Method and system for detecting and identifying dynamic target of underground parking lot

Abstract

The invention discloses a method and a system for detecting and identifying a dynamic target of an underground parking lot. The method comprises the following steps: constructing a twin neural network; the method comprises the following steps: dividing a wireless signal covered area of an underground parking lot into a plurality of grid points, coding according to positions, and respectively collecting data of each grid point when no dynamic target exists and data of each grid point when the dynamic target exists; processing the collected data, and combining the processed data with manual features and convolutional features to jointly train a twin neural network; and inputting data of a to-be-detected grid point into the trained twin neural network, judging whether a dynamic target exists in the grid point or surrounding grid points thereof, and if so, further judging the position of the dynamic target. According to the invention, fusion of wireless sensing and visual data is realized, and the detection and identification efficiency of the underground parking lot on the dynamic target is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of target detection and recognition, and in particular to a method and system for detecting and recognizing dynamic targets in an underground parking lot. Background Art

[0002] As part of public transportation, underground parking is an important part of realizing the intelligent transportation system. With the continuous development of artificial intelligence and intelligent driving, underground parking will inevitably tend to be digital and intelligent in the future. However, in underground parking environments, due to the large number of obstructions, the perception of dynamic targets is often weak, so how to solve the problem of dynamic target detection and recognition in underground parking environments has become a difficult problem.

[0003] In the current prior art, the detection and recognition of dynamic targets are mainly achieved through wireless sensors, visual sensors, etc. combined with deep learning. In the pedestrian detection method using wireless sensors, dynamic targets are usually detected in indoor environments based on the WiFi strength value RSSI, but the accuracy of its detection and positioning is low. In order to combat signal interference and attenuation, channel state information CSI is introduced to improve the accuracy of detection and positioning. Compared with RSSI, CSI provides amplitude and phase information of subcarriers of different frequencies. Due to its strong stability and high spatial resolution, it is easy to use fingerprint positioning technology and has low deployment cost. In the pedestrian detection method using visual sensors, computer vision technology and deep learning technology are often used to enhance the detection, recognition and positioning capabilities of targets in the scene, but the method has low real-time performance and is easy to block the target to be detected, thereby reducing the robustness of the system. Based on the defects of the prior art, the present invention provides a method and system for detecting and identifying dynamic targets in underground parking lots. Summary of the invention

[0004] The main purpose of the present invention is to provide a method and system for detecting and identifying dynamic targets in an underground parking lot, realize the fusion of wireless sensing and visual data, and improve the detection and identification efficiency of dynamic targets in an underground parking lot.

[0005] The technical solution adopted by the present invention is: a method for detecting and identifying dynamic targets in underground parking lots, including: constructing a twin neural network; further, the twin neural network framework includes a feature extraction module and a Mamba network, and realizes dynamic target feature detection and recognition by fusing BEBLID features, wireless signal features and the Mamba network. Among them, the feature extraction module extracts image key points in combination with the BEBLID features; the wireless signal features are extracted by collecting offline perception maps to enhance the accuracy and robustness of dynamic target positioning; the Mamba network is used for deep semantic feature extraction of images, so that the feature expression is more comprehensive and has stronger discrimination ability. Compared with non-twin neural networks, the twin network focuses on comparing the similarity of two inputs, which makes it very effective in matching and retrieval tasks. By learning similarity metrics, it can judge the relationship between input pairs, effectively learn the similarity between input pairs, and finally measure the similarity of input samples through the similarity calculation module.

[0006] The passable area of ​​the underground parking lot is divided into multiple grid points to obtain a grid set [u, v], where u is the number of grids in the vertical direction and v is the number of grids in the horizontal direction. The grid points are encoded according to their positions, and data of each grid point with and without dynamic targets is collected to form an offline perception map. The offline perception map is the wireless signal features and image data of the traversed grids collected offline.

[0007] Process the collected data, combine the processed data with manual features and convolutional features, and use them together to train the twin neural network;

[0008] The data of the grid point to be detected is input into the trained twin neural network to determine whether there is a dynamic target in the grid point or the grid points around it. If there is, the position of the dynamic target is further determined. Furthermore, the dynamic target detection and recognition method proposed in the present invention includes two stages: the first stage classifies and determines whether there is a dynamic target, which is a classification determination; the second stage determines the position of the current dynamic target.

