Scene object detection method based on deep learning fusion of Gaussian mixture model

Abstract

The present invention discloses a deep learning fusion scene target detection method based on a mixed Gaussian model, the steps of which are as follows: using a mixed Gaussian model to continuously adapt a surveillance video frame sequence of a real-time operation of an airport scene to obtain a static scene background frame image; using RGB to YUV color space conversion to filter out the shadow portion in the static scene foreground image and separate the moving target; making the mixed Gaussian model's sensitivity to scene moving target detection adaptable to the real-time video surveillance screen of its specific application; judging whether the YOLO benchmark detection model meets the index requirements of detection recall rate and detection accuracy; if not, continuing to judge until the index requirements are met; and using a fusion target detection algorithm to obtain the final detection result. The present invention combines a moving target detection algorithm with a deep learning method to improve the accuracy and robustness of airport scene target recognition, so as to achieve the purpose of accurately acquiring scene targets and conducting conflict warning management.

Description

Technical Field

[0001] The present invention belongs to the technical field of airport scene target detection and recognition, and specifically relates to a scene target detection method based on deep learning fusion of a mixed Gaussian model. Background Art

[0002] With the booming domestic civil aviation market, airports are experiencing increasing traffic loads and a growing variety of targets on the ground, placing higher demands on ground operational safety and conflict warning. Detecting and identifying targets on the ground is essential for ensuring operational safety and preventing conflicts.

[0003] In recent years, video-based target detection methods have rapidly developed. These methods utilize optical sensors to capture real-time video sequences of the airfield and identify moving targets within them, enabling conflict warning and ensuring operational safety. Compared to traditional methods of active surveillance of targets on the airfield, video-based target detection methods are highly efficient, fast, and cost-effective. Airport target recognition requires highly robust detection algorithms that accurately identify all types of unusual, dynamic, and static targets on the airfield.

[0004] Existing technologies for extracting moving objects in images fall into two main categories: one is based on traditional computer vision methods, such as frame subtraction, optical flow, and background subtraction. These methods, based on statistics, identify background and moving foreground elements in image sequences or video frames, offering high recognition rates and the ability to identify objects regardless of their type. The other is based on deep learning, which uses image annotation, dataset creation, and training to generate efficient and accurate object detection models.

[0005] The above-mentioned existing technologies for extracting moving objects in images have the following defects:

[0006] Traditional computer vision methods are usually greatly affected by changes in lighting and shadows, and have a high rate of false detection; while deep learning-based methods can only detect target types in closed sets and cannot detect category targets that do not appear in the dataset. Summary of the Invention

[0007] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a deep learning fusion scene target detection method based on a mixed Gaussian model, combining the motion target detection algorithm with the deep learning method to improve the accuracy and robustness of airport scene target recognition, so as to achieve the purpose of accurately acquiring scene targets and conducting conflict warning management.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] The present invention provides a method for scene target detection based on deep learning fusion of mixed Gaussian models, comprising the following steps:

[0010] 1) Using a mixed Gaussian model to continuously adapt the real-time surveillance video frame sequence of the airport scene to obtain static scene background frame images;

[0011] 2) Using RGB to YUV color space conversion, the shadow part of the foreground image of the static scene is filtered out by setting a screening method to separate the moving target;

[0012] 3) According to the motion fitting relationship between the local brightness channel and the scene moving target in the YUV color space, the frame difference interval n of the mixed Gaussian model is set so that the sensitivity of the mixed Gaussian model to the scene moving target detection can adapt to the real-time video surveillance screen of its specific application;

[0013] 4) Using a mixed Gaussian model, detect the real-time surveillance video, detect a set of moving targets in the scene, add annotation information, and use this as a basic training set based on the deep learning method YOLO to train a YOLO benchmark detection model. Determine whether the YOLO benchmark detection model meets the detection recall rate and detection accuracy index requirements; if not, repeat step 4) until it meets the index requirements;

[0014] 5) Use the fusion target detection algorithm to obtain the final detection result.

