Intelligent security method and system based on machine vision

Abstract

The invention discloses an intelligent security and protection method and system based on machine vision, and relates to the technical field of intelligent security and protection. Collecting video data in the target area, and performing frame extraction processing on the video data to obtain an image recognition set; performing denoising processing on images in the image recognition set to obtain an effective image set; substituting each image in the effective image set into a preset model to obtain a corresponding detection object in each image; and according to the detection object corresponding to each image in the effective image set, establishing a change track of the detection object in the time sequence, carrying out abnormal behavior detection according to the change track, and giving an alarm if an abnormal behavior is detected. According to the method, the image quality is improved through image denoising processing, then the preset model is introduced to enhance the recognition capability of the partially-shielded target, and after the object is recognized, the time sequence track of the detected object is established, so that the behavior mode of the object is analyzed, abnormal behaviors are detected more easily, the risk of misinformation or missing report is reduced, and the accuracy of system recognition is improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent security technology, and in particular relates to an intelligent security method and system based on machine vision. Background Art

[0002] With the rapid development of artificial intelligence (AI), traditional security systems are gradually evolving towards intelligent systems. Traditional security measures, which rely on manual inspections and infrared sensing, suffer from high false alarm rates, slow response times, and poor environmental adaptability. In recent years, machine vision, a key branch of AI, has demonstrated tremendous potential in image recognition, object detection, and behavioral analysis. Applying machine vision technology to security systems helps achieve efficient identification and rapid response to unusual events.

[0003] Patent CN117523443B discloses a community intelligent security system based on machine vision and its implementation method. It obtains high-definition video of the window sill of a building through a shooting unit, continuously captures instantaneous images with a fixed time interval as a step size, and evenly divides the instantaneous image into cell areas with equal step sizes. The pixel mean of the corresponding cell area in the instantaneous image is obtained, and a pixel mean matrix of the cell area is generated. The pixel mean difference matrix of the cell area between the initial image and the intermediate image and the intermediate image and the node image is calculated. According to the calculation results of the pixel mean difference matrix of the cell area, it is determined whether parabolic motion has occurred. When it is determined that a falling object has occurred, an early warning signal is generated, and an alarm operation is performed when a falling object occurs. When a falling object occurs, an alarm warning can be issued in time to remind pedestrians below the building to take emergency avoidance.

[0004] However, in actual application scenarios, existing defense systems may partially block target objects such as people and objects due to environmental occlusions, such as obstructions, insufficient light, and covert behavior, resulting in reduced system recognition accuracy and the risk of false alarms or missed alarms. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem that target objects such as people and objects may be partially obscured due to environmental obstructions, such as obstructions, insufficient light, covert behavior, etc., resulting in reduced system recognition accuracy and the risk of false alarms or missed alarms, and to propose an intelligent security method and system based on machine vision.

[0006] In a first aspect of the present invention, a machine vision-based intelligent security method is first proposed, the method comprising:

[0007] Collecting video data within a target area, and performing frame extraction processing on the video data to obtain an image recognition set;

[0008] Performing denoising on the images in the image recognition set to obtain a valid image set;

[0009] Substituting each image in the valid image set into a preset model to obtain the corresponding detection object in each image;

[0010] According to the detection object corresponding to each image in the effective image set, a change trajectory of the detection object in a time series is established, abnormal behavior detection is performed based on the change trajectory, and an alarm is issued if abnormal behavior is detected.

[0011] Optionally, performing denoising on the images in the image recognition set to obtain a valid image set includes:

[0012] For each image in the image recognition set, convert the image into YUV space to obtain a Y channel, a U channel, and a V channel;

[0013] Performing dual-tree complex wavelet transform to denoise the Y channel to obtain a denoised Y channel, and reconstructing the denoised Y channel through inverse dual-tree complex wavelet transform to obtain a reconstructed Y channel;

[0014] Perform guided filtering on the U channel and the V channel through the denoised Y channel to obtain the reconstructed U channel and the reconstructed V channel;

[0015] Perform RGB conversion on the reconstructed Y channel, the reconstructed U channel, and the reconstructed V channel to obtain valid images, and obtain a valid image set based on all valid images.