[0009] According to the above technical solution, a channel that is robust to dynamic targets is selected, and when there is no dynamic target in the grid point, the data of the grid point includes the amplitude value and image data of the signal state information of the channel;

[0010] When there is a dynamic target in a grid point, the data of the grid point includes the amplitude value of the signal state information of the channel and the image data of the dynamic target, and the image data of the dynamic target includes the category of the dynamic target and the position of the dynamic target.

[0011] According to the above technical solution, the method for processing the collected data specifically includes: synchronizing the collected data, removing abnormal values ​​in the amplitude value of the signal state information, applying smoothing filtering to pre-process the original data of the signal state information, and defining the amplitude value as [v1 v2 v3] or [v1'v2'v3'] and presetting the value range according to whether there is a dynamic target in the grid point; mapping the amplitude value [v1 v2 v3] or [v1'v2'v3'] to a pixel value. Furthermore, the method for mapping the amplitude value [v1 v2 v3] or [v1'v2'v3'] to a pixel value includes: using the grid coordinates as the pixel point coordinates, constructing a ternary amplitude space with the three-channel values ​​in the grid, mapping the three-channel amplitude values ​​[v1 v2 v3] to the three channels of the RGB image respectively, and changing the value range accordingly. The mapping formula is as follows:

[0012]

[0013] Among them, v1, v2, and v3 correspond to the amplitude values ​​of the red (R), green (G), and blue (B) channels of the RGB image, respectively. channel Indicates the RGB value of the CSI data corresponding to this channel, v channel is the amplitude value.

[0014] According to the above technical solution, the combined manual features include BEBLID features, and the convolutional features include features trained by the Mamba model. Furthermore, the manual features can also select features such as SIFT and ORB, and the convolutional feature trainer can also use deep learning models such as YOLO. However, in the underground parking environment, the traditional deep learning method does not take into account the structured characteristics of dynamic obstacles, and the recognition of occluded and small targets is not ideal. Therefore, the present invention selects BEBLID features and Mamba models to achieve the best detection and recognition effect of dynamic targets.

[0015] According to the above technical solution, the objective function of training the twin neural network includes: in, Represents x j The features of the pth image in x j is the data sample of the grid point with dynamic target, x i is the data sample of the grid point without dynamic target, x k is the data sample where there is no dynamic target at this grid point but there are dynamic targets at other grid points, f B is a manual feature, f M is the convolution feature. Furthermore, [v1v2v3] Trepresents transpose. For the multi-dimensional, multi-subcarrier amplitude value v, in each grid, the acquisition tool can be used to collect the CSI amplitude information of 30 subcarriers. Therefore, transpose is used to convert the amplitude value v into a column vector, which is convenient for comparison with f B 、f M Features are fused.

[0016] According to the above technical solution, the method of using the trained twin neural network to determine whether there is a dynamic target in the grid point to be tested and the grid points around it specifically includes: using a classification determination algorithm, according to the data of the grid point to be tested, calculating the similarity probability between the grid point to be tested and the grid point without a dynamic target inside, the grid point with a dynamic target inside, and the grid point without a dynamic target inside but with a dynamic target in other grid points, the classification determination algorithm includes Among them, d(x n ) represents the data of the grid point to be tested and the sample x n The characteristic distance, F n is the maximum value of the feature distance, p(m|X n ) represents the current grid point m to be tested and the sample x n The similarity probability, α n is the Gaussian weight under different samples, and its value is obtained from the distribution of sample spacing. The selection range of n includes i, j, and k;

[0017] According to the size of the similarity probability, it is determined whether there is a dynamic target in the grid point to be tested.

[0018] According to the above technical solution, if there is a dynamic target in the grid point or the grid points around it, the method for determining the code of the grid point closest to the dynamic target and the position of the dynamic target specifically includes: selecting d(x n ) corresponds to the offline perception map node, which corresponds to the grid point closest to the dynamic target, obtains the code of the grid point to be measured, and obtains the position of the current dynamic target according to the code. Furthermore, since the category of the dynamic target has been marked in advance, the category information of the dynamic target can also be output.

[0019] Another aspect of the present invention provides a detection and recognition system for dynamic targets in an underground parking lot, comprising a wireless sensor module, a visual perception module, a roadside module and a power module; wherein:

[0020] The wireless sensing module is used to generate wireless signals; the roadside module is used to receive wireless signals and process and store the signals; the visual perception module is used to collect image data using photographic equipment such as cameras; and the power supply module is used to provide energy support for the above modules.

