Deep learning-based real-time monitoring method for parking lot space status

Abstract

The present application relates to the technical field of image processing, more particularly, the present application relates to a parking lot space state real-time monitoring method based on deep learning, which comprises inputting each frame of monitoring image of the parking lot into a pre-trained neural network, identifying each parking space area in each frame of monitoring image and the original occupancy probability of each parking space area; obtaining the historical environmental turbulence degree of a target parking space, combining the original occupancy probability of the target parking space with the difference of its occupancy state in the last frame of monitoring image, calculating the state inertia factor of the target parking space, determining the spatial topological constraint deviation of the target parking space based on the distance between the target parking space and each neighboring parking space and the original occupancy probability deviation; using the state inertia factor and the spatial topological constraint deviation to correct the original occupancy probability to obtain the final occupancy probability for determining the occupancy state of the target parking space, and the present application improves the accuracy of parking space occupancy state monitoring.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to a deep learning-based method for real-time monitoring of parking space status. Background Technology

[0002] With the continuous growth of urban car ownership, automated parking management has become a key aspect of alleviating traffic congestion and improving parking efficiency. In intelligent parking systems, computer vision-based parking space detection technology has become the mainstream application.

[0003] Existing parking space status monitoring technologies generally adopt mainstream target detection models such as the YOLO series. They collect images of the parking space area through monitoring cameras deployed in parking lots, perform real-time inference calculations on the image frames, output the probability of parking space occupancy, and determine whether the parking space is occupied or vacant based on the probability, thereby achieving automated identification of parking space status.

[0004] However, in the actual application of real-time dynamic monitoring in parking lots, single-frame video images are prone to environmental noise interference such as pedestrians crossing and changes in vehicle lighting. Such interference directly destroys the feature integrity of the parking space area, causing the single-frame parking space occupancy probability output by the target detection model to fluctuate drastically at the judgment threshold boundary. This leads to irregular and frequent jumps in parking space status between idle and occupied. Such state jumps not only cause confusion in the billing logic of the backend automatic billing system (such as duplicate billing, missed billing, billing delay, etc.), but also significantly increase the concurrent read and write pressure of the backend database, reducing the overall operational stability of the system.

[0005] Furthermore, the parking spaces in the parking lot are closely distributed and have inherent physical connections. When large vehicles are parked in adjacent spaces, or when vehicles are parked at an angle (crossing or pressing the lines), the image features of the target parking space will be obscured or interfered with due to the perspective of the surveillance camera, causing the target detection algorithm to make false or false detections of the status of that parking space. Summary of the Invention

[0006] To address the issues of frequent changes in parking space status due to environmental noise causing billing confusion and system instability in real-time parking lot monitoring, and the problems of densely distributed parking spaces and interference from adjacent vehicles leading to false positives and false negatives in the algorithm, this invention proposes a deep learning-based real-time parking space status monitoring method. This method includes the following steps:

[0007] Acquire several frames of monitoring images of the parking lot, input each frame of monitoring image into a pre-trained neural network, and identify each parking space in each monitoring image and the original occupancy probability of each parking space.

[0008] Take any parking space in the current frame of the monitoring image as the target parking space; based on the fluctuation of the original occupancy probability of the target parking space in multiple consecutive historical frames, obtain the historical environmental turbulence level of the target parking space; based on the historical environmental turbulence level and the difference between the original occupancy probability of the target parking space and the occupancy state of the target parking space in the previous frame of the monitoring image, obtain the state inertia factor of the target parking space.

[0009] Based on the distance between the target parking space and its neighboring parking spaces, a spatial distance attenuation coefficient is obtained; based on the spatial distance attenuation coefficient and the deviation of the original occupancy probability between the target parking space and its neighboring parking spaces, a spatial topology constraint deviation of the target parking space is obtained; the original occupancy probability of the target parking space is corrected using the state inertia factor and the spatial topology constraint deviation to obtain the final occupancy probability of the target parking space.

[0010] The occupancy status of the target parking space is determined based on the comparison result between the final occupancy probability and the preset basic judgment threshold.