[0016] Optionally, substituting each image in the valid image set into a preset model to obtain a corresponding detection object in each image includes:

[0017] Substituting the first image into the first convolution module and the second convolution module in sequence to obtain a first convolution feature; the first image is any one of the valid image set;

[0018] Substituting the first convolution image into a preset feature enhancement module to obtain a first enhanced feature;

[0019] Substituting the first enhanced feature into the third convolution module to obtain a second convolution feature, and substituting the second convolution feature into the preset feature enhancement module to obtain a second enhanced feature;

[0020] Substituting the second enhanced feature into the third convolution module to obtain a third convolution feature, substituting the second convolution feature into the preset feature enhancement module to obtain a third enhanced feature, and substituting the third enhanced feature into the pooling module to obtain a first pooling feature;

[0021] Substitute the first enhanced feature and the first pooled feature into the attention enhancement module to obtain a first attention feature, and substitute the first attention feature into the C2f module to obtain a first fusion feature;

[0022] Substituting the first fused feature and the second enhanced feature into the attention enhancement module to obtain a second attention feature, and substituting the second attention feature into the C2f module to obtain a second fused feature;

[0023] Substitute the second fusion feature into the detection head to obtain the detection object.

[0024] Optionally, the attention enhancement module includes a first input module, a second input module and a first output module;

[0025] The first input module receives a first input feature, the second input module receives a second input feature, and concatenates the first input feature and the second input feature to obtain a first concatenated feature;

[0026] Substituting the first splicing feature into the first target convolution module to obtain an output splicing feature;

[0027] Substituting the output splicing features into the first attention enhancement module and the second attention enhancement module respectively to obtain the first attention output feature and the second attention output feature;

[0028] Multiply the first attention output feature and the second attention output feature element-wise, and normalize them using a Sigmoid function to obtain a fusion weight;

[0029] Performing feature fusion on the first attention output feature and the second input feature according to the fusion weight to obtain a first attention fusion feature;

[0030] Performing feature fusion on the second attention output feature and the first input feature according to the fusion weight to obtain a second attention fusion feature;

[0031] The first attention fusion feature and the second attention fusion feature are fused to obtain an output feature, and the output feature is output from the first output module.

[0032] Optionally, establishing a change trajectory of the detection object in a time series, and performing abnormal behavior detection based on the change trajectory includes:

[0033] The detection object corresponding to each image in the valid image set is associated with the object by using the Hungarian algorithm, an independent ID is created for each object that appears, and a trajectory change is established;

[0034] For each ID, the Fourier descriptor of the trajectory corresponding to the ID is extracted, and the abnormal behavior corresponding to the ID is determined based on the Fourier descriptor and preset rules.

[0035] In a second aspect of the present invention, an intelligent security system based on machine vision is proposed, comprising:

[0036] A frame extraction processing module is used to collect video data in the target area and perform frame extraction processing on the video data to obtain an image recognition set;

[0037] A denoising module, configured to perform denoising on the images in the image recognition set to obtain a valid image set;

[0038] A detection object determination module, configured to substitute each image in the valid image set into a preset model to obtain a detection object corresponding to each image;

[0039] The abnormal behavior detection module is used to establish a change trajectory of the detection object in a time series based on the detection object corresponding to each image in the valid image set, perform abnormal behavior detection based on the change trajectory, and issue an alarm if abnormal behavior is detected.

[0040] Optionally, the denoising module includes:

[0041] An image conversion module is used to convert each image in the image recognition set into a YUV space to obtain a Y channel, a U channel, and a V channel;

[0042] A Y channel reconstruction module is used to perform dual-tree complex wavelet transform denoising on the Y channel to obtain a denoised Y channel, and reconstruct the denoised Y channel by inverse dual-tree complex wavelet transform to obtain a reconstructed Y channel;

[0043] A guided filtering module is used to perform guided filtering on the U channel and the V channel through the denoised Y channel to obtain a reconstructed U channel and a reconstructed V channel;

[0044] The RGB conversion module is used to perform RGB conversion on the reconstructed Y channel, the reconstructed U channel and the reconstructed V channel to obtain a valid image, and obtain a valid image set based on all the valid images.

[0045] Optionally, the detection object determination module includes:

[0046] a first convolution feature determination module, configured to sequentially substitute a first image into a first convolution module and a second convolution module to obtain a first convolution feature; the first image being any one of the valid image set;

[0047] a first enhanced feature determination module, configured to substitute the first convolution image into a preset feature enhancement module to obtain a first enhanced feature;

[0048] a second enhanced feature determination module, configured to substitute the first enhanced feature into a third convolution module to obtain a second convolution feature, and substitute the second convolution feature into the preset feature enhancement module to obtain a second enhanced feature;

[0049] a first pooling feature determination module, configured to substitute the second enhanced feature into a third convolution module to obtain a third convolution feature, substitute the second convolution feature into the preset feature enhancement module to obtain a third enhanced feature, and substitute the third enhanced feature into a pooling module to obtain a first pooling feature;

[0050] A first fusion feature determination module is configured to substitute the first enhanced feature and the first pooled feature into an attention enhancement module to obtain a first attention feature, and substitute the first attention feature into a C2f module to obtain a first fusion feature;

[0051] A second fusion feature determination module is configured to substitute the first fusion feature and the second enhancement feature into the attention enhancement module to obtain a second attention feature, and substitute the second attention feature into the C2f module to obtain a second fusion feature;

[0052] The detection object generation module is used to substitute the second fusion feature into the detection head to obtain the detection object.