[0021] According to the above technical solution, the wireless sensing module includes a wireless signal transmitter and a gain antenna, wherein the wireless signal transmitter is used to generate a wireless signal, and the gain antenna is used to enhance the generated wireless signal. Furthermore, based on the scenario and cost considerations, the wireless signal transmitter is arranged on the pillars of the parking lot, and the range of the arrangement spacing and the arrangement height are preset. If there are no pillars for installation, the beam hanging method is adopted, the hanger is fixed on the ceiling, and the wireless signal transmitter is installed under the hanger. In order to reduce interference with driving and pedestrians, the range of the installation height is preset.

[0022] According to the above technical solution, the roadside module includes a wireless signal receiver, a data synchronization unit, a data calculation unit and a data storage unit; wherein:

[0023] The wireless signal receiver is used to receive the wireless signal generated by the wireless signal transmitter. When a dynamic target enters its monitoring range, the movement of the dynamic target will disturb the propagation of the wireless signal, thereby causing the amplitude value of the received wireless signal to change; the data synchronization unit is used to synchronize the wireless signal data and the image data of cameras with different viewing angles; the data calculation unit is used to process data, the amount of which depends on the area of ​​the underground parking lot and the processing capacity of the data calculation unit itself; the data storage unit is used to store data.

[0024] The beneficial effects of the present invention are as follows: the underground parking lot dynamic target detection and recognition method and system provided by the present invention divide the underground parking lot into grids, collect data of each grid in different states, and complete the training of the twin neural network by combining the BEBLID feature and the Mamba model. In actual application, the twin neural network is used to analyze the data of the grid point to be tested to determine whether there is a dynamic target in the grid point or its surrounding grid points and the position of the dynamic target.

[0025] Compared with the prior art: the existing dynamic target detection and recognition methods based on wireless sensing usually use the strength value RSSI of WiFi, but the accuracy of its detection and positioning is low, and it is difficult to resist signal interference and attenuation. The present invention introduces channel state information CSI, provides amplitude and phase information of subcarriers of different frequencies, has strong stability, high spatial resolution, and low deployment cost; the existing dynamic target detection and recognition methods based on visual sensing usually use computer vision technology or deep learning technology to enhance the target detection, recognition and positioning capabilities in the scene, but the method has low real-time performance, is easy to block the detection target, and thus leads to low robustness. The present invention integrates BEBLID features and Mamba model features to enhance the detection and recognition effects of dynamic targets.

[0026] Furthermore, the present invention maps the amplitude value of the channel signal to a pixel value, converts the physical characteristics of the wireless signal into an image form, combines the traditional wireless signal data with the image processing method in computer vision technology, and uses it for twin neural network feature training, effectively normalizing and mapping the signal data to the color range of the image to make it conform to the input format of conventional image processing. It can not only retain the original signal amplitude information, but also use the spatial and visual characteristics of the image data to improve the accuracy of target detection.

[0027] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0029] Figure 1 is a flow chart of a method for detecting and identifying dynamic targets in an underground parking lot according to an embodiment of the present invention;

[0030] Figure 2 is a structural diagram of a dynamic target detection and recognition system for an underground parking lot according to an embodiment of the present invention;

[0031] Figure 3 It is a framework diagram of a twin neural network in a method for detecting and identifying dynamic targets in an underground parking lot according to an embodiment of the present invention;

[0032] Figure 4 is a functional implementation flow chart of the detection and recognition system for dynamic targets in an underground parking lot according to an embodiment of the present invention;

[0033] Figure 5 4 is a layout diagram of an underground parking lot according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0035] It should be noted that the illustrations provided in the embodiments of the present invention are only used to illustrate the basic concept of the present invention in a schematic manner. Therefore, the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type may also be more complicated.

[0036] In the present invention, it is also necessary to explain that, if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, if the terms "first" and "second" appear, they are only used for description and distinction purposes, and cannot be understood as indicating or implying relative importance.

[0037] Example 1

[0038] This embodiment provides a method for detecting and identifying dynamic targets in an underground parking lot. The process is as follows: Figure 1 As shown, it includes: S1, building a twin neural network, the framework of the twin neural network is as follows Figure 3 shown.