[0015] Furthermore, step 1) specifically includes: using a Gaussian mixture model starting from the first image of the airport scene monitoring video frame sequence as the background, and continuously adapting subsequent images to obtain an accurate static scene background frame image; the Gaussian mixture model is defined as follows:

[0016]

[0017] in, represents the probability density function of the random variable x vector, X T is the image matrix, T is the period, BG is the background, and FG is the foreground; Represents the weight of the mth Gaussian distribution, which satisfies the sum of 1; represents the mth random variable x vector obeys the Gaussian distribution, and its mean is The standard deviation is I represents the identity matrix with diagonal 1; the parameter algorithm in period T is as follows:

[0018]

[0019] in, α is the update coefficient, is the update rate, i.e. 1 / T is updated per cycle; For a new sample, the initial value of m is 1. Start from the first image as the background and adapt to subsequent images. The steps are as follows:

[0020] 11) Read video frame: read each frame from the video as input data;

[0021] 12) Establishing a Gaussian mixture model: For each pixel, the values ​​of the three RGB channels are used as a vector and the vector is modeled using a Gaussian mixture model. During the modeling process, the number of Gaussian distributions in the Gaussian mixture model is set, and the mean, variance, and weight of each distribution are initialized.

[0022] 13) Update the Gaussian mixture model: For each pixel in the current frame, calculate its distance to each Gaussian distribution and select the distribution with the smallest distance as its category; based on the category of the current pixel, update the mean, variance, and weight of the corresponding Gaussian distribution;

[0023] 14) Static background screening: The static background of the video is screened out by determining whether the category of each pixel is stable. A threshold is set, and when the category of a pixel remains unchanged in multiple consecutive frames, it is considered to be a static background.

[0024] 15) Generate background image: For all the static background pixels that are screened out, use the average or median of their values ​​in the previous frames as the pixel value of the background image.

[0025] Furthermore, step 2) specifically includes: converting the surveillance video frame of the static scene from RGB space to YUV space, and analyzing the shadow area to determine if it has the following characteristics: a. the brightness of the pixels in the shadow area is lower than the brightness of the background pixels and the target pixel area; b. the chromaticity of the shadow area is equal to the chromaticity of the background pixels; c. the gradient difference between the boundary between the pixels in the shadow area and the background area is small, while the gradient difference between the boundary between the target area and the background area is large;

[0026] According to the above three features a, b, and c, a static background image is generated by separation using a mixed Gaussian model, namely the BG component in formula (1); the difference between the current frame and the background frame is calculated:

[0027] FG=X T -BG (3)

[0028] Get the foreground image FG, calculate the brightness channel, chroma channel and gradient difference in the foreground image, and compare them with the thresholds of the shadow area and non-shadow area to determine the shadow area. Combined with edge detection, the shadow area is eliminated.

[0029] Furthermore, the step 2) is described in detail as the following steps:

[0030] 21) RGB to YUV conversion: Convert the RGB value of each pixel to the corresponding YUV value, where Y represents brightness and U and V represent chrominance;

[0031] 22) Filtering shadows: Based on the scene characteristics and requirements, set a threshold to filter out the shadows in the image; use a threshold-based binarization method to convert the image into black and white, and then process the image through morphological operations to mark the shadows as white and the non-shadows as black;

[0032] 23) Filtering: Set the values ​​of all pixels marked as shadows in the foreground image to the background color or median value, thereby removing the shadows;

[0033] 24) Convert to RGB: Convert the processed YUV image to an RGB image to obtain a foreground image without shadows.

[0034] Furthermore, the step 3) specifically includes: assuming that the frame difference interval used by the mixed Gaussian model is n, n>2, and the specific numerical value adopts an adaptive algorithm, specifically: converting the image from RGB space to YUV space, performing preliminary grid division on the operation scene image, screening out the operation area and the non-operation area, and completing grid convergence and merging according to the operation history data; further performing brightness histogram statistics on the local and regional brightness channels, analyzing the brightness change law caused by the historical target motion, fitting the motion-brightness distribution histogram curve, calculating the mixed Gaussian model frame difference interval n according to the fitting curve, and filtering out false detections caused by local small displacements.