[0053] Optionally, the attention enhancement module includes a first input module, a second input module and a first output module;

[0054] A first splicing feature determination module, configured to receive a first input feature from the first input module, receive a second input feature from the second input module, and splice the first input feature and the second input feature to obtain a first splicing feature;

[0055] an output splicing feature determination module, configured to substitute the first splicing feature into a first target convolution module to obtain an output splicing feature;

[0056] an attention feature output module, configured to substitute the output splicing feature into the first attention enhancement module and the second attention enhancement module respectively to obtain a first attention output feature and a second attention output feature;

[0057] a fusion weight determination module, configured to multiply the first attention output feature and the second attention output feature element-wise and normalize them using a Sigmoid function to obtain a fusion weight;

[0058] a first attention fusion feature determination module, configured to perform feature fusion on the first attention output feature and the second input feature according to the fusion weight to obtain a first attention fusion feature;

[0059] A second attention fusion feature determination module is used to perform feature fusion on the second attention output feature and the first input feature according to the fusion weight to obtain a second attention fusion feature;

[0060] An output feature determination module is used to perform feature fusion on the first attention fusion feature and the second attention fusion feature to obtain an output feature, and the output feature is output from the first output module.

[0061] Optionally, the abnormal behavior detection module includes:

[0062] An object association module is used to associate the detection objects corresponding to each image in the valid image set using the Hungarian algorithm, create an independent ID for each object that appears, and establish a trajectory change;

[0063] The abnormal behavior determination module is used to extract the Fourier descriptor of the trajectory corresponding to each ID, and determine the abnormal behavior corresponding to the ID based on the Fourier descriptor and preset rules.

[0064] Beneficial effects of the present invention:

[0065] The present invention proposes an intelligent security method based on machine vision, which collects video data in a target area, extracts frames from the video data to obtain an image recognition set, performs denoising on the images in the image recognition set to obtain a valid image set, substitutes each image in the valid image set into a preset model to obtain the corresponding detection object in each image, establishes a change trajectory of the detection object in a time series based on the detection object corresponding to each image in the valid image set, performs abnormal behavior detection based on the change trajectory, and issues an alarm if abnormal behavior is detected. After improving the image quality through image denoising, the preset model is inserted to enhance the recognition ability of partially occluded targets. After the object is identified, a time series trajectory of the detection object is established to analyze its behavior pattern, making it easier to detect abnormal behavior, reducing the risk of false positives or false negatives, and improving the accuracy of system recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The present invention will be further described below with reference to the accompanying drawings.

[0067] Figure 1 A flowchart of an intelligent security method based on machine vision provided by an embodiment of the present invention;

[0068] Figure 2 A framework diagram of a machine vision-based intelligent security system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0070] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0071] The embodiment of the present invention provides an intelligent security method based on machine vision. Figure 1 , Figure 1 A flowchart of an intelligent security method based on machine vision provided by an embodiment of the present invention. The method includes the following steps:

[0072] S101, collecting video data within the target area, and performing frame extraction processing on the video data to obtain an image recognition set;

[0073] S102, performing denoising processing on the images in the image recognition set to obtain a valid image set;

[0074] S103, substituting each image in the valid image set into a preset model to obtain the corresponding detection object in each image;

[0075] S104, establishing a change trajectory of the detection object in a time series based on the detection object corresponding to each image in the valid image set, performing abnormal behavior detection based on the change trajectory, and issuing an alarm if abnormal behavior is detected.

[0076] An intelligent security method based on machine vision provided by an embodiment of the present invention improves image quality through image denoising processing and then introduces a preset model to enhance the recognition ability of partially occluded targets. After the object is recognized, a time series trajectory of the detection object is established to analyze its behavior pattern, thereby making it easier to detect abnormal behavior, reducing the risk of false alarms or missed alarms, and improving the accuracy of system recognition.

[0077] In one implementation, by extracting frames from the collected video data according to a preset time period, the number of processed video frames can be effectively reduced, thereby reducing the data volume and computing load. This makes subsequent image processing and recognition work more efficient, while shortening the overall processing delay; the preset time period is determined by technical personnel; after extracting frames, an image recognition set is obtained, and here all obtained are RGB images.

[0078] In one implementation, denoising is performed on the image set after frame extraction to improve image quality and ensure that the subsequent recognition model inputs clear and clean data, thereby improving detection accuracy; if no abnormal behavior is detected, no action is taken.