[0011] The innovation of this invention lies in firstly, based on the fluctuation of the original occupancy probability of the target parking space in multiple consecutive historical frames, obtaining the historical environmental turbulence of the target parking space and introducing a state inertia factor, effectively smoothing out instantaneous noise interference such as pedestrian crossings and changes in light and shadow, and avoiding irregular and frequent jumps in the parking space status between vacancy and occupancy; furthermore, by calculating the spatial distance attenuation coefficient and probability deviation between the target parking space and its neighboring parking spaces, the spatial topological constraint deviation is obtained, which can identify and correct the abnormal increase in the original occupancy probability caused by adjacent large vehicles crossing the line or obstructing the space, thereby accurately determining the true status of the parking space.

[0012] Preferably, the acquisition of the historical environmental volatility of the target parking space includes:

[0013] ;

[0014] In the formula, This represents the historical environmental turbulence level of the t-th parking space in the current frame of the monitoring image; This represents the number of historical frames preceding the current monitored frame. This represents the original occupancy probability of the t-th parking space in the (m+1)-th historical frame image preceding the current frame monitoring image; This represents the original occupancy probability of the t-th parking space in the m-th historical frame image preceding the current frame monitoring image; Represents the time weight of the m-th historical frame image preceding the current frame image; || represents the absolute value symbol.

[0015] Based on the fluctuation of the original occupancy probability of the target parking space in multiple consecutive historical frames, the degree of historical environmental turbulence can be quantified, which can effectively suppress the phenomenon of parking space status jump caused by environmental interference noise.

[0016] Preferably, the time weight of the m-th historical frame image preceding the current frame monitoring image includes:

[0017] , This represents the number of historical frame images preceding the current frame image; m represents the index number of the historical frame image. This represents the time weight of the m-th historical frame image preceding the current frame's monitored image.

[0018] Giving higher weight to historical frames that are closer to the current monitoring frame makes the results of historical environmental turbulence more accurate.

[0019] Preferably, obtaining the state inertia factor of the target parking space includes:

[0020] , This represents the state inertia factor of the t-th parking space in the current frame of the monitoring image; This represents the original occupancy probability of the t-th parking space in the current frame of the monitoring image; This represents the degree of environmental disturbance in the t-th parking space under the current frame of the monitoring image; This represents the occupancy status value of the t-th parking space in the previous frame of the current monitoring image; This represents the preset sensitivity adjustment coefficient; || represents the minimum bias constant; || represents the absolute value sign.

[0021] Preferably, obtaining the spatial distance attenuation coefficient includes:

[0022] ;

[0023] In the formula, This represents the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image; This represents the Euclidean distance between the center point of the t-th parking space and the center point of its n-th neighboring parking spaces; This represents the width of a standard parking space.

[0024] Preferably, obtaining the spatial topology constraint deviation of the target parking space includes:

[0025] ;

[0026] In the formula, This represents the spatial topological constraint deviation of the t-th parking space in the current frame of the monitoring image. This represents the number of neighboring parking spaces of the t-th parking space; This represents the original occupancy probability of the t-th parking space in the current frame of the monitoring image; Represents the original occupancy probability of the nth neighboring parking space of the t-th parking space in the current frame of the monitoring image; || represents the absolute value sign; Represents the logarithmic function; This represents the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image.

[0027] This effectively avoids the problem of incorrect parking space status recognition caused by physical obstruction between parking spaces.

[0028] Preferably, obtaining the final occupancy probability of the target parking space includes:

[0029] ;

[0030] In the formula, This represents the final occupancy probability of the t-th parking space in the current frame of the monitoring image; This represents the original occupancy probability of the t-th parking space in the current frame of the monitoring image; This represents the state inertia factor of the t-th parking space in the current frame of the monitoring image; This represents the spatial topological constraint deviation of the t-th parking space in the current frame of the monitoring image. This represents the preset spatial correction intensity coefficient.

[0031] By reconstructing the original occupancy probability, the accuracy of parking space status determination in complex environments is significantly improved.

[0032] Preferably, obtaining the occupancy status value of the t-th parking space in the previous frame of the current frame monitoring image includes:

[0033] Preset basic judgment threshold If the original occupancy probability of the t-th parking space in the previous frame of the current monitoring image is greater than or equal to the basic judgment threshold, the occupancy status value of the t-th parking space in the previous frame of the current monitoring image is 1; otherwise, the occupancy status value of the t-th parking space in the previous frame of the current monitoring image is 0.