[0039] Specifically, the twin neural network framework includes a feature extraction module and a Mamba network, which realizes dynamic target feature detection and recognition by integrating BEBLID features, wireless signal features and the Mamba network. The feature extraction module extracts image key points in combination with BEBLID features; wireless signal features are extracted by collecting offline perception maps to enhance the accuracy and robustness of dynamic target positioning; the Mamba network is used to extract deep semantic features of images, making feature expression more comprehensive and with stronger discrimination capabilities.

[0040] Compared with non-twin neural networks, twin networks focus on comparing the similarity of two inputs, which makes them very effective in matching and retrieval tasks. By learning similarity metrics, it can determine the relationship between input pairs, effectively learn the similarity between input pairs, and finally measure the similarity of input samples through the similarity calculation module. Therefore, this embodiment selects the twin neural network structure for data analysis.

[0041] S2. Divide the accessible area of ​​the underground parking lot into multiple grid points and encode them according to the location. Collect data of each grid point when there is no dynamic target and when there is a dynamic target, respectively, to form an offline perception map.

[0042] Specifically, an offline perception map is constructed, which is the wireless signal characteristics and image data of the traversed grid collected offline. The passable area of ​​the underground parking lot is gridded, and each grid is encoded to determine the position of each grid, and a grid set [u, v] is obtained, where u is the number of grids in the vertical direction and v is the number of grids in the horizontal direction.

[0043] Different from the traditional method of collecting average signal strength, the present invention collects 149-channel, 153-channel and 157-channel signals which are more robust to dynamic targets.

[0044] When there is no dynamic target in the grid point, the data of the grid point includes the amplitude value of the signal status information CSI of channels 149, 153 and 157 of the grid point and the image data; when there is a dynamic target in the grid point, the data of the grid point includes the amplitude value of the signal status information CSI of channels 149, 153 and 157 of the grid point, and the image data of the dynamic target, and the image data of the dynamic target includes the category of the dynamic target and the location of the dynamic target.

[0045] S3. Process the collected data, combine the processed data with manual features and convolutional features, and use them together to train the twin neural network.

[0046] Specifically, the collected wireless signal and image data are synchronized, the outliers in the amplitude value of the signal status information CSI are removed, and the smoothing filter is applied to preprocess the original CSI data. If there is a dynamic target in the grid point, the amplitude value is defined as [v1 v2 v3]; if there is no dynamic target in the grid point, the amplitude value is defined as [v1' v2' v3'], and the value ranges of v1, v2, v3, v1', v2', and v3' are all [0,50].

[0047] The amplitude value [v1 v2 v3] or [v1' v2' v3'] is mapped to a pixel value, and the value range of v1, v2, v3, v1', v2', and v3' is converted from [0,50] to [0,255]. The mapping formula is as follows: Among them, v1, v2, and v3 correspond to the amplitude values ​​of the red (R), green (G), and blue (B) channels of the RGB image, respectively. channel Indicates the RGB value of the CSI data corresponding to this channel, v channel is the amplitude value.

[0048] The processed amplitude value information is combined with manual features and convolutional features to train the twin neural network, wherein the manual features select BEBLID features, and the convolutional features select features trained with the Mamba model. Furthermore, the manual features can also select features such as SIFT and ORB, and the convolutional feature trainer can also use deep learning models such as YOLO. However, in underground parking environments, traditional deep learning methods do not take into account the structured features of dynamic obstacles, so the recognition of occluded and small targets is not ideal. Therefore, this embodiment selects BEBLID features and the Mamba model to optimize the detection and recognition of dynamic targets.

[0049] In this embodiment, the detection and recognition of dynamic targets is divided into two stages. The first stage classifies and determines whether there is a dynamic target. The second stage determines that if there is a dynamic target, the offline perception map node corresponding to the grid with the smallest distance is selected to determine the location of the current dynamic target.

[0050] The objective function of the two-stage training of the twin neural network is defined as: in, Represents x j The features of the pth image in x j is the data sample of the grid point with dynamic target, x i is the data sample of the grid point without dynamic target, x k is the data sample where there is no dynamic target at this grid point but there are dynamic targets at other grid points, f B is the BEBLID feature, f M Features trained for the Mamba model.

[0051] S4. Input the data of the grid point to be detected into the trained twin neural network to determine whether there is a dynamic target in the grid point or the grid points around it. If so, further determine the position of the dynamic target.