[0079] In one implementation, each valid image is substituted into a pre-trained model for detection, which can accurately identify the detection object in the image and establish a trajectory based on the change of the detection object over time. By analyzing the trajectory, it is possible to effectively determine whether there is abnormal behavior, thereby reducing false alarms and missed alarms, making anomaly detection more precise and reliable.

[0080] In one implementation, based on the changing trajectory of the detection object in consecutive frames, behavioral changes can be captured in real time, thereby achieving timely detection of abnormal behavior. Once the system detects abnormal behavior, an alarm is immediately issued.

[0081] In one embodiment, performing denoising on images in the image recognition set to obtain a valid image set includes:

[0082] For each image in the image recognition set, convert the image to YUV space to obtain Y channel, U channel and V channel;

[0083] Performing dual-tree complex wavelet transform to denoise the Y channel to obtain a denoised Y channel, and reconstructing the denoised Y channel through inverse dual-tree complex wavelet transform to obtain a reconstructed Y channel;

[0084] Perform guided filtering on the U channel and the V channel through the denoised Y channel to obtain the reconstructed U channel and the reconstructed V channel;

[0085] Perform RGB conversion on the reconstructed Y channel, the reconstructed U channel, and the reconstructed V channel to obtain valid images, and obtain a valid image set based on all valid images.

[0086] In one implementation, luminance information contains most structural details such as edges and textures. Applying a dual-tree complex wavelet transform (DTCWT) to the Y channel can effectively separate noise from image features at multiple scales and directions, achieving significant denoising results.

[0087] In one implementation, the U / V channels usually contain less structural information, and direct processing can easily lead to color distortion. Using the Y channel to guide filtering can smooth the colors while preserving edge consistency.

[0088] In one implementation, the dual-tree complex wavelet transform has the advantages of strong directional selectivity and good translation invariance, making it more suitable for image denoising tasks than traditional wavelet transform. Compared with independent processing of RGB channels, the YUV space is more in line with human visual characteristics and is more effective in improving image quality. During color restoration, instead of directly performing strong denoising on the U / V channels, guided filtering is used to reduce color blocking effects or color casts and enhance color reproduction.

[0089] In one embodiment, substituting each image in the valid image set into a preset model to obtain the corresponding detection object in each image includes:

[0090] Substituting the first image into the first convolution module and the second convolution module in sequence to obtain a first convolution feature; the first image is any one in the valid image set;

[0091] Substituting the first convolution image into a preset feature enhancement module to obtain a first enhanced feature;

[0092] Substituting the first enhanced feature into the third convolution module to obtain the second convolution feature, and substituting the second convolution feature into the preset feature enhancement module to obtain the second enhanced feature;

[0093] Substituting the second enhanced feature into the third convolution module to obtain the third convolution feature, substituting the second convolution feature into the preset feature enhancement module to obtain the third enhanced feature, and substituting the third enhanced feature into the pooling module to obtain the first pooling feature;

[0094] Substitute the first enhanced feature and the first pooled feature into the attention enhancement module to obtain the first attention feature, and substitute the first attention feature into the C2f module to obtain the first fusion feature;

[0095] Substitute the first fusion feature and the second enhancement feature into the attention enhancement module to obtain the second attention feature, and substitute the second attention feature into the C2f module to obtain the second fusion feature;

[0096] Substitute the second fusion feature into the detection head to obtain the detection object.

[0097] In one implementation, by using different convolution modules and preset feature enhancement modules in sequence, various feature information from low-level to high-level in the image can be extracted, and the feature information can be refined and enhanced layer by layer.

[0098] In one implementation, the preset feature enhancement module includes an X edge extractor, a Y edge extractor, and three convolution blocks. When the input feature enters the preset feature enhancement module, it will pass through the X edge extractor, the Y edge extractor, and the first convolution block respectively to obtain the x feature, the y feature, and the convolution feature. These three features are then spliced to obtain a spliced feature. This spliced feature will pass through the second convolution block to obtain a convolved feature. The convolved feature is then averaged with the input feature and passed through the third convolution block to obtain the output feature of the preset feature enhancement module. The first convolution block performs two layers of 3×3 convolution, the second convolution block performs a single layer of 1×1 convolution, and the third convolution block performs a single layer of 3×3 convolution plus Sigmoi d. The output channel is consistent with the input. The preset feature enhancement module can strengthen key information during feature transmission, help maintain image details, and thus improve detection accuracy.

[0099] In one implementation, the first enhanced feature, the second enhanced feature, and the subsequent features processed by the pooling and attention enhancement modules are fused, which helps to comprehensively utilize information at all levels and from multiple angles, making the information used for target detection more comprehensive; by fusing different attention features with convolutional features through the C2f module, local details and global context information can be effectively integrated, thereby improving the robustness of feature expression.

