Parking lot abnormal event detection system based on multi-source data fusion

Abstract

The application discloses a parking lot abnormal event detection system based on multi-source data fusion and belongs to the technical field of vehicle collision detection. The system comprises a data acquisition module, a vehicle type magnetic field normalization module, a potential abnormal event identification module, an abnormal event data analysis module and an abnormal event determination module. The system collects the geomagnetic time series data and monitoring video data of vehicles parked on adjacent two parking spaces, normalizes the two-parking-space geomagnetic time series data according to vehicle types, constructs a cross-parking-space multi-dimensional coupling feature after normalization, identifies potential abnormal events, and adaptively weights and fuses the geomagnetic event classification results and video motion saliency scores corresponding to the potential abnormal events based on quality evaluation results. The system can timely and accurately discover vehicle scratching and collision events in a low-illumination underground parking lot, and solves the problems of low detection accuracy and high false negative rate of abnormal events such as vehicle scratching and collision in a low-illumination underground parking lot in the prior art.

Description

Technical Field

[0001] This invention relates to the field of vehicle collision detection technology, and in particular to a vehicle parking lot abnormal event detection system based on multi-source data fusion. Background Technology

[0002] With the continuous growth of motor vehicle ownership in cities, minor collisions and scrapes in parking lots are frequent. Moreover, most shopping mall underground parking lots have high traffic volume, low lighting, and pillars that often obstruct the view of surveillance cameras, making it difficult to detect minor collisions and scrapes in parking spaces in a timely manner, thus causing certain economic losses to the parking lot.

[0003] Currently, parking lots typically use one-to-one or one-to-many video parking space detectors to detect whether accidents have occurred between vehicles.

[0004] For example, the invention patent announcement CN119252077B, which discloses a collision detection method, system, medium, and program product based on 3D projection modeling, establishes a collision enclosure cuboid model corresponding to each vehicle based on the entry monitoring video when the vehicle enters the parking lot, and then detects the projection overlap points between the cuboid models corresponding to each vehicle to determine whether an accident has occurred. Another example is the invention patent announcement CN111091718B, which discloses a parking monitoring and early warning method and system, which detects vehicle crossing the line using a parking space camera above the parking space to determine whether a collision has occurred.

[0005] The above-disclosed technical solutions have at least the following technical problems:

[0006] In traditional methods, surveillance videos in low-light underground parking lots are easily affected by vehicle color, headlight shadows, and overexposure of headlights, resulting in unclear images, decreased detection accuracy, or even failure to detect accidents, leading to missed detections of minor collisions and other accidents. To address these issues, this invention proposes a solution. Summary of the Invention

[0007] This application provides a parking lot abnormal event detection system based on multi-source data fusion, which at least solves the problems of low detection accuracy and high false negative rate of abnormal events such as vehicle scratches and collisions in parking spaces in low-light underground parking lots in related technologies.

[0008] This application provides a parking lot anomaly event detection system based on multi-source data fusion, including:

[0009] The data acquisition module is used to acquire real-time geomagnetic time-series data of the first and second parking spaces, which are collected by triaxial geomagnetic sensors deployed in adjacent first and second parking spaces and transmitted via narrowband Internet of Things, as well as monitoring video data associated with the parking spaces.

[0010] The vehicle model magnetic field normalization module is used to obtain the vehicle model normalized geomagnetic time series data based on the initial geomagnetic time series data after the vehicle is stably parked in the parking space.

[0011] The potential abnormal event identification module is used to identify the occurrence of potential abnormal events and generate trigger signals by constructing cross-parking space multidimensional coupling features based on the vehicle type normalized geomagnetic time series data of the first and second parking spaces.

[0012] The abnormal event data analysis module is used to respond to trigger signals, extract geomagnetic event segments and synchronous video event segments within a preset time window when potential abnormal events occur, perform parallel analysis on geomagnetic event segments and video event segments, and obtain geomagnetic event classification results for geomagnetic event segments, quality assessment results for video event segments, and video motion saliency scores, respectively.