[0034] Preferably, determining the occupancy status of the target parking space based on the comparison result between the final occupancy probability and a preset basic determination threshold includes:

[0035] If the final occupancy probability of the t-th parking space in the current frame of the monitoring image minus the basic judgment threshold is greater than or equal to 0, the final judgment status of the t-th parking space in the current frame of the monitoring image is 1, indicating that the t-th parking space in the current frame of the monitoring image is in an occupied state; if the final occupancy probability of the t-th parking space in the current frame of the monitoring image minus the basic judgment threshold is less than 0, the final judgment status of the t-th parking space in the current frame of the monitoring image is 0, indicating that the t-th parking space in the current frame of the monitoring image is in an idle state.

[0036] Preferably, acquiring several frames of monitoring images of the parking lot includes:

[0037] The preset frame extraction step size is 200ms. The real-time monitoring video of the parking lot is obtained by using the surveillance cameras deployed in the parking lot. The FFmpeg tool is used to divide the real-time monitoring video of the parking lot into several monitoring images according to the preset frame extraction step size of 200ms.

[0038] The present invention has the following beneficial effects: The purpose of the present invention is to first obtain the historical environmental turbulence of the target parking space based on the fluctuation of the original occupancy probability of the target parking space in multiple consecutive historical frames of images, and based on the historical environmental turbulence and the difference between the original occupancy probability of the target parking space and the occupancy status of the target parking space in the previous frame of monitoring image, obtain the state inertia factor of the target parking space, effectively smooth out instantaneous noise interference such as pedestrian crossing and light and shadow changes, and avoid irregular and frequent jumps in the parking space status between idle and occupied; furthermore, by calculating the spatial distance attenuation coefficient and probability deviation between the target parking space and its neighboring parking spaces, the spatial topology constraint deviation is obtained, which can identify and correct the abnormal increase in the original occupancy probability caused by adjacent large vehicles crossing the line or obstructing the space, thereby accurately judging the true status of the parking space. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the steps of the real-time monitoring method for parking space status based on deep learning, according to an embodiment of the present invention.

[0040] Figure 2 This is a rendering of the existing parking space monitoring technology;

[0041] Figure 3 This is a diagram illustrating the parking space monitoring effect of the present invention. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0043] Please see Figure 1 The diagram illustrates a flowchart of a deep learning-based real-time parking space status monitoring method according to an embodiment of the present invention. The method includes the following steps:

[0044] S001. Acquire several frames of monitoring images of the parking lot, input each frame of monitoring image into a pre-trained neural network, and identify each parking space area in each frame of monitoring image and the original occupancy probability of each parking space area.

[0045] In this embodiment of the invention, surveillance cameras deployed in the parking lot are used to acquire real-time monitoring video of the parking lot, and FFmpeg is used to divide the real-time monitoring video of the parking lot into several frames of monitoring images according to the frame extraction step size; in this embodiment of the invention, the preset frame extraction step size is 200ms.

[0046] Each frame of the surveillance image is input into the trained neural network to obtain each parking space area and its original occupancy probability in each frame of the surveillance image. In this embodiment, the neural network used is MobileNet, and the dataset for training this neural network is obtained as follows:

[0047] A large number of surveillance images are collected, and parking spaces are manually marked in each image using bounding boxes. For any given parking space, if there are no vehicles in the area, the original occupancy probability is marked as 0; if there are complete vehicles in the area, the original occupancy probability is marked as 1; if there are incomplete vehicles in the area due to reasons such as vehicles crossing lines, obstructed views, or pedestrians crossing, the original occupancy probability is marked as 0.5. This marking result is recorded as the label for each surveillance image. A large number of surveillance images and their corresponding labels are collected to form a dataset. The neural network is trained using this dataset, and the loss function used during training is the binary cross-entropy loss. The specific training process is well-known in neural networks, and this embodiment will not elaborate on the specific training process.

[0048] For the t-th parking space region in the current frame of the monitoring image, obtain the Euclidean distance between the centroid positions of the t-th parking space region and each other parking space region. Record the 8 parking space regions with the smallest distances as the neighboring parking spaces of the t-th parking space region. It should be noted that the parking spaces mentioned in the following instructions refer to parking space regions.