[0052] Specifically, the constructed offline perception map is used to collect data in real time to detect and identify dynamic targets. When a dynamic target enters a grid point, the CSI amplitude value collected in the grid point will change. At the same time, the image data of the dynamic target is collected, and the collected signal and image data are synchronized. The 149-channel, 153-channel, and 157-channel signals in the current wireless signal are extracted, and the signal amplitude value is mapped to a pixel value as the feature of the wireless signal.

[0053] The collected wireless signal features are input into the trained twin neural network, and the classification judgment algorithm is used to calculate the similarity probability between the grid point to be tested and the grid point without dynamic targets, the grid point with dynamic targets, and the grid point without dynamic targets but with dynamic targets in other grid points. This embodiment is improved on the basis of the Bayesian estimation method and proposes a new classification judgment algorithm:

[0054]

[0055] Among them, d(x i ) represents the data of the grid point to be tested and the sample x i The characteristic distance, F i is the characteristic distance d(x i ), p(m|X i ) represents the current grid point m to be tested and the sample x i The similarity probability, α i is the Gaussian weight under different samples, and its value is obtained through the distribution of sample spacing;

[0056] d(x j ) represents the data of the grid point to be tested and the sample x j The characteristic distance, F j is the characteristic distance d(x j ), p(m|X j ) represents the current grid point m to be tested and the sample x j The similarity probability, α j is the Gaussian weight under different samples, and its value is obtained through the distribution of sample spacing;

[0057] d(x k ) represents the data of the grid point to be tested and the sample x k The characteristic distance, F k is the characteristic distance d(x k ), p(m|X k ) represents the current grid point m to be tested and the sample x k The similarity probability, α k is the Gaussian weight under different samples, and its value is obtained through the distribution of sample spacing.

[0058] The category with the highest output probability is: M * = argmax[p(m|M i ),p(m|M j ),p(m|M k )]

[0059] According to the size of the similarity probability, that is, which sample the grid point to be tested is closest to, it is determined whether there is a dynamic target in the grid point to be tested.

[0060] The above formula can be used to determine whether there is a dynamic target in the current input data, complete the first stage of classification and judgment, and achieve the purpose of detecting dynamic targets.

[0061] Continue to the second stage of judgment and select the smallest d(x j ) is the nearest node, and the code of the corresponding grid point is retrieved to determine the location of the current dynamic target. At the same time, since the category of the dynamic target has been marked in advance, the category information of the target can also be output.

[0062] Example 2

[0063] This embodiment provides a system for detecting and identifying dynamic targets in an underground parking lot. The system executes the method for detecting and identifying dynamic targets in an underground parking lot described in Embodiment 1. The structure of the system is as follows: Figure 2 As shown, it includes a wireless sensor module, a visual perception module, a roadside module and a power module.

[0064] The wireless sensor module is used to generate a wireless signal.

[0065] Specifically, the wireless sensor module includes a wireless signal transmitter and a booster antenna. The wireless signal transmitter is used to generate a wireless signal, and the booster antenna is used to enhance the generated wireless signal.

[0066] In this embodiment, a WiFi signal is selected as a wireless signal for detecting and identifying dynamic targets, and the WiFi signal meets the IEEE 802.11b / g / n standard, allowing hardware devices that meet the above standards to access.

[0067] It is understandable that the number of wireless sensor modules deployed is proportional to the perception accuracy and inversely proportional to the deployment cost. Therefore, in order to reasonably deploy WiFi signal transmitters based on scenarios and costs, the deployment spacing between transmitters is set to 2-3 meters. At the same time, its deployment height is also limited. For dynamic targets in existing underground parking environments, there are generally pedestrians, non-motor vehicles, motor vehicles, etc., and their heights are generally distributed in the range of 0.5-2 meters. The commonly used deployment height of WiFi signal transmitters can be 1-2 meters. It can be installed on the columns of the parking lot according to local conditions. If there are no columns to install, the beam hanging method can be used to fix the hanger on the ceiling and install the WiFi signal transmitter under the hanger. To avoid interference from vehicles and pedestrians, the installation height of the WiFi signal transmitter is set to 2.0-2.2 meters.

[0068] The visual perception module is used to collect image data using photographic equipment, including a camera.

[0069] The roadside module is used to receive the WiFi signal transmitted by the WiFi signal transmitter, and process and store the signal.