[0013] The abnormal event determination module is used to adaptively weight and fuse the geomagnetic event classification results and video motion saliency scores based on the quality assessment results of video event segments to obtain the final event determination result. Based on the event determination result, it confirms whether there are any abnormal events such as scratches or collisions in the parking lot. If so, it outputs alarm information.

[0014] Furthermore, the normalized geomagnetic time series data of the vehicle model includes the following acquisition steps:

[0015] Acquire initial geomagnetic time series data and match it with a pre-set vehicle type-magnetic field feature database to determine the vehicle type category of the vehicle to be placed;

[0016] Based on the determined vehicle type, select the corresponding vehicle type normalization factor from the preset normalization factor table;

[0017] Based on the initial geomagnetic time series data and vehicle model normalization factor, normalized geomagnetic time series data for vehicle models are obtained through normalization methods.

[0018] Furthermore, the method of identifying potential abnormal events by constructing multi-dimensional coupling features across parking spaces includes:

[0019] Real-time calculation of the differential signal, cross-correlation coefficient, and magnetic field gradient between the normalized geomagnetic time series data of the first and second parking spaces;

[0020] The differential signal, cross-correlation coefficient, and magnetic field gradient are fused into a multi-dimensional coupled feature vector across parking spaces, and the corresponding norm value is obtained.

[0021] The norm value of the multidimensional coupling feature vector across parking spaces is compared with the normal disturbance threshold. When the norm value exceeds the normal disturbance threshold, it is determined that a potential abnormal event has occurred, and a trigger signal is generated.

[0022] Furthermore, the steps for constructing the multi-dimensional coupled feature vector across parking spaces include:

[0023] Obtain the normalized geomagnetic vector corresponding to time t from the normalized geomagnetic time series data of the first and second parking spaces;

[0024] At the same time, the differential signal and cross-correlation coefficient between the normalized geomagnetic vectors of the first parking space and the second parking space, the magnetic field gradient of the first parking space signal, and the magnetic field gradient of the second parking space signal are concatenated into a multi-dimensional vector, which serves as the multi-dimensional coupling feature vector across parking spaces.

[0025] Furthermore, the analysis of the geomagnetic event fragments includes:

[0026] Segments of geomagnetic events are fed into a pre-trained, lightweight neural network model designed for processing time-series data;

[0027] The lightweight neural network model outputs the probability values ​​of geomagnetic event segments as normal disturbances, scraping, and collisions, forming the geomagnetic event classification results.

[0028] Furthermore, the quality assessment results based on video event segments, including adaptive weighted fusion of geomagnetic event classification results and video motion saliency scores, include:

[0029] The image contrast and information entropy of video event segments are obtained, and the video quality score is obtained by weighted summation.

[0030] Based on the video quality score, the geomagnetic weight coefficient and the video weight coefficient are determined through a preset mapping function;

[0031] The final event confidence score is calculated using the following weighted fusion formula and compared with the judgment threshold to obtain the final event judgment result:

[0032] S f inal=wma g ·Pma g +wvid·Pvid

[0033] Among them, S f inal is the final event confidence score, Pma g The probability of scraping or collision in the geomagnetic event classification results is represented by Pvid, which is the video motion saliency score, and wma is the video motion saliency score. g wvid and wvid are the geomagnetic weight coefficient and the video weight coefficient, respectively, and their sum is 1.

[0034] Furthermore, the video weighting coefficient w vid The determination process is as follows:

[0035]

[0036] Among them, Q vid The video quality score is calculated by evaluating the average contrast (Cav) of the target region image within a video clip. g We obtain Qvid by weighted summation of the information entropy H, i.e., Qvid = α·Cav g +β·H, where α and β are preset weights, f(Q) vid ) is the Sigmoid mapping function, e is the natural constant, k is the gain coefficient of the function, and Q0 is the reference inflection point of video quality.