[0049] S002. Based on the fluctuation of the original occupancy probability of the target parking space in multiple consecutive historical frames, determine the degree of historical environmental turbulence of the target parking space, and calculate the state inertia factor of the target parking space by combining the difference between the original occupancy probability of the target parking space and its occupancy status in the previous frame monitoring image.

[0050] It should be noted that existing technologies presuppose a fixed judgment threshold and simply compare the original occupancy probability of a parking space inferred from a single frame image with this threshold to directly determine whether the parking space is occupied or vacant. This type of parking space status detection lacks effective resistance to various noise interferences in the complex environment of a parking lot. It is known that in an objective physical scene, the process of a vehicle entering or leaving a parking space has continuous characteristics, while environmental interference noise (such as pedestrians crossing, changes in vehicle light and shadow, etc.) is characterized by isolated high-frequency instantaneous pulse signals. Therefore, this invention first bases its detection on the wave pattern of the original occupancy probability of a parking space in historical frame images. The system dynamically assesses the intensity of environmental disturbances affecting the historical state of parking spaces, obtaining the degree of historical environmental turbulence. It then integrates the degree of historical environmental turbulence with abrupt changes in the original occupancy probability of the parking space to obtain the parking space's state inertia factor. If the intensity of environmental disturbances affecting the historical state of the parking space is low, then abrupt changes in the original occupancy probability are likely caused by environmental noise, resulting in a smaller state inertia factor. The state inertia factor effectively enhances the ability to maintain the original state of the parking space under noise interference scenarios, preventing erroneous switching of the parking space state triggered by instantaneous noise pulses, and ensuring the accuracy and stability of parking space state determination.

[0051] In this embodiment of the invention, the historical environmental turbulence level of the t-th parking space under the current frame monitoring image is obtained:

[0052] ;

[0053] In the formula, This represents the historical environmental turbulence level of the t-th parking space in the current frame of the monitoring image; This represents the number of historical frames preceding the current monitored frame. This represents the original occupancy probability of the t-th parking space in the (m+1)-th historical frame image preceding the current frame monitoring image; This represents the original occupancy probability of the t-th parking space in the m-th historical frame image preceding the current frame monitoring image; The time weight of the m-th historical frame image preceding the current frame monitoring image is represented by ||; the absolute value symbol is represented by ||. In this embodiment of the invention, the number of historical frame images preceding the current frame monitoring image is preset to K=5. In other embodiments, the implementer may preset the value of K according to the specific implementation situation.

[0054] Wherein, the time weight of the m-th historical frame image preceding the current frame monitoring image. , This represents the number of historical frames preceding the current frame; m represents the index number of the historical frame, ranging from 1 to K. The purpose of setting the weights is to make the historical environmental fluctuations more real-time, giving higher weights to historical frames that are closer to the current frame.

[0055] This represents the change in the original occupancy probability of the parking space in the m-th historical frame image preceding the current frame image; The larger the value, the greater the complexity of the historical fluctuations of the t-th parking space status under the current frame monitoring image, and the greater the influence of environmental interference noise on the t-th parking space status under the current frame monitoring image.

[0056] In this embodiment of the invention, the state inertia factor of the t-th parking space in the current frame monitoring image is obtained:

[0057] ;

[0058] In the formula, This represents the state inertia factor of the t-th parking space in the current frame of the monitoring image; This represents the original occupancy probability of the t-th parking space in the current frame of the monitoring image; This represents the degree of environmental disturbance in the t-th parking space under the current frame of the monitoring image; This represents the occupancy status value of the t-th parking space in the previous frame of the current monitoring image; This represents the preset sensitivity adjustment coefficient; || represents the minimum bias constant, used to prevent the denominator from being zero; || represents the absolute value sign.

[0059] Preset basic judgment threshold If the original occupancy probability of the t-th parking space in the previous frame of the current monitoring image is greater than or equal to the basic judgment threshold, the occupancy status value of the t-th parking space in the previous frame of the current monitoring image is 1; otherwise, the occupancy status value of the t-th parking space in the previous frame of the current monitoring image is 0.