[0100] In one implementation, all convolution modules are existing modules in the YOLOv8 detection framework. By repeatedly entering the preset feature enhancement module and feature reprocessing of the convolution module, the target features can be continuously enhanced, making the description of the target more accurate during the detection process. Through multi-path feature extraction and fusion, the system can better distinguish the target from the background and enhance the correlation between local features, thereby reducing the probability of false detection or missed detection; the detection head is the Head in the YOLOv8 detection framework.

[0101] In one implementation, the attention enhancement module can automatically focus on salient areas in the image, reduce background interference, and ensure that the detection head inputs features that have been enhanced, so that it can still effectively detect targets in complex backgrounds. The attention mechanism filters out noise information, enhances valuable features, realizes effective compression and utilization of information, and improves overall detection performance.

[0102] In one embodiment, the attention enhancement module includes a first input module, a second input module, and a first output module;

[0103] The first input module receives a first input feature, the second input module receives a second input feature, and concatenates the first input feature and the second input feature to obtain a first concatenated feature;

[0104] Substitute the first splicing feature into the first target convolution module to obtain the output splicing feature;

[0105] Substitute the output splicing features into the first attention enhancement module and the second attention enhancement module respectively to obtain the first attention output feature and the second attention output feature;

[0106] Multiply the first attention output feature and the second attention output feature element by element, and normalize them through the Sigma id function to obtain the fusion weight;

[0107] Perform feature fusion on the first attention output feature and the second input feature according to the fusion weight to obtain the first attention fusion feature;

[0108] The second attention output feature and the first input feature are fused according to the fusion weight to obtain the second attention fusion feature;

[0109] The first attention fusion feature and the second attention fusion feature are fused to obtain an output feature, and the output feature is output from the first output module.

[0110] In one implementation, the first attention enhancement module includes two convolutional layers, and the second attention enhancement module includes adaptive average pooling and two convolutional layers. The two attention modules respectively extract important information from two perspectives or channels, fully exploring the potential structures and patterns. The attention mechanism weights the input, which helps to "filter noise" and "enhance key areas" to better identify the target object.

[0111] In one implementation, the two input features are concatenated to form a richer joint representation, introducing cross-modal synergy and improving information capacity; convolution operations are performed on the concatenated features to extract local spatial relationships or high-order features, introducing nonlinear transformations and local perception to improve the discriminability of the fused features; and attention enhancement is performed on the output concatenated features to extract key areas or dimensions, thereby focusing more on more effective information fragments.

[0112] In one implementation, the first attention fusion feature is generated by the "first attention output and the second input feature", and the second attention fusion feature is generated by the "second attention output and the first input feature". This "cross-fusion" design improves the interaction depth between the two channels and strengthens the semantic complementation; by splicing the first input feature and the second input feature, the integrity of the original feature is retained and information aliasing caused by early addition is avoided.

[0113] In one implementation, weights are dynamically allocated through a dual attention mechanism to retain important features and suppress noise, and then a cross-fusion mechanism is used to enhance feature complementarity, reduce information loss, and improve the efficiency of feature fusion. The fusion weight is normalized by the Sigma ID function and used as the weight of the first attention output feature and the second attention output feature. Then, 1 minus the fusion weight is used as the weight of the first input feature and the second input feature. The first attention fusion feature and the second attention fusion feature are fused to average the features.

[0114] In one embodiment, establishing a change trajectory of the detection object in a time series and performing abnormal behavior detection based on the change trajectory includes:

[0115] The Hungarian algorithm is used to associate the detection objects corresponding to each image in the valid image set, create an independent ID for each object that appears, and establish trajectory changes;

[0116] For each ID, the Fourier descriptor of the trajectory corresponding to the ID is extracted, and the abnormal behavior corresponding to the ID is determined based on the Fourier descriptor and preset rules.

[0117] In one implementation, the Hungarian algorithm is used to assign detection boxes, ensuring the uniqueness and continuity of each target in the image and being able to handle multi-target situations. A unique ID is assigned to each target, establishing a complete trajectory, supporting individual analysis, avoiding confusion between multiple object information, and tracing back the target's motion history to provide contextual information for behavioral analysis.

[0118] In one implementation, the trajectory is mapped into Fourier space features, making the trajectory measurable and comparable, compressing trajectory information, and removing noise.