[0035] Furthermore, the step 4) specifically includes:

[0036] 41) Use a mixed Gaussian model to detect a set of moving targets in real-time video surveillance scene images;

[0037] 42) Add annotation information and save the moving target set and the corresponding real annotation box as the basic training set;

[0038] 43) Use the YOLO model based on deep learning method to train the basic training set to obtain the YOLO benchmark detection model;

[0039] 44) Test the YOLO benchmark detection model and evaluate its performance on a set of test datasets, including recall, accuracy, and mAP; if any indicators do not meet the requirements, make adjustments;

[0040] 45) If the YOLO benchmark detection model does not meet the requirements, improve the detection performance by increasing the amount of training data, adjusting model parameters, and changing the data augmentation method; continue training and testing until the requirements are met.

[0041] Furthermore, the step 5) specifically includes: performing target detection on the real-time video surveillance image using the YOLO benchmark detection model and the mixed Gaussian model respectively, fusing the two detection results after obtaining them, and classifying the fusion results; performing corresponding processing based on the classification results to improve the detection accuracy and robustness of the fusion target detection algorithm for scene targets.

[0042] Furthermore, the step 5) specifically includes:

[0043] 51) Use the YOLO benchmark detection model and the mixed Gaussian model to detect moving targets in real-time video surveillance images, and obtain two detection results;

[0044] 52) Fusing the two detection results; using weighted average or voting methods to fuse;

[0045] 53) Classify the fusion results and divide the detected targets into true targets and false positive targets; classify the targets based on their size, shape, color, position and motion information;

[0046] 54) For false positive targets, further processing is performed by introducing scene background modeling or negative sample mining to reduce the false detection rate;

[0047] 55) Further processing is performed on real targets based on actual needs, including target tracking and behavior analysis;

[0048] 56) Evaluate the detection accuracy and robustness of the fusion results, including recall rate, precision rate, F value, and mAP; based on the evaluation results, gradually adjust the weights and thresholds of the YOLO benchmark detection model and the mixed Gaussian model in the fusion target detection algorithm to improve the performance and robustness of the fusion target detection algorithm.

[0049] Furthermore, the fusion result includes:

[0050] Repeated targets: targets detected by both the Gaussian mixture model and the YOLO benchmark detection model. According to the Gaussian mixture model principle, this target is a moving target.

[0051] Targets detected by the Gaussian mixture model but not by the YOLO benchmark detection model: These targets are also moving targets and are not present in the YOLO benchmark detection model training dataset or are insufficiently trained.

[0052] Targets detected by the YOLO benchmark detection model but not detected by the mixed Gaussian model: This target is a stationary target, which is further subdivided into: stationary targets misdetected by the YOLO benchmark detection model; stationary targets correctly detected by the YOLO benchmark detection model.

[0053] Beneficial effects of the present invention:

[0054] The present invention uses a deep learning fusion target detection algorithm based on a mixed Gaussian model to perform target detection and extraction on panoramic videos, which can realize dynamic and static target and open set target detection. The fusion target detection algorithm has the characteristics of good robustness, strong adaptability, and can be optimized and upgraded according to work history data.

[0055] By using background modeling based on a mixture Gaussian model, the foreground object and background in the video can be effectively separated, thereby improving the accuracy and robustness of object detection.

[0056] By integrating deep learning with target detection algorithms, we can simultaneously detect both static and dynamic targets and open sets of targets, improving the accuracy and applicability of target detection. Meanwhile, tracking-based methods can improve the continuity and stability of target detection.

[0057] The accuracy and reliability of target detection can be further improved through post-processing, including target classification. At the same time, the algorithm can be optimized and upgraded based on historical working data to improve its adaptability and iteration capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 Schematic diagram of the method of the present invention. DETAILED DESCRIPTION

[0059] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and drawings. The contents mentioned in the embodiments are not intended to limit the present invention.