[0070] Specifically, the roadside module includes a WiFi signal receiver, a data synchronization unit, a data calculation unit and a data storage unit. For underground parking lots of different areas, the number of roadside modules is different due to the different number of wireless sensor modules deployed and the different processing capabilities of the calculation units.

[0071] In the roadside module, the WiFi signal receiver is used to receive signals from the WiFi signal transmitter. When the WiFi signal propagates in the air, its amplitude value will attenuate with the distance. Therefore, in this embodiment, the WiFi signal receiver in a single roadside module only receives WiFi signals from several nearby nodes in a directional manner. At the same time, there is a certain limit on the deployment height of the WiFi signal receiver, and the deployment height range is set to 0.5-1 meter.

[0072] The data synchronization unit is used to synchronize the WiFi data received by the WiFi signal receiver and the image data collected by the visual perception module.

[0073] The calculation unit is used to process the collected data, remove abnormal values ​​in the amplitude value of the signal state information, apply smoothing filtering to pre-process the original data of the signal state information, and map the amplitude value information of the wireless signal to pixel values.

[0074] The storage unit is used to provide data storage space.

[0075] The power supply module is used to provide energy support for the above modules.

[0076] According to the structure of the underground parking lot dynamic target detection and recognition system, the system executes the process of the underground parking lot dynamic target detection and recognition method described in Example 1 as follows: Figure 4 As shown, including:

[0077] T1. Pre-build offline perception maps, complete the training of twin neural networks, and classify data samples.

[0078] T2. The WiFi signal receiver in the roadside module receives the CSI amplitude value of the WiFi signal, and the visual module observes and collects image data.

[0079] T3. The data synchronization unit synchronizes the collected wireless signal and image data, obtains the signal amplitude values ​​of channels 149, 153 and 157 in the current wireless signal, and processes them.

[0080] T4. Input the processed signal amplitude value and image data into the twin neural network, and combine it with the offline perception map to complete the detection and recognition of dynamic targets.

[0081] Example 3

[0082] This embodiment provides an underground parking lot, which is arranged as follows Figure 5 As shown, it includes the underground parking lot dynamic target detection and recognition system described in Example 2, a number of pillars, a parking area and a road area.

[0083] Among them, the several pillars are set up around the road area, and one or more of WIFI signal transmitters, roadside modules, visual perception modules and power supply modules are installed on some of the pillars.

[0084] The purpose of installing WIFI signal transmitters is to form a wireless signal field throughout the underground parking lot. The number of WIFI signal transmitters is proportional to their perception accuracy and installation cost. Therefore, considering the cost, it is necessary to reasonably arrange the WiFi signal transmitters, and set the spacing between the transmitters to 2-3 meters. At the same time, there are certain restrictions on the height of the arrangement on the column. For the dynamic targets in the underground parking environment in this embodiment, there are generally pedestrians, non-motor vehicles, motor vehicles, etc., and their heights are generally distributed in 0.5-2 meters, so the height of the WiFi signal transmitter can be selected to be 1-2 meters.

[0085] When WiFi signals propagate in the air, their amplitude value will attenuate with distance. Therefore, in this embodiment, the WiFi signal receiver in a single roadside unit only receives WiFi signals from several nearby pillars in a directional manner. At the same time, there is a certain limit on the installation height of the WiFi signal receiver. In this embodiment, it is installed on the pillar at a height of 0.5-1 meter from the ground.

[0086] The visual perception modules are arranged on the pillars at different angles to ensure that the entire image of the road area can be captured.

[0087] The power module is arranged in the column and is connected with other modules to provide energy support.

[0088] In summary, the present invention provides a method and system for detecting and identifying dynamic targets in an underground parking lot. By pre-building an offline perception map, characteristic information on the presence or absence of dynamic targets in the passage area of ​​the underground parking lot is collected, and the training of a twin neural network is completed. Dynamic targets can be detected and identified based on signal amplitude values ​​and image data, thereby realizing the fusion of wireless sensing and visual data, and effectively improving the detection and identification efficiency of dynamic targets in underground parking lots.

[0089] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, and two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.