[0119] In one implementation, the preset rules can be restricted area intrusion detection, wandering identification, snatching behavior, stalking behavior, drunken behavior, etc.; restricted area intrusion detection is triggered when the low-frequency component (first component) of the trajectory suddenly increases and the movement direction points to the restricted area, the amplitude of the first component is greater than twice the historical average (rapidly approaching the restricted area) and the phase angle change rate is less than 10 degrees / frame (straight line motion); wandering identification is triggered when multiple low-frequency components (second to fourth components) have continuous high energy and low speed, the amplitude of the second component divided by the amplitude of the first component is greater than 0.6 (non-linear motion), and the total energy of the third and fourth components is greater than 30% of the total energy. %; snatching behavior is triggered when the energy of the high-frequency component explodes and the speed changes suddenly, the amplitude of the fifth component is greater than 3 times the average of the previous 10 frames (violent shaking), and the phase angle of the first component changes suddenly by more than 90 degrees (sharp turn); tailing behavior is triggered when the Fourier descriptors of two trajectories are highly similar and the distance is too close, and the cosine similarity of the two Fourier descriptors is greater than 0.9 (synchronous movement); drunken behavior is triggered when all components show irregular fluctuations and the speed gradually decreases, the entropy of the energy distribution of each component is greater than 1.5, and the attenuation rate of the amplitude of the first component per frame is greater than 5%. All of the above preset rules will issue corresponding alarms after being triggered.

[0120] Based on the same inventive concept, the present invention also provides an intelligent security system based on machine vision. Figure 2 , Figure 2 A framework diagram of a machine vision-based intelligent security system provided by an embodiment of the present invention includes:

[0121] A frame extraction processing module is used to collect video data in the target area and perform frame extraction processing on the video data to obtain an image recognition set;

[0122] A denoising module, used to denoise the images in the image recognition set to obtain a valid image set;

[0123] A detection object determination module is used to substitute each image in the valid image set into a preset model to obtain the corresponding detection object in each image;

[0124] The abnormal behavior detection module is used to establish the change trajectory of the detection object in the time series based on the detection object corresponding to each image in the valid image set, perform abnormal behavior detection based on the change trajectory, and issue an alarm if abnormal behavior is detected.

[0125] An intelligent security system based on machine vision provided by an embodiment of the present invention improves image quality through image denoising processing and then introduces a preset model to enhance the recognition ability of partially occluded targets. After the object is recognized, a time series trajectory of the detection object is established to analyze its behavior pattern, thereby making it easier to detect abnormal behavior, reducing the risk of false alarms or missed alarms, and improving the accuracy of system recognition.

[0126] In one embodiment, the denoising module includes:

[0127] An image conversion module is used to convert each image in the image recognition set into a YUV space to obtain a Y channel, a U channel, and a V channel;

[0128] A Y channel reconstruction module is used to perform dual-tree complex wavelet transform denoising on the Y channel to obtain a denoised Y channel, and reconstruct the denoised Y channel by inverse dual-tree complex wavelet transform to obtain a reconstructed Y channel;

[0129] A guided filtering module is used to perform guided filtering on the U channel and the V channel through the denoised Y channel to obtain a reconstructed U channel and a reconstructed V channel;

[0130] The RGB conversion module is used to perform RGB conversion on the reconstructed Y channel, the reconstructed U channel and the reconstructed V channel to obtain a valid image, and obtain a valid image set based on all the valid images.

[0131] In one embodiment, the detection object determination module includes:

[0132] a first convolution feature determination module, configured to sequentially substitute the first image into the first convolution module and the second convolution module to obtain a first convolution feature; the first image is any one in the valid image set;

[0133] A first enhanced feature determination module, configured to substitute the first convolution image into a preset feature enhancement module to obtain a first enhanced feature;

[0134] A second enhanced feature determination module is configured to substitute the first enhanced feature into the third convolution module to obtain a second convolution feature, and substitute the second convolution feature into the preset feature enhancement module to obtain a second enhanced feature;

[0135] A first pooling feature determination module is configured to substitute the second enhanced feature into the third convolution module to obtain a third convolution feature, substitute the second convolution feature into the preset feature enhancement module to obtain a third enhanced feature, and substitute the third enhanced feature into the pooling module to obtain the first pooling feature;

[0136] A first fusion feature determination module is used to substitute the first enhanced feature and the first pooled feature into the attention enhancement module to obtain a first attention feature, and substitute the first attention feature into the C2f module to obtain a first fusion feature;

[0137] A second fusion feature determination module is used to substitute the first fusion feature and the second enhancement feature into the attention enhancement module to obtain a second attention feature, and substitute the second attention feature into the C2f module to obtain a second fusion feature;

[0138] The detection object generation module is used to substitute the second fusion feature into the detection head to obtain the detection object.