[0060] Reference Figure 1 As shown, the present invention is a deep learning fusion scene target detection method based on a mixed Gaussian model, the steps are as follows:

[0061] 1) A Gaussian mixture model (GMM) is used to continuously adapt a sequence of real-time surveillance video frames of the airport scene to obtain a static scene background frame image. Specifically, the Gaussian mixture model (GMM) is used to continuously adapt subsequent images starting from the first image in the sequence of airport scene surveillance video frames to obtain an accurate static scene background frame image. The Gaussian mixture model is defined as follows:

[0062]

[0063] in, represents the probability density function of the random variable x vector, X Tis the image matrix, T is the period, BG is the background, and FG is the foreground; Represents the weight of the mth Gaussian distribution, which satisfies the sum of 1; N(x) represents the mth random variable x vector obeys the Gaussian distribution (normal distribution), and its mean is The standard deviation is I represents the identity matrix with diagonal 1; the parameter algorithm in period T is as follows:

[0064]

[0065] in, α is the update coefficient, is the update rate, i.e. 1 / T is updated per cycle; For a new sample, the initial value of m is 1. Start from the first image as the background and adapt to subsequent images. The steps are as follows:

[0066] 11) Read video frames: Read each frame from the video as input data (implemented using computer vision libraries such as OpenCV);

[0067] 12) Establishing a Gaussian mixture model: For each pixel, the values ​​of the three RGB channels are used as a vector and the vector is modeled using a Gaussian mixture model. During the modeling process, the number of Gaussian distributions in the Gaussian mixture model is set (usually 3-5), and the mean, variance, and weight of each distribution are initialized.

[0068] 13) Update the Gaussian mixture model: For each pixel in the current frame, calculate its distance to each Gaussian distribution and select the distribution with the smallest distance as its category; based on the category of the current pixel, update the mean, variance, and weight of the corresponding Gaussian distribution;

[0069] 14) Static background screening: The static background of the video is screened out by determining whether the category of each pixel is stable. A threshold is set, and when the category of a pixel remains unchanged in multiple consecutive frames, it is considered to be a static background.

[0070] 15) Generate background image: For all the static background pixels that are screened out, use the average or median of their values ​​in the previous frames as the pixel value of the background image.

[0071] 2) Using RGB to YUV color space conversion, a screening method is set to filter out shadows in the foreground image of a static scene and isolate moving targets. Specifically, the method involves converting surveillance video frames of the static scene from RGB space to YUV space and analyzing the shadow areas to see if they have the following characteristics: a. The brightness of the pixels in the shadow area is lower than that of the background pixels and the target pixel area; b. The chromaticity of the shadow area is equal to the background pixel color (i.e., the local shadow area and background color are equal); c. The gradient difference between the boundary between the pixels in the shadow area and the background area is small, while the gradient difference between the boundary between the target area and the background area is large.

[0072] According to the above three features a, b, and c, a static background image is generated by separation using a mixed Gaussian model, namely the BG component in formula (1); the difference between the current frame and the background frame is calculated:

[0073] FG=X T -BG (3)

[0074] Get the foreground image FG, calculate the brightness channel, chroma channel and gradient difference in the foreground image, and compare them with the thresholds of the shadow area and non-shadow area to determine the shadow area. Combined with edge detection, the shadow area is eliminated.

[0075] The following steps are described in detail:

[0076] 21) RGB to YUV conversion: Convert the RGB value of each pixel to the corresponding YUV value, where Y represents brightness and U and V represent chrominance;

[0077] 22) Filtering shadows: Based on the scene characteristics and requirements, set a threshold to filter out the shadows in the image; use a threshold-based binarization method to convert the image into a black and white image, and then process the image through morphological operations (such as dilation and erosion) to mark the shadow parts as white and the non-shadow parts as black;

[0078] 23) Filtering: Set the values ​​of all pixels marked as shadows in the foreground image to the background color or median value, thereby removing the shadows;

[0079] 24) Convert to RGB: Convert the processed YUV image to an RGB image to obtain a foreground image without shadows.

