people detected, A region is the area of ​​the monitoring area, D crowd is the calculated crowd density;

[0237] Time periodicity: adjust the system sensitivity according to the time period. The system records the average activity level every hour and establishes a 24-hour activity pattern model to identify normal activity patterns in different time periods;

[0238] S602, parameter adaptive adjustment: Based on the detected environmental changes, the system automatically adjusts the following parameters:

[0239] Preprocessing parameters:

[0240] Gaussian filter standard deviation adjustment: Dynamically adjust the filter strength as lighting conditions change:

[0241]

[0242] Among them, σ base is the benchmark standard deviation, α1 is the first adjustment coefficient, I light is the current light intensity, I ref is the reference brightness value, σ gauss is the adjusted Gaussian filter standard deviation;

[0243] Adjust the target value of illumination normalization, and adjust it adaptively according to the ambient light intensity:

[0244]

[0245] Among them, μ base is the reference target brightness, β1 is the second adjustment coefficient, I light is the current light intensity, I ref is the reference brightness value, μ target is the adjusted target brightness value;

[0246] Motion detection parameters:

[0247] Gaussian background model learning rate adjustment, according to the background complexity and crowd density:

[0248] α GMM =α base ×(1+γ1×C bg +δ1×D crowd );

[0249] Among them, α base is the baseline learning rate, γ1 and δ1 are the third and fourth adjustment coefficients respectively, C bg is the background complexity, D crowd is the crowd density, α GMM is the adjusted Gaussian mixture model learning rate;

[0250] Abnormal judgment parameters:

[0251] Adjust the anomaly score threshold based on time factor and crowd density:

[0252] T i ′=T i ×(1-∈×(S time -0.5)-ζ×(D crowd -D ref ));

[0253] Among them, T i is the original threshold, ∈ and ζ are adjustment coefficients (default is 0.2 and 0.3 respectively), S time is the time periodicity factor (range 0-1), D crowd is the current crowd density, Dref is the reference crowd density, T i ′ is the adjusted threshold;

[0254] S603, online model update. The system supports online incremental learning of the model to adapt to environmental changes and new intrusion behaviors:

[0255] Sample collection and screening:

[0256] Automatically collect high-confidence samples, and the system automatically saves boundary samples whose anomaly scores are close to the threshold;

[0257] Security personnel feedback samples, add the real abnormal and false positive samples confirmed by security personnel into the training set;

[0258] Sample quality assessment and screening: screening high-quality samples through feature diversity and representativeness analysis;

[0259] Incremental learning mechanism:

[0260] Memory replay: retains representative samples of the original training data and participates in training together with new samples to prevent catastrophic forgetting;

[0261] Parameter migration: retain most of the model parameters and only fine-tune the last few layers to maintain model stability;

[0262] Regularization constraints, applying elastic weight integration techniques to balance new and old knowledge;

[0263] Update strategy:

[0264] Regular updates: the system performs automatic incremental learning once a week to update model parameters;

[0265] Triggered update: when the cumulative number of new samples reaches the preset threshold (100 samples by default), the update is triggered;

[0266] Distributed update: Model updates are performed on the server side, and the updated model parameters are distributed to edge devices;

[0267] In one embodiment of the present invention, an application example of the aforementioned deep learning-based intelligent detection method for a warehouse intrusion is provided: This example focuses on upgrading the security system of a museum's cultural relics warehouse, which stores a variety of precious cultural relics and requires high security protection;

[0268] Application Environment: The museum's underground cultural relic storage room, approximately 500 square meters, is divided into five independent areas for storing different types of cultural relics. The lighting conditions within the storage room are complex, including natural light during the day, emergency lighting at night, and special light sources for display. Personnel flow in the storage room includes routine staff inspections, cultural relic management, and expert research visits.

[0269] System Deployment: 20 high-definition cameras and 12 infrared sensors were installed in the warehouse, covering all key areas and entrances. Edge computing equipment was deployed in the warehouse control room, and servers were deployed in the museum's data center, integrating with the existing security system.

[0270] The system performance comparison data is shown in the following table:

[0271]

[0272] The performance of the system under different lighting conditions is shown in the following table:

[0273] Lighting conditions Detection accuracy of traditional security systems Detection accuracy of the system of the present invention Improvement ratio Normal daylight 82.3％ 97.1％ 18.0％ Low-light environment 65.7％ 94.8％ 44.3％ Night emergency lighting 58.2％ 93.5％ 60.7％ Light mutation 51.4％ 92.7％ 80.4％ Local strong light interference 60.8％ 91.9％ 51.2％

[0274] The trend of the system's incremental learning effect over time is shown in the following table:

[0275] Runtime Initial Model 1 month later 3 months later 6 months later 12 months later Detection accuracy 92.5％ 94.1％ 95.3％ 96.2％ 97.8％ False alarm rate 5.6％ 4.2％ 3.5％ 2.8％ 1.9％ New anomaly recognition rate Baseline +15.3％ +28.7％ +42.1％ +56.4％

[0276] Application results: Six months after the system was deployed, it successfully alerted 12 potential security incidents, including: identifying abnormal behavior of unauthorized personnel in specific areas; detecting staff members' illegal handling of cultural relics; discovering abnormal warehouse environment (temperature and humidity fluctuations caused by personnel emergency handling); and identifying suspicious behavior patterns of multiple people collaborating.