[0060] In this embodiment of the invention, a preset sensitivity adjustment coefficient is used. , In other embodiments, implementers may preset the values ​​of the sensitivity adjustment coefficient and the minimum bias constant according to the specific implementation situation;

[0061] If the historical fluctuation complexity of the t-th parking space status in the current frame monitoring image is smaller ( (relatively small), and the original occupancy probability of the t-th parking space in the current frame monitoring image. A mutation occurred, leading to The larger the value of , the more rapidly the exponent term will decrease due to the numerator term, thus... Reduce, considering the original occupancy probability The sudden change is most likely due to environmental interference and noise.

[0062] S003. Determine the spatial distance attenuation coefficient based on the distance between the target parking space and its neighboring parking spaces; determine the spatial topology constraint deviation of the target parking space based on the spatial distance attenuation coefficient and the original occupancy probability deviation between the target parking space and its neighboring parking spaces; and perform nonlinear fusion correction on the original occupancy probability of the target parking space using the state inertia factor and the spatial topology constraint deviation to obtain the final occupancy probability.

[0063] It should be noted that parking spaces within a parking lot have inherent physical relationships in their spatial distribution. For example, when a large vehicle is parked in parking space A, due to the perspective of the camera, the vehicle's body may obstruct the markings of the adjacent parking space B. This can cause an abnormally high output probability in the parking space status recognition model for parking space B, resulting in distortion. Existing technologies treat each parking space as an independent unit for isolated judgment, failing to identify and resolve such visual coupling interference between parking spaces, easily leading to misjudgments of parking space status. Therefore, this invention obtains spatial topological constraint deviation based on the deviation between the original occupancy probability of the current parking space and its adjacent parking spaces, thereby effectively avoiding the problem of incorrect parking space status recognition caused by physical obstruction between parking spaces.

[0064] In this embodiment of the invention, the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame monitoring image is:

[0065] ;

[0066] In the formula, This represents the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image; This represents the Euclidean distance between the center point of the t-th parking space and the center point of its n-th neighboring parking spaces; This represents the width of a standard parking space; the closer the t-th parking space is to its n-th neighboring parking space, the larger the value of the spatial distance attenuation coefficient.

[0067] Obtain the spatial topology constraint deviation of the t-th parking space in the current frame of the monitoring image:

[0068] ;

[0069] In the formula, This represents the spatial topological constraint deviation of the t-th parking space in the current frame of the monitoring image. This represents the number of neighboring parking spaces of the t-th parking space; This represents the original occupancy probability of the t-th parking space in the current frame of the monitoring image; Represents the original occupancy probability of the nth neighboring parking space of the t-th parking space in the current frame of the monitoring image; || represents the absolute value sign; Represents the logarithmic function; This represents the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image;

[0070] The larger the value, the closer the t-th parking space is to its n-th neighboring parking space in the current frame of the monitored image; and the closer the neighboring parking space is to the t-th parking space, the higher the probability of them overlapping, obstructing each other, or having light and shadow coupling. The greater the weight;

[0071] If both the t-th parking space and its n-th neighboring parking space are parked in the current frame of the monitoring image, then the original occupancy probability of the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image is higher. The value of approaches 0, making The value is close to 0, indicating that the original occupancy probability of the t-th parking space in the current frame monitoring image is not affected by its neighboring parking spaces;

[0072] When the original occupancy probability of the nth neighboring parking space of the t-th parking space in the current frame of the monitoring image increases, and the original occupancy probability of the t-th parking space in the current frame of the monitoring image increases accordingly due to the occupancy of its nth neighboring parking space, The value is greater than 0, at this time The value is positive. The larger the spatial topology constraint deviation of the t-th parking space in the current frame monitoring image, the more it indicates that the original occupancy probability of the t-th parking space in the current frame monitoring image is affected by its neighboring parking spaces.

[0073] It should be noted that relying solely on the original occupancy probability of a parking space in the current frame of the monitoring image to determine the parking space status is insufficient in terms of adaptability and anti-interference capability, and cannot effectively cope with the complex and ever-changing actual parking lot scenarios. Therefore, this step nonlinearly fuses the obtained state inertia factor spatial topology constraint deviation with the original occupancy probability feature to obtain the final occupancy probability of the parking space in the current frame of the monitoring image, which is used to determine the occupancy status of the current parking space.