[0037] Furthermore, the output alarm information includes the type of abnormal event, the time of occurrence, the parking space number, and associated storage of geomagnetic event segments and video event segments.

[0038] This application utilizes real-time geomagnetic time-series data and monitoring video data corresponding to vehicles parked in adjacent parking spaces. Geomagnetic time-series data is extracted after the vehicles have parked and stabilized. The geomagnetic time-series data for both parking spaces is normalized according to vehicle type, avoiding interference from diverse vehicle types. Based on the normalized values, a multi-dimensional coupling feature across parking spaces is constructed, elevating traditional isolated parking space status monitoring to dynamic capture of relational events. This allows for sensitive identification of magnetic field interactions caused by scratches and collisions, and preliminary identification of potential abnormal events. Finally, geomagnetic event segments and synchronized video event segments are analyzed in parallel for potential abnormal events. By combining the quality assessment results of the video event segments, an adaptive weighted fusion of geomagnetic event classification results and video motion saliency scores is performed to intelligently assess the reliability of the video source and dynamically adjust the weights of geomagnetic analysis and video analysis in the final decision. When the video is clear, video details are fused to improve the accuracy of vehicle abnormal event identification; when the video is blurry, the stability of geomagnetic analysis is relied upon more to identify vehicle abnormal events, ensuring timely and accurate detection of vehicle scratches and collisions in low-light underground parking lots. Attached Figure Description

[0039] Figure 1 A flowchart of a parking lot abnormal event detection system based on multi-source data fusion provided in this application embodiment. Detailed Implementation

[0040] This application provides a parking lot anomaly event detection system based on multi-source data fusion, which solves the problems of low detection accuracy and high false negative rate of abnormal events such as vehicle scratches and collisions in low-light underground parking lots in the prior art. By utilizing the parking lot's three-axis geomagnetic sensor and monitoring camera, it achieves low-cost and high-precision detection of abnormal events such as vehicle scratches and collisions.

[0041] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0042] like Figure 1 As shown in the figure, this application provides a parking lot anomaly event detection system based on multi-source data fusion, including:

[0043] The data acquisition module is used to acquire real-time geomagnetic time-series data of the first and second parking spaces, which are collected by triaxial geomagnetic sensors deployed in adjacent first and second parking spaces and transmitted via narrowband Internet of Things, as well as monitoring video data associated with the parking spaces.

[0044] In this embodiment, a three-axis geomagnetic sensor deployed beneath two adjacent parking spaces (parking space A and parking space B) collects real-time time-series data of the geomagnetic field intensity along the X, Y, and Z axes at a fixed frequency (e.g., 100Hz). Through a built-in narrowband Internet of Things (NB-IoT) communication module, the collected raw geomagnetic data is reported to a cloud-based monitoring and management platform in real-time or in batches using a low-power, high-penetration method. Simultaneously, the monitoring and management platform records the video stream addresses of the cameras associated with the parking spaces.

[0045] Narrowband Internet of Things (NB-IoT) technology is used to enable low-power, deep-coverage data transmission from triaxial geomagnetic sensors in environments with poor signal coverage, such as underground parking lots.

[0046] The vehicle model magnetic field normalization module is used to obtain the vehicle model-specific normalized geomagnetic time series data based on the initial geomagnetic time series data after the vehicle has been stably parked in the parking space.

[0047] In this embodiment, geomagnetic data is first used to determine the parking space status, including whether it is vacant, occupied, or has been entered / exited by a vehicle. Once the vehicle has come to a complete stop, static geomagnetic data for a specific period (e.g., 5 seconds) is captured to generate the vehicle's initial geomagnetic time-series data. By comparing the initial geomagnetic time-series data with a pre-defined vehicle type-magnetic field feature database, the approximate type of the parked vehicle is identified, such as a small sedan, a mid-size SUV, or a large commercial vehicle. A corresponding normalization coefficient is then selected based on the vehicle type to normalize the amplitude of subsequent dynamic magnetic field data, eliminating the differences in the fundamental magnetic field caused by different vehicle types. The vehicle type-magnetic field feature database stores geomagnetic data for different vehicle types when they are static.