[0139] In one embodiment, the attention enhancement module includes a first input module, a second input module, and a first output module;

[0140] A first splicing feature determination module is configured to receive a first input feature from a first input module, receive a second input feature from a second input module, and splice the first input feature and the second input feature to obtain a first splicing feature;

[0141] An output splicing feature determination module, configured to substitute the first splicing feature into the first target convolution module to obtain an output splicing feature;

[0142] An attention feature output module, configured to substitute the output concatenated features into the first attention enhancement module and the second attention enhancement module respectively to obtain a first attention output feature and a second attention output feature;

[0143] A fusion weight determination module is used to multiply the first attention output feature and the second attention output feature element by element and normalize them through the Sigmoid function to obtain a fusion weight;

[0144] A first attention fusion feature determination module is used to perform feature fusion on the first attention output feature and the second input feature according to the fusion weight to obtain the first attention fusion feature;

[0145] A second attention fusion feature determination module is used to perform feature fusion on the second attention output feature and the first input feature according to the fusion weight to obtain a second attention fusion feature;

[0146] The output feature determination module is used to perform feature fusion on the first attention fusion feature and the second attention fusion feature to obtain an output feature, and the output feature is output from the first output module.

[0147] In one embodiment, the abnormal behavior detection module includes:

[0148] The object association module is used to associate the detection objects corresponding to each image in the valid image set using the Hungarian algorithm, create an independent ID for each object that appears, and establish trajectory changes;

[0149] The abnormal behavior determination module is used to extract the Fourier descriptor of the trajectory corresponding to each ID, and determine the abnormal behavior corresponding to the ID based on the Fourier descriptor and preset rules.

[0150] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. An intelligent security method based on machine vision, characterized in that: The method comprises: Collecting video data within a target area, and performing frame extraction processing on the video data to obtain an image recognition set; Performing denoising on the images in the image recognition set to obtain a valid image set; Substituting each image in the valid image set into a preset model to obtain the corresponding detection object in each image; According to the detection object corresponding to each image in the effective image set, a change trajectory of the detection object in a time series is established, abnormal behavior detection is performed based on the change trajectory, and an alarm is issued if abnormal behavior is detected.

2. The intelligent security method based on machine vision according to claim 1, characterized in that: Performing denoising on the images in the image recognition set to obtain a valid image set includes: For each image in the image recognition set, convert the image into YUV space to obtain a Y channel, a U channel, and a V channel; Performing dual-tree complex wavelet transform to denoise the Y channel to obtain a denoised Y channel, and reconstructing the denoised Y channel through inverse dual-tree complex wavelet transform to obtain a reconstructed Y channel; Perform guided filtering on the U channel and the V channel through the denoised Y channel to obtain the reconstructed U channel and the reconstructed V channel; Perform RGB conversion on the reconstructed Y channel, the reconstructed U channel, and the reconstructed V channel to obtain valid images, and obtain a valid image set based on all valid images.

3. The intelligent security method based on machine vision according to claim 1, characterized in that: Substituting each image in the valid image set into the preset model to obtain the corresponding detection object in each image includes: Substituting the first image into the first convolution module and the second convolution module in sequence to obtain a first convolution feature; the first image is any one of the valid image set; Substituting the first convolution image into a preset feature enhancement module to obtain a first enhanced feature; Substituting the first enhanced feature into the third convolution module to obtain a second convolution feature, and substituting the second convolution feature into the preset feature enhancement module to obtain a second enhanced feature; Substituting the second enhanced feature into the third convolution module to obtain a third convolution feature, substituting the second convolution feature into the preset feature enhancement module to obtain a third enhanced feature, and substituting the third enhanced feature into the pooling module to obtain a first pooling feature; Substitute the first enhanced feature and the first pooled feature into the attention enhancement module to obtain a first attention feature, and substitute the first attention feature into the C2f module to obtain a first fusion feature; Substituting the first fused feature and the second enhanced feature into the attention enhancement module to obtain a second attention feature, and substituting the second attention feature into the C2f module to obtain a second fused feature; Substitute the second fusion feature into the detection head to obtain the detection object.

4. The intelligent security method based on machine vision according to claim 3, characterized in that: The attention enhancement module includes a first input module, a second input module and a first output module; The first input module receives a first input feature, the second input module receives a second input feature, and concatenates the first input feature and the second input feature to obtain a first concatenated feature; Substituting the first splicing feature into the first target convolution module to obtain an output splicing feature; Substituting the output splicing features into the first attention enhancement module and the second attention enhancement module respectively to obtain the first attention output feature and the second attention output feature; Multiply the first attention output feature and the second attention output feature element-wise, and normalize them using the Sigmoid function to obtain a fusion weight; Performing feature fusion on the first attention output feature and the second input feature according to the fusion weight to obtain a first attention fusion feature; Performing feature fusion on the second attention output feature and the first input feature according to the fusion weight to obtain a second attention fusion feature; The first attention fusion feature and the second attention fusion feature are fused to obtain an output feature, and the output feature is output from the first output module.