[0080] 3) According to the motion fitting relationship between the local brightness channel and the scene moving target in the YUV color space, the frame difference interval n of the mixed Gaussian model is set so that the sensitivity of the mixed Gaussian model to the scene moving target detection can adapt to the real-time video surveillance screen of its specific application;

[0081] Since the Gaussian mixture model relies on the frame difference method to achieve foreground and background subtraction, the frame difference interval will affect the sensitivity of the GMM method. If the frame difference interval is too small, small noise changes will be mistakenly detected as moving targets, while if the frame difference interval is too large, the detection sensitivity of real moving targets will be reduced. Therefore, selecting an appropriate frame difference interval can help improve the detection robustness of the Gaussian mixture model method.

[0082] Specifically, it includes: assuming that the frame difference interval used by the mixed Gaussian model is n, n>2, and the specific value adopts an adaptive algorithm, specifically: converting the image from RGB space to YUV space, performing preliminary grid division on the operation scene image, screening out the operation area and non-operation area, and completing grid convergence and merging based on the operation history data; further performing brightness histogram statistics on the local and regional brightness channel (Y channel), analyzing the brightness change law caused by historical target motion, fitting the motion-brightness distribution histogram curve, and calculating the mixed Gaussian model frame difference interval n according to the fitting curve to filter out false detections caused by local small displacements.

[0083] 4) Using a mixed Gaussian model, detect the real-time surveillance video, detect a set of moving targets in the scene, add annotation information, and use it as a basic training set based on the deep learning method YOLO to train a YOLO benchmark detection model, and determine whether the YOLO benchmark detection model meets the requirements of detection recall rate and detection accuracy; if not, repeat step 4) until it meets the requirements; specifically including:

[0084] 41) Use a mixed Gaussian model to detect a set of moving targets in real-time video surveillance scene images;

[0085] 42) Add annotation information and save the moving target set and the corresponding real annotation box as the basic training set;

[0086] 43) Use the YOLO model based on deep learning method to train the basic training set to obtain the YOLO benchmark detection model;

[0087] 44) Test the YOLO benchmark detection model and evaluate its performance on a set of test datasets, including recall, accuracy, and mAP; if any indicators do not meet the requirements, make adjustments;

[0088] 45) If the YOLO benchmark detection model does not meet the requirements, improve the detection performance by increasing the amount of training data, adjusting model parameters, and changing the data augmentation method; continue training and testing until the requirements are met.

[0089] 5) Use the fusion target detection algorithm to obtain the final detection result; specifically, the YOLO benchmark detection model and the mixed Gaussian model are used to perform target detection on the real-time video surveillance image, and the two detection results are fused and classified. According to the classification results, corresponding processing is performed to improve the detection accuracy and robustness of the fusion target detection algorithm for scene targets.

[0090] 51) Use the YOLO benchmark detection model and the mixed Gaussian model to detect moving targets in real-time video surveillance images, and obtain two detection results;

[0091] 52) Fusing the two detection results; using weighted average or voting methods to fuse;

[0092] 53) Classify the fusion results and divide the detected targets into true targets and false positive targets; classify the targets based on their size, shape, color, position and motion information (such as the position and motion of scene objects and people);

[0093] 54) For false positive targets, further processing is performed by introducing scene background modeling or negative sample mining to reduce the false detection rate;

[0094] 55) Further processing is performed on real targets based on actual needs, including target tracking and behavior analysis;

[0095] 56) Evaluate the detection accuracy and robustness of the fusion results, including recall rate, precision rate, F value, and mAP; based on the evaluation results, gradually adjust the weights and thresholds of the YOLO benchmark detection model and the mixed Gaussian model in the fusion target detection algorithm to improve the performance and robustness of the fusion target detection algorithm.

[0096] Furthermore, at each iteration, the fusion results are observed and classified; then, the misclassified samples are further analyzed, the algorithm is retested and evaluated, and optimization is continued until the desired performance requirements are met.

[0097] The fusion result includes:

[0098] Repeated targets: targets detected by both the Gaussian mixture model and the YOLO benchmark detection model. According to the Gaussian mixture model principle, this target is a moving target.

[0099] Targets detected by the Gaussian mixture model but not by the YOLO benchmark detection model: These targets are also moving targets and are not present in the YOLO benchmark detection model training dataset or are insufficiently trained.