[0090] The order of execution of each step in the above embodiment does not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0091] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A method for detecting and identifying dynamic targets in an underground parking lot, characterized in that: include: Build a twin neural network; The area of ​​the underground parking lot covered by wireless signals is divided into multiple grid points, and the points are coded according to their locations. The data of each grid point when there is no dynamic target and when there is a dynamic target are collected respectively. Process the collected data, combine the processed data with manual features and convolutional features, and use them together to train the twin neural network; The data of the grid point to be tested is input into the trained twin neural network to determine whether there is a dynamic target in the grid point or its surrounding grid points. If so, the position of the dynamic target is further determined.

2. The underground parking lot dynamic target detection and recognition method according to claim 1 is characterized in that: Selecting a channel that is robust to dynamic targets, when there is no dynamic target in a grid point, the data of the grid point includes the amplitude value and image data of the signal state information of the channel; When there is a dynamic target in a grid point, the data of the grid point includes the amplitude value of the signal state information of the channel and the image data of the dynamic target, and the image data of the dynamic target includes the category of the dynamic target and the position of the dynamic target.

3. The underground parking lot dynamic target detection and recognition method according to claim 2 is characterized in that: The method for processing the collected data specifically includes: synchronizing the collected data, removing abnormal values ​​in the amplitude value of the signal state information, applying smoothing filtering to preprocess the original data of the signal state information, and defining the amplitude value as [v1 v2 v3] or [v1'v2'v3'] and presetting the value range according to whether there is a dynamic target in the grid point, and mapping the amplitude value [v1 v2 v3] or [v1'v2'v3'] to a pixel value.

4. The underground parking lot dynamic target detection and recognition method according to claim 1 is characterized in that: The combined manual features include BEBLID features, and the convolutional features include features trained by the Mamba model.

5. The underground parking lot dynamic target detection and recognition method according to claim 3 is characterized in that: The objective functions for training the twin neural network include: in, Represents x j The features of the pth image in x j is the data sample of the grid point with dynamic target, x i is the data sample of the grid point without dynamic target, x k is the data sample where there is no dynamic target at this grid point but there are dynamic targets at other grid points, f B is a manual feature, f M is the convolution feature.

6. The underground parking lot dynamic target detection and recognition method according to claim 5 is characterized in that: The method for determining whether there is a dynamic target in the grid point to be tested and the grid points around it by using the trained twin neural network specifically includes: using a classification determination algorithm, according to the data of the grid point to be tested, calculating the similarity probability between the grid point to be tested and the grid point without a dynamic target inside, the grid point with a dynamic target inside, and the grid point without a dynamic target inside but with a dynamic target in other grid points, wherein the classification determination algorithm includes Among them, d(x n ) represents the data of the grid point to be tested and the sample x n The characteristic distance, F n is the maximum value of the feature distance, p(mX n ) represents the current grid point m to be tested and the sample x n The similarity probability, α n is the Gaussian weight under different samples, and its value is obtained from the distribution of sample spacing. The selection range of n includes i, j, and k; According to the size of the similarity probability, it is determined whether there is a dynamic target in the grid point to be tested.

7. The underground parking lot dynamic target detection and recognition method according to claim 6 is characterized in that: The method for further determining the position of the dynamic target specifically includes: selecting the smallest d(x n ) is used as the grid point to be measured that is closest to the dynamic target, and the code of the grid point to be measured is obtained to obtain the position of the current dynamic target accordingly.

8. A detection and recognition system for dynamic targets in underground parking lots, characterized in that: The system implements the underground parking lot dynamic target detection and recognition method described in claim 1, including a wireless sensor module, a visual perception module, a roadside module and a power module; wherein: The wireless sensor module is used to generate wireless signals; the roadside module is used to receive the wireless signals generated by the wireless sensor module and process and store the signals; the visual perception module is used to collect image data; and the power supply module is used to provide energy support for the above modules.

9. The underground parking lot dynamic target detection and recognition system according to claim 8, characterized in that: The wireless sensor module includes a wireless signal transmitter and a boost antenna. The wireless signal transmitter is used to generate a wireless signal, and the boost antenna is used to enhance the generated wireless signal.

10. The underground parking lot dynamic target detection and recognition system according to claim 9, characterized in that: The roadside module includes a wireless signal receiver, a data synchronization unit, a data calculation unit and a data storage unit; wherein: The wireless signal receiver is used to receive the wireless signal generated by the wireless signal transmitter; the data synchronization unit is used to synchronize data; the data calculation unit is used to process data; and the data storage unit is used to store data.

Images