[0074] In this embodiment of the invention, the final occupancy probability of the t-th parking space in the current frame monitoring image is obtained:

[0075] ;

[0076] In the formula, This represents the final occupancy probability of the t-th parking space in the current frame of the monitoring image; This represents the original occupancy probability of the t-th parking space in the current frame of the monitoring image; This represents the state inertia factor of the t-th parking space in the current frame of the monitoring image; This represents the spatial topological constraint deviation of the t-th parking space in the current frame of the monitoring image. This represents a preset spatial correction intensity coefficient. In this embodiment of the invention, the preset spatial correction intensity coefficient... In other embodiments, implementers may pre-set according to specific implementation conditions. The value of the spatial correction intensity coefficient is used to enhance the system's ability to suppress and correct interference from neighboring parking spaces, by combining with... Multiply to quantify the degree to which the original occupancy probability of the t-th parking space in the current frame of the monitoring image is affected by the interference from neighboring parking spaces;

[0077] The smaller the value, the greater the probability of the original occupancy of the t-th parking space in the current frame of the monitoring image, and the more likely it is to be environmental interference noise. In this case, based on... The smaller the original occupancy probability of the t-th parking space in the current frame of the monitoring image, the less likely the vehicle will trigger a state change; conversely, the smaller the probability, the less likely the vehicle will trigger a state change. The larger the value, the greater the probability that the t-th parking space is originally occupied in the current frame of the monitored image; the more likely the parking space is to be occupied. In this case, based on... The original occupancy probability of the t-th parking space in the current frame monitoring image is not adjusted;

[0078] This is used to quantify the degree to which the original occupancy probability of the t-th parking space in the current frame of the surveillance image is affected by interference from neighboring parking spaces. The larger the value, the stronger the visual interference of the neighboring parking spaces on the t-th parking space (such as vehicle crossing the line, detection area occlusion, light and shadow coupling, etc.), and the more significant the impact of interference on the original occupancy probability.

[0079] S004. Based on the comparison result between the final occupancy probability and the preset basic judgment threshold, determine the occupancy status of the target parking space.

[0080] It should be noted that the present invention uses the difference between the final occupancy probability and the preset basic judgment threshold as the core judgment criterion to achieve accurate definition of the parking space status.

[0081] In this embodiment of the invention, a preset basic judgment threshold is used. If the final occupancy probability of the t-th parking space in the current frame of the monitoring image minus the basic judgment threshold is greater than or equal to 0, the final judgment status of the t-th parking space in the current frame of the monitoring image is 1, indicating that the t-th parking space in the current frame of the monitoring image is in an occupied state; if the final occupancy probability of the t-th parking space in the current frame of the monitoring image minus the basic judgment threshold is less than 0, the final judgment status of the t-th parking space in the current frame of the monitoring image is 0, indicating that the t-th parking space in the current frame of the monitoring image is in an idle state.

[0082] Figure 2 It is the monitoring effect of existing technology. Figure 2The blue dashed line represents the original occupancy probability output by the model. As you can see in the figure, there are multiple high-frequency noise interferences (representing pedestrians crossing the line, etc.) and a period of continuous fluctuation at a mid-to-high level (representing adjacent vehicles crossing the line, etc.). Figure 2 The red step line represents the judgment result of the existing technology. It can be clearly seen that when pedestrians pass by and adjacent cars cross the line, the parking space status changes frequently and drastically between free and occupied, which will cause the billing system to count incorrectly.

[0083] Figure 3 This is the parking space monitoring effect of the present invention. Figure 3 The blue dashed line represents the original occupancy probability output by the model; Figure 3 The green step line represents the judgment result of this invention. When pedestrians are crossing, the state inertia factor smooths out multiple high-frequency noise interferences, and the parking space status does not change between idle and occupied. During the interference of adjacent vehicles crossing the line, the spatial topology constraint deviation accurately offsets the abnormally increased probability, so that the parking space status always remains idle until the vehicle actually enters the parking space and the probability stabilizes before switching the status. The overall output of the green step line presents a clean step shape, perfectly eliminating false alarms and state jumps.