5. The intelligent security method based on machine vision according to claim 1, characterized in that: Establishing a change trajectory of the detection object in a time series and performing abnormal behavior detection based on the change trajectory includes: The detection object corresponding to each image in the valid image set is associated with the object by using the Hungarian algorithm, an independent ID is created for each object that appears, and a trajectory change is established; For each ID, the Fourier descriptor of the trajectory corresponding to the ID is extracted, and the abnormal behavior corresponding to the ID is determined based on the Fourier descriptor and preset rules.

6. An intelligent security system based on machine vision, characterized in that: The system comprises: A frame extraction processing module is used to collect video data in the target area and perform frame extraction processing on the video data to obtain an image recognition set; A denoising module, configured to perform denoising on the images in the image recognition set to obtain a valid image set; A detection object determination module, configured to substitute each image in the valid image set into a preset model to obtain a detection object corresponding to each image; The abnormal behavior detection module is used to establish a change trajectory of the detection object in a time series based on the detection object corresponding to each image in the valid image set, perform abnormal behavior detection based on the change trajectory, and issue an alarm if abnormal behavior is detected.

7. The intelligent security system based on machine vision according to claim 6, characterized in that: The denoising module includes: An image conversion module is used to convert each image in the image recognition set into a YUV space to obtain a Y channel, a U channel, and a V channel; A Y channel reconstruction module is used to perform dual-tree complex wavelet transform denoising on the Y channel to obtain a denoised Y channel, and reconstruct the denoised Y channel by inverse dual-tree complex wavelet transform to obtain a reconstructed Y channel; A guided filtering module is used to perform guided filtering on the U channel and the V channel through the denoised Y channel to obtain a reconstructed U channel and a reconstructed V channel; The RGB conversion module is used to perform RGB conversion on the reconstructed Y channel, the reconstructed U channel and the reconstructed V channel to obtain a valid image, and obtain a valid image set based on all the valid images.

8. The intelligent security system based on machine vision according to claim 6, characterized in that: The detection object determination module includes: a first convolution feature determination module, configured to sequentially substitute a first image into a first convolution module and a second convolution module to obtain a first convolution feature; the first image being any one of the valid image set; a first enhanced feature determination module, configured to substitute the first convolution image into a preset feature enhancement module to obtain a first enhanced feature; a second enhanced feature determination module, configured to substitute the first enhanced feature into a third convolution module to obtain a second convolution feature, and substitute the second convolution feature into the preset feature enhancement module to obtain a second enhanced feature; a first pooling feature determination module, configured to substitute the second enhanced feature into a third convolution module to obtain a third convolution feature, substitute the second convolution feature into the preset feature enhancement module to obtain a third enhanced feature, and substitute the third enhanced feature into a pooling module to obtain a first pooling feature; A first fusion feature determination module is configured to substitute the first enhanced feature and the first pooled feature into an attention enhancement module to obtain a first attention feature, and substitute the first attention feature into a C2f module to obtain a first fusion feature; a second fusion feature determination module, configured to substitute the first fusion feature and the second enhancement feature into the attention enhancement module to obtain a second attention feature, and substitute the second attention feature into the C2f module to obtain a second fusion feature; The detection object generation module is used to substitute the second fusion feature into the detection head to obtain the detection object.

9. The intelligent security system based on machine vision according to claim 8, characterized in that: The attention enhancement module includes a first input module, a second input module and a first output module; A first splicing feature determination module is configured to receive a first input feature from the first input module, receive a second input feature from the second input module, and splice the first input feature and the second input feature to obtain a first splicing feature; an output splicing feature determination module, configured to substitute the first splicing feature into a first target convolution module to obtain an output splicing feature; An attention feature output module, configured to substitute the output splicing feature into the first attention enhancement module and the second attention enhancement module respectively to obtain a first attention output feature and a second attention output feature; a fusion weight determination module, configured to multiply the first attention output feature and the second attention output feature element-wise and normalize them using a Sigmoid function to obtain a fusion weight; a first attention fusion feature determination module, configured to perform feature fusion on the first attention output feature and the second input feature according to the fusion weight to obtain a first attention fusion feature; A second attention fusion feature determination module is used to perform feature fusion on the second attention output feature and the first input feature according to the fusion weight to obtain a second attention fusion feature; An output feature determination module is used to perform feature fusion on the first attention fusion feature and the second attention fusion feature to obtain an output feature, and the output feature is output from the first output module.

10. The intelligent security system based on machine vision according to claim 6, characterized in that: The abnormal behavior detection module includes: An object association module is used to associate the detection objects corresponding to each image in the valid image set using the Hungarian algorithm, create an independent ID for each object that appears, and establish a trajectory change; The abnormal behavior determination module is used to extract the Fourier descriptor of the trajectory corresponding to each ID, and determine the abnormal behavior corresponding to the ID based on the Fourier descriptor and preset rules.

Images