[0048] The potential abnormal event identification module is used to identify the occurrence of potential abnormal events and generate trigger signals by constructing cross-parking space multidimensional coupling features based on the vehicle type normalized geomagnetic time series data of the first and second parking spaces.

[0049] In this embodiment, normalized geomagnetic time-series data of two adjacent parking spaces (parking space A and parking space B) are continuously monitored. By statistically analyzing the correlation index between the differences between the two sensor signals, a multi-dimensional coupling feature vector across parking spaces is constructed. This vector can sensitively reflect the interaction of the magnetic fields between the two parking spaces. When the multi-dimensional coupling feature vector value exceeds the normal disturbance threshold, a potential abnormal event is determined to have occurred, and the abnormal event data analysis module is immediately triggered.

[0050] The aforementioned normal disturbance threshold is dynamically adjusted and confirmed based on the environmental noise of different parking lots. Specifically, this can be achieved by: collecting geomagnetic background noise data in different areas and time periods of the parking lot when no abnormal vehicle events occur; normalizing the collected background noise data to ensure it is in the same data format as the norm of the multi-dimensional coupled feature vector across parking spaces; calculating the statistical characteristics of the normalized background noise data, such as the mean and standard deviation; determining the initial normal disturbance threshold based on these statistical characteristics, for example, setting it as the mean plus a certain multiple of the standard deviation; dynamically adjusting the normal disturbance threshold based on actual detection results during system operation to balance the false alarm rate and the false negative rate; periodically repeating the above steps and updating the normal disturbance threshold according to the latest background noise conditions to ensure it always matches the current parking lot environment.

[0051] The abnormal event data analysis module is used to respond to trigger signals, extract geomagnetic event segments and synchronous video event segments within a preset time window when potential abnormal events occur, perform parallel analysis on the geomagnetic event segments and video event segments, and obtain geomagnetic event classification results for geomagnetic event segments, quality assessment results for video event segments, and video motion saliency scores, respectively.

[0052] In this embodiment, once a potential abnormal event is triggered, the system captures high-frequency geomagnetic data and synchronized monitoring video clips within a specific time window before and after the trigger point. The specific time window can be set to 2 seconds before and 5 seconds after the trigger time.

[0053] Geomagnetic analysis: The analysis of geomagnetic event segments includes: inputting geomagnetic event segments into a pre-trained lightweight neural network model for processing time-series data; the lightweight neural network model outputs the probability values ​​of geomagnetic event segments being classified as normal disturbances, scraping, and collisions, forming geomagnetic event classification results.

[0054] Lightweight neural network models are either Temporal Convolutional Networks (TCN) or Lightweight Long Short-Term Memory Networks (LSTM).

[0055] Specifically, this involves extracting time-series features from geomagnetic time-series data and constructing a training set containing labels for three types of events: normal disturbances, scraping, and collisions. These features and their corresponding labels are then input into a lightweight long short-term memory (LSTM) network, which is trained using a cross-entropy loss function to learn the mapping relationship between features and event labels. After training, the trained model is deployed to the target platform, enabling it to classify new geomagnetic event fragments in real time and output event classification probabilities.

[0056] Video analysis: Based on the quality assessment results of video event segments, an adaptive weighted fusion of geomagnetic event classification results and video motion saliency scores is performed, including:

[0057] The image contrast and information entropy of video event segments are obtained, and the video quality score is obtained by weighted summation.

[0058] Quality assessment of video clips: This is achieved by evaluating the average contrast (Cav) of the target region image within the video clip. g The video quality score Qvid is obtained by weighting and summing the information entropy H, i.e., Qvid = α·Cav. g +β·H, where α and β are preset weights.