[0100] Targets detected by the YOLO benchmark detection model but not detected by the mixed Gaussian model: This target is a stationary target, which is further subdivided into: stationary targets misdetected by the YOLO benchmark detection model; stationary targets correctly detected by the YOLO benchmark detection model.

[0101] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be considered as the scope of protection of the present invention.

Claims

1. A deep learning fusion scene target detection method based on a mixture Gaussian model, characterized in that: Here are the steps: 1) Using a mixed Gaussian model to continuously adapt the real-time surveillance video frame sequence of the airport scene to obtain static scene background frame images; 2) Using RGB to YUV color space conversion, the shadow part of the foreground image of the static scene is filtered out by setting a screening method to separate the moving target; 3) According to the motion fitting relationship between the local brightness channel and the scene moving target in the YUV color space, the frame difference interval n of the mixed Gaussian model is set so that the sensitivity of the mixed Gaussian model to the scene moving target detection can adapt to the real-time video surveillance screen of its specific application; 4) Using a mixed Gaussian model, detect the real-time surveillance video, detect a set of moving targets in the scene, add annotation information, and use this as a basic training set based on the deep learning method YOLO to train a YOLO benchmark detection model. Determine whether the YOLO benchmark detection model meets the detection recall rate and detection accuracy index requirements; if not, repeat step 4) until it meets the index requirements; 5) Use the fusion target detection algorithm to obtain the final detection result; The step 2) specifically includes: converting the surveillance video frame of the static scene from RGB space to YUV space, and analyzing the shadow area to determine whether it has the following characteristics: a. the brightness of the pixels in the shadow area is lower than the brightness of the background pixels and the target pixel area; b. the chromaticity of the shadow area is equal to the background pixel color; c. the gradient difference between the boundary between the pixels in the shadow area and the background area is small, while the gradient difference between the boundary between the target area and the background area is large; Based on the above three features a, b, and c, a static background image, namely the BG component, is generated by separation using a mixed Gaussian model; the difference between the current frame and the background frame is calculated: FG=X T -BG (3) Get the foreground image FG, calculate the brightness channel, chroma channel and gradient difference in the foreground image, compare it with the threshold of the shadow area and non-shadow area, determine the shadow area, combine it with edge detection, and eliminate the shadow area; The step 3) specifically includes: assuming that the frame difference interval used by the mixed Gaussian model is n, where n>2, and an adaptive algorithm is used for the specific value, specifically: converting the image from RGB space to YUV space, performing preliminary grid division on the operation scene image, screening out the operation area and the non-operation area, and completing grid convergence and merging based on the operation history data; further performing brightness histogram statistics on the local, regional brightness channels, analyzing the brightness change pattern caused by the historical target motion, fitting the motion-brightness distribution histogram curve, calculating the mixed Gaussian model frame difference interval n based on the fitted curve, and filtering out false detections caused by local small displacements; The step 4) specifically includes: 41) Use a mixed Gaussian model to detect a set of moving targets in real-time video surveillance scene images; 42) Add annotation information and save the moving target set and the corresponding real annotation box as the basic training set; 43) Use the YOLO model based on deep learning method to train the basic training set to obtain the YOLO benchmark detection model; 44) Test the YOLO benchmark detection model and evaluate its performance on a set of test datasets, including recall, accuracy, and mAP; if any indicators do not meet the requirements, make adjustments; 45) If the YOLO benchmark detection model does not meet the requirements, improve the detection performance by increasing the amount of training data, adjusting model parameters, and changing the data augmentation method; continue training and testing until the requirements are met.