[0084] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of parking space status in a parking lot based on deep learning, characterized in that, include: Acquire several frames of monitoring images of the parking lot, input each frame of monitoring image into a pre-trained neural network, and identify each parking space in each monitoring image and the original occupancy probability of each parking space. Select any parking space in the current frame of the monitored image as the target parking space; Based on the fluctuation of the original occupancy probability of the target parking space in multiple consecutive historical frames, the historical environmental volatility of the target parking space is obtained, including: , This represents the historical environmental turbulence level of the t-th parking space in the current frame of the monitoring image. This represents the number of historical frames preceding the current monitored frame. , These represent the original occupancy probabilities of the t-th parking space in the (m+1)th and mth historical frames preceding the current monitoring image, respectively. This represents the time weight of the m-th historical frame image preceding the current frame image, where || represents the absolute value sign; Based on the historical environmental instability and the difference between the original occupancy probability of the target parking space and its occupancy status in the previous frame of the monitoring image, the state inertia factor of the target parking space is obtained, including: , , These represent the state inertia factor and the original occupancy probability of the t-th parking space in the current frame of the monitoring image. This represents the occupancy status value of the t-th parking space in the previous frame of the current monitoring image. To preset the sensitivity adjustment coefficient, It is the minimum bias constant; Based on the distance between the target parking space and its neighboring parking spaces, a spatial distance attenuation coefficient is obtained; based on the spatial distance attenuation coefficient and the deviation of the original occupancy probability between the target parking space and its neighboring parking spaces, the spatial topology constraint deviation of the target parking space is obtained, including: , , These represent the spatial topological constraint deviation of the t-th parking space and the original occupancy probability of the n-th neighboring parking space, respectively, in the current frame of the monitoring image. This represents the number of neighboring parking spaces of the t-th parking space. It is a logarithmic function. This is the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image; The original occupancy probability of the target parking space is corrected using the state inertia factor and spatial topology constraint deviation to obtain the final occupancy probability of the target parking space, including: , Let be the final occupancy probability of the t-th parking space in the current frame of the monitoring image. The preset spatial correction strength coefficient; The occupancy status of the target parking space is determined based on the comparison between the final occupancy probability and the preset basic judgment threshold.

2. The real-time parking space status monitoring method based on deep learning according to claim 1, characterized in that, The time weight of the m-th historical frame image preceding the current frame monitoring image includes: , This represents the number of historical frame images preceding the current frame image; m represents the index number of the historical frame image. This represents the time weight of the m-th historical frame image preceding the current frame's monitored image.

3. The method for real-time monitoring of parking space status based on deep learning according to claim 1, characterized in that, The process of obtaining the spatial distance attenuation coefficient includes: ； In the formula, This represents the spatial distance attenuation coefficient between the t-th parking space and its n-th neighboring parking space in the current frame of the monitoring image; This represents the Euclidean distance between the center point of the t-th parking space and the center point of its n-th neighboring parking spaces; This represents the width of a standard parking space.

4. The method for real-time monitoring of parking space status based on deep learning according to claim 1, characterized in that, The acquisition of the occupancy status value of the t-th parking space in the previous frame of the current frame monitoring image includes: Preset basic judgment threshold If the original occupancy probability of the t-th parking space in the previous frame of the current monitoring image is greater than or equal to the basic judgment threshold, the occupancy status value of the t-th parking space in the previous frame of the current monitoring image is 1; otherwise, the occupancy status value of the t-th parking space in the previous frame of the current monitoring image is 0.

5. The real-time monitoring method for parking space status based on deep learning according to claim 1 or 4, characterized in that, The step of determining the occupancy status of the target parking space based on the comparison result between the final occupancy probability and the preset basic determination threshold includes: If the final occupancy probability of the t-th parking space in the current frame of the monitoring image minus the basic judgment threshold is greater than or equal to 0, the final judgment status of the t-th parking space in the current frame of the monitoring image is 1, indicating that the t-th parking space in the current frame of the monitoring image is in an occupied state; if the final occupancy probability of the t-th parking space in the current frame of the monitoring image minus the basic judgment threshold is less than 0, the final judgment status of the t-th parking space in the current frame of the monitoring image is 0, indicating that the t-th parking space in the current frame of the monitoring image is in an idle state.

6. The method for real-time monitoring of parking space status based on deep learning according to claim 1, characterized in that, The acquisition of several frames of monitoring images of the parking lot includes: The preset frame extraction step size is 200ms. The real-time monitoring video of the parking lot is obtained by using the surveillance cameras deployed in the parking lot. The FFmpeg tool is used to divide the real-time monitoring video of the parking lot into several monitoring images according to the preset frame extraction step size of 200ms.

Images