[0059] In the parking lot abnormal event detection system, different combinations of α and β were used to conduct event detection experiments. The detection accuracy, false alarm rate, and other indicators under each combination were statistically analyzed and compared. Finally, the weight combination that performed best in this scenario was selected. For example, when α = 0.7 and β = 0.3, the system achieved the highest detection accuracy and the lowest false alarm rate, and this weight combination was used as the preset weight.

[0060] Based on the video quality score, the geomagnetic weight coefficient and the video weight coefficient are determined through a preset mapping function;

[0061] The final event confidence score is calculated using the following weighted fusion formula and compared with the judgment threshold to obtain the final event judgment result:

[0062] S f inal=wma g ·Pma g +wvid·Pvid

[0063] Among them, S f inal is the final event confidence score, Pma g The probability of scraping or collision in the geomagnetic event classification results is represented by Pvid, which is the video motion saliency score, and wma is the video motion saliency score. g and wvid are the geomagnetic weight coefficient and the video weight coefficient, respectively, and their sum is 1. Their values ​​are dynamically determined by the video quality score.

[0064] Among them, the video weight coefficient w vid The determination process is as follows:

[0065]

[0066] Where, f(Q) vid ) is the Sigmoid mapping function, which ensures a smooth transition of weight values ​​between 0 and 1. e is the natural constant, k is the gain coefficient of the function, which is used to control the drasticness of weight changes, and Q0 is the reference inflection point of video quality. When the video quality is higher than Q0, the video weight increases significantly.

[0067] The abnormal event determination module is used to adaptively weight and fuse the geomagnetic event classification results and video motion saliency scores based on the quality assessment results of video event segments to obtain the final event determination result. Based on the event determination result, it confirms whether there are any abnormal events such as scratches or collisions in the parking lot. If so, it outputs alarm information.

[0068] The output alarm information includes the type of abnormal event, the time of occurrence, the parking space number, and associated storage of geomagnetic event segments and video event segments.

[0069] In this embodiment, when the video quality is high: the weight of the video analysis results is increased, and geomagnetic analysis is used as the main reference to jointly determine the nature of the event.

[0070] When the video quality is low, such as when it is too dark or overexposed: the weight is almost entirely allocated to the geomagnetic analysis results, and the video is only used as supplementary evidence.

[0071] By comparing the final score obtained through weighted fusion with the preset judgment threshold, the system makes a final conclusion on potential abnormal events, such as confirming scratches, confirming collisions, or judging them as normal disturbances, and generates alarm information with accompanying geomagnetic event clips and video event clips to push to the management personnel.

[0072] Furthermore, the normalized geomagnetic time series data for vehicle models includes the following acquisition steps:

[0073] Acquire initial geomagnetic time series data and match it with a pre-set vehicle type-magnetic field feature database to determine the vehicle type category of the vehicle to be placed;

[0074] Based on the determined vehicle type, select the corresponding vehicle type normalization factor from the preset normalization factor table;

[0075] Based on the initial geomagnetic time series data and vehicle model normalization factor, normalized geomagnetic time series data for vehicle models are obtained through normalization methods.

[0076] Optionally, the subsequently acquired geomagnetic time series data can be normalized using the following formula to generate vehicle model normalized geomagnetic time series data:

[0077]

[0078] Among them, M norm (t) is the normalized triaxial geomagnetic vector at time t, M raw (t) represents the original triaxial geomagnetic vector collected at time t. Svt is the mean of the static three-axis geomagnetic vectors after the vehicle has stopped. yp e represents the vehicle model normalization factor. A pre-defined normalization factor table stores the normalization factors of different vehicle models relative to a benchmark vehicle model. The vehicle model normalization factor is pre-calibrated by statistically analyzing the geomagnetic time-series data of various vehicle types in the standard vehicle model sample library under static conditions and calculating their deviation from the benchmark vehicle model. This unifies all vehicle models into the same category, facilitating subsequent comparisons between vehicles in two parking spaces.