2. The method for scene target detection based on deep learning fusion of mixed Gaussian models according to claim 1 is characterized in that: The step 1) specifically includes: using a Gaussian mixture model starting from the first image of the airport scene monitoring video frame sequence as the background, and continuously adapting subsequent images to obtain an accurate static scene background frame image; the Gaussian mixture model is defined as follows: in, represents the probability density function of the random variable x vector, X T is the image matrix, T is the period, BG is the background, and FG is the foreground; represents the weight of the mth Gaussian distribution, which satisfies the sum of 1; N(x) represents the mth random variable x vector obeys the Gaussian distribution, and its mean is The standard deviation is I represents the identity matrix with diagonal 1; the parameter algorithm in period T is as follows: in, α is the update coefficient, is the update rate, i.e. 1 / T is updated per cycle; For a new sample, the initial value of m is 1. Start from the first image as the background and adapt to subsequent images. The steps are as follows: 11) Read video frame: read each frame from the video as input data; 12) Establishing a Gaussian mixture model: For each pixel, the values ​​of the three RGB channels are used as a vector and the vector is modeled using a Gaussian mixture model. During the modeling process, the number of Gaussian distributions in the Gaussian mixture model is set, and the mean, variance, and weight of each distribution are initialized. 13) Update the Gaussian mixture model: For each pixel in the current frame, calculate its distance to each Gaussian distribution and select the distribution with the smallest distance as its category; based on the category of the current pixel, update the mean, variance, and weight of the corresponding Gaussian distribution; 14) Static background screening: The static background of the video is screened out by determining whether the category of each pixel is stable. A threshold is set, and when the category of a pixel remains unchanged in multiple consecutive frames, it is considered to be a static background. 15) Generate background image: For all the static background pixels that are screened out, use the average or median of their values ​​in the previous frames as the pixel value of the background image.

3. The method for scene target detection based on deep learning fusion of mixed Gaussian models according to claim 1, characterized in that: The step 2) is described in detail as the following steps: 21) RGB to YUV conversion: Convert the RGB value of each pixel to the corresponding YUV value, where Y represents brightness and U and V represent chrominance; 22) Filtering shadows: Based on the scene characteristics and requirements, set a threshold to filter out the shadows in the image; use a threshold-based binarization method to convert the image into black and white, and then process the image through morphological operations to mark the shadows as white and the non-shadows as black; 23) Filtering: Set the values ​​of all pixels marked as shadows in the foreground image to the background color or median value, thereby removing the shadows; 24) Convert to RGB: Convert the processed YUV image to an RGB image to obtain a foreground image without shadows.

4. The method for scene target detection based on deep learning fusion of mixed Gaussian models according to claim 1, characterized in that: The step 5) specifically includes: performing target detection on the real-time video surveillance image using the YOLO benchmark detection model and the mixed Gaussian model respectively, fusing the two detection results after obtaining them, and classifying the fusion results; performing corresponding processing based on the classification results to improve the detection accuracy and robustness of the fusion target detection algorithm for scene targets.

5. The method for scene target detection based on deep learning fusion of mixed Gaussian models according to claim 1, characterized in that: The step 5) specifically includes: 51) Use the YOLO benchmark detection model and the mixed Gaussian model to detect moving targets in real-time video surveillance images, and obtain two detection results; 52) Fusing the two detection results; using weighted average or voting methods to fuse; 53) Classify the fusion results and divide the detected targets into true targets and false positive targets; classify the targets based on their size, shape, color, position and motion information; 54) For false positive targets, further processing is performed by introducing scene background modeling or negative sample mining to reduce the false detection rate; 55) Further processing is performed on real targets based on actual needs, including target tracking and behavior analysis; 56) Evaluate the detection accuracy and robustness of the fusion results, including recall rate, precision rate, F value, and mAP; based on the evaluation results, gradually adjust the weights and thresholds of the YOLO benchmark detection model and the mixed Gaussian model in the fusion target detection algorithm to improve the performance and robustness of the fusion target detection algorithm.

6. The method for scene target detection based on deep learning fusion of mixed Gaussian models according to claim 5, characterized in that: The fusion results include: Repeated targets: targets detected by both the Gaussian mixture model and the YOLO benchmark detection model. According to the Gaussian mixture model principle, this target is a moving target. Targets detected by the Gaussian mixture model but not by the YOLO benchmark detection model: These targets are also moving targets and are not present in the YOLO benchmark detection model training dataset or are insufficiently trained. Targets detected by the YOLO benchmark detection model but not detected by the mixed Gaussian model: This target is a stationary target, which is further subdivided into: stationary targets misdetected by the YOLO benchmark detection model; stationary targets correctly detected by the YOLO benchmark detection model.

Images