[0277] The system also demonstrated excellent environmental adaptability, maintaining stable performance under conditions such as changes in lighting and fluctuations in crowd density. Through continuous optimization through incremental learning, the false alarm rate has decreased month by month. The work efficiency of security personnel has been improved, and they can focus on high-risk events marked by the system, reducing the workload of routine monitoring. By comparison with traditional security systems, it can be seen that the intelligent detection system for warehouse intrusion of the present invention has achieved certain improvements in various key indicators, especially reducing the false alarm rate and missed alarm rate, which is crucial for the protection of high-value cultural relics. The system's adaptive ability ensures stable performance under different environmental conditions, and the incremental learning mechanism enables the system to continuously improve its detection capabilities as the running time increases.

[0278] In practical applications, the system also exhibits the following characteristics:

[0279] Scalability: The framework design supports flexible deployment in warehouses of different sizes, from small cultural relics warehouses to large storage facilities.

[0280] Robustness: The system is tolerant to partial camera failures and maintains overall monitoring effectiveness through data fusion.

[0281] Low maintenance: Self-learning and adaptive mechanisms reduce the need for manual maintenance and parameter adjustments, lowering operating costs.

[0282] Privacy protection: The system only stores videos related to abnormal events, reducing unnecessary data storage and privacy risks.

[0283] In addition, the system can also be expanded according to actual needs, such as integration with access management systems, adding voice alarm modules, integrating IoT devices, etc., to provide customized solutions for different types of warehouse security.

[0284] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A deep learning-based intelligent detection method for warehouse intrusion, characterized in that: The following steps are involved: Acquire warehouse monitoring information through multi-source sensing equipment and pre-process the collected video data. The warehouse monitoring information includes video data and heat data; Extract spatiotemporal features from preprocessed video data, including spatial features based on deep convolutional neural networks and temporal features based on optical flow methods; Build a two-stream fusion model based on deep convolutional neural networks and long short-term memory networks to integrate spatial and temporal features; Analyze the extracted spatiotemporal features based on deep learning models to identify intrusion behaviors in the warehouse, including behavior classification and anomaly score calculation; Based on a multi-level threshold judgment mechanism, different levels of warnings are triggered, and trajectory analysis technology is combined to enhance decision-making accuracy; The detection parameters are adaptively adjusted according to environmental changes, and abnormal behavior detection and early warning can be stably achieved under different conditions.

2. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 1 is characterized in that: In the step of obtaining warehouse monitoring information through multi-source sensing equipment, multiple cameras and auxiliary infrared sensors are deployed in key areas of the warehouse to collect video and heat information in real time; Among them, cameras with full HD resolution or higher resolution are used, with a frame rate of no less than 25 frames per second, and infrared sensors are installed at key locations of the warehouse entrance and windows to detect changes in human body heat and provide auxiliary judgment information for video analysis.

3. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 1 is characterized in that: The steps for preprocessing the collected video data include: Video decoding and frame extraction, image denoising and enhancement, illumination normalization, and moving target detection; illumination normalization uses an adaptive illumination normalization algorithm, specifically: Among them, I(x,y) represents the pixel value of the original image at the coordinate (x,y), I norm (x, y) represents the normalized pixel value, μ represents the average brightness value of the original image, σ represents the standard deviation of the original image, and μ target and σ target are the target mean brightness and standard deviation, respectively.

4. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 3 is characterized in that: Moving target detection uses an improved mixed Gaussian background modeling algorithm and applies morphological operations, including dilation and erosion, to the detected moving targets to remove noise and fill holes. The dilation operation is defined as: The erosion operation is defined as: Among them, A represents the binary image, representing the foreground mask; B represents the structural element, defining the shape and size of the morphological operation; B z Represents the result of translating the structural element B to point z; represents the expansion operator; Represents the erosion operator.

5. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 1 is characterized in that: In the step of extracting spatiotemporal features from preprocessed video data: Spatial feature extraction uses a pre-trained ResNet-50 network, removing the last fully connected layer and retaining the feature maps extracted by the convolutional layer; For each input video frame, the size is adjusted to 224×224×3, and a 2048-dimensional feature vector is obtained; Temporal feature extraction uses the optical flow method to calculate the motion vector between multiple consecutive frames of target area images, and uses the Farneback algorithm to calculate the dense optical flow field. The horizontal and vertical optical flow fields are calculated for each pair of consecutive frames, and the optical flow fields of 10 consecutive frames are stacked to form an optical flow volume with a size of 224×224×20.

6. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 1 is characterized in that: The steps to build a dual-stream fusion model include: Construct a spatial stream network, receive the processed spatial feature vector, and output a 512-dimensional spatial stream feature vector through a fully connected layer and activation function; Construct a temporal flow network, use a 3D convolutional neural network and a long short-term memory network to process the temporal information in the optical flow volume, and output a 128-dimensional temporal flow feature vector; A feature fusion network is constructed to concatenate 512-dimensional spatial features and 128-dimensional temporal features into a 640-dimensional vector. Through the self-attention layer and the fully connected layer, a 256-dimensional fused feature vector and the probability distribution of behavior categories are output.

7. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 1 is characterized in that: In the step of identifying intrusion behavior in the warehouse, the behaviors in the warehouse are divided into the following categories: normal walking, normal operation, wandering behavior, climbing behavior, abnormal operation, intrusion behavior, object handling, and other abnormal behaviors. The anomaly score is calculated based on the probability distribution of the model output. The comprehensive anomaly score calculation formula is: S anomaly =α·S category +β·S duration +γ·S region +δ·S time ; Among them, α, β, γ, and δ are the weight coefficients of behavior category, behavior duration, occurrence area, and time sensitivity, respectively, satisfying α+β+γ+δ=1; S anomaly is the final calculated comprehensive anomaly score.

8. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 1 is characterized in that: The steps for triggering different levels of warnings based on the multi-level threshold judgment mechanism are as follows: When the anomaly score S anomaly <is less than T1, it is at the normal level, and the behavior is recorded in the system log but no alarm is triggered; When T1 ≤ S anomaly <When T2, it is the attention level, and the system marks it as an event that needs attention; When T2 ≤ S anomaly When T3, it is at the warning level, and a warning notice is sent to the security personnel; When S anomaly When ≥T3, it is an emergency level and triggers a high-level alarm; Among them, T1, T2, and T3 are the first, second, and third thresholds of the anomaly score, respectively.

9. The method for intelligent detection of warehouse intrusion based on deep learning according to claim 8 is characterized in that: The steps to enhance decision-making accuracy by combining trajectory analysis technology include: Use Kalman filter to track the detected moving target; Calculate the average velocity, trajectory curvature, and dwell time characteristics of the trajectory; Calculate the trajectory anomaly score S based on the extracted features trajectory ; Combine the trajectory anomaly score with the behavior anomaly score to form the final comprehensive anomaly score: S final =λ1S anomaly +(1-λ1)S trajectory ; Among them, λ1 is the trade-off coefficient, S anomaly is the final calculated comprehensive anomaly score, S trajectory is the calculated trajectory anomaly score, S final is the final calculated comprehensive anomaly score.

10. A deep learning-based intelligent detection system for a warehouse intrusion, configured to execute the steps of a deep learning-based intelligent detection method for a warehouse intrusion according to any one of claims 1 to 9, characterized in that: include: Multi-source sensor acquisition module, used to obtain warehouse monitoring information, including high-definition cameras and infrared sensors; The multi-source sensing acquisition module is responsible for deploying multiple full-HD cameras and infrared sensors in key areas of the storage room to collect video streams and thermal data in real time; The video preprocessing module is used to decode, denoise, enhance, normalize illumination, and detect moving targets on the collected video data. The video preprocessing module receives the original video stream and decodes it, applies a Gaussian filter algorithm to remove noise, implements adaptive illumination normalization, and performs moving target detection using an improved mixed Gaussian background modeling algorithm. Feature extraction module, used to extract spatiotemporal features from preprocessed video data; The feature extraction module uses a pre-trained ResNet-50 network to extract spatial features and generate a 2048-dimensional feature vector. It also uses the Farneback algorithm to calculate dense optical flow fields to extract temporal features and form optical flow volumes, providing a complete spatiotemporal feature representation for subsequent deep learning models. The deep learning model module is used to build a dual-stream fusion model and perform behavior recognition and anomaly score calculation. The deep learning model module consists of three sub-modules: spatial stream network, temporal stream network, and feature fusion network. It fuses spatial and temporal features through a self-attention mechanism. The early warning decision module is used to trigger different levels of early warnings based on a multi-level threshold judgment mechanism; The early warning decision module uses the calculated anomaly score, combined with trajectory analysis technology, and a multi-level threshold judgment mechanism to classify abnormal events into four levels: normal, attention, warning, and emergency. It then triggers corresponding early warning measures to ensure the timeliness and accuracy of safety responses. The environmental adaptation module is used to automatically adjust detection parameters according to environmental changes; the environmental adaptation module continuously monitors the environmental conditions of the storage room and automatically adjusts the detection threshold, feature weights and model parameters through a feedback learning mechanism to ensure that the system can maintain stable detection performance under various environmental conditions and improve the robustness and adaptability of the system.

Images