[0079] In this embodiment, the vehicle type is first identified using the static magnetic field characteristics after the vehicle has come to a complete stop. Based on this, the geomagnetic data is normalized to address the fundamental problem that vehicles of different sizes and chassis generate varying magnetic field signal intensities, making it difficult to standardize the detection threshold. This reduces the cost of obtaining the normal disturbance threshold. Furthermore, after normalizing the vehicle model data, the cost of obtaining training data for the lightweight neural network model is also significantly reduced; only the corresponding training set of the benchmark vehicle model needs to be obtained.

[0080] Furthermore, potential abnormal events can be identified by constructing multi-dimensional coupling features across parking spaces, including:

[0081] Real-time statistical analysis of the differential signal, cross-correlation coefficient, and magnetic field gradient between the normalized geomagnetic time-series data of the first and second parking spaces. (This represents the maximum value of the first derivative of the magnetic field strength H with respect to the displacement x);

[0082] The differential signal, cross-correlation coefficient, and magnetic field gradient are fused into a multi-dimensional coupled feature vector across parking spaces, and the corresponding norm value is obtained.

[0083] The norm value of the multidimensional coupling feature vector across parking spaces is compared with the normal disturbance threshold. When the norm value exceeds the normal disturbance threshold, it is determined that a potential abnormal event has occurred, and a trigger signal is generated.

[0084] Furthermore, the steps for constructing the multi-dimensional coupled feature vector across parking spaces include:

[0085] Obtain the normalized geomagnetic vector corresponding to time t from the normalized geomagnetic time series data of the first and second parking spaces;

[0086] At the same time, the differential signal and cross-correlation coefficient between the normalized geomagnetic vectors of the first parking space and the second parking space, the magnetic field gradient of the first parking space signal, and the magnetic field gradient of the second parking space signal are concatenated into a multi-dimensional vector, which serves as the multi-dimensional coupling feature vector across parking spaces.

[0087] Optionally, multi-dimensional coupled feature vectors across parking spaces:

[0088]

[0089] Where t is time t, F(t) is the cross-parking space multidimensional coupling feature vector at time t, and M A (t) and M B (t) represents the normalized geomagnetic vectors corresponding to the first and second parking spaces at time t, respectively. A M B (t) represents the cross-correlation coefficient between two geomagnetic vectors at time t, used to measure the degree of synchronous change in signal morphology. and These are the geomagnetic gradient vectors of the two geomagnetic vectors at time t, used to characterize the instantaneous rate of change of the signal.

[0090] In this embodiment, two adjacent geomagnetic sensors are regarded as a coupled system. By constructing a multidimensional feature vector containing differential, cross-correlation, and gradient, it can not only monitor the presence or absence of vehicles in parking spaces, but also sensitively capture the weak but unique interaction between the two magnetic fields caused by external forces such as scratches and collisions, thereby achieving a leap from "state detection" to "relationship event detection".

[0091] In summary, this embodiment does not simply add the geomagnetic and video results together, but first performs a quality diagnosis on the video data source itself. Based on the diagnosed video clarity, the importance of both geomagnetic and video evidence in the final decision is dynamically adjusted. This ensures that in good lighting conditions, video details can be used to improve accuracy; and in poor lighting conditions or when video is unreliable, geomagnetic analysis can be relied upon decisively. This allows for the most robust and reliable judgment in any parking lot environment, solving the problem of single information sources failing in complex environments in the prior art.

[0092] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0093] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0094] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0095] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0096] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0097] In conclusion, 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 spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A parking lot anomaly event detection system based on multi-source data fusion, characterized in that, include: The data acquisition module is used to acquire real-time geomagnetic time-series data of the first and second parking spaces, which are collected by triaxial geomagnetic sensors deployed in adjacent first and second parking spaces and transmitted via narrowband Internet of Things, as well as monitoring video data associated with the parking spaces. The vehicle model magnetic field normalization module is used to obtain the vehicle model normalized geomagnetic time series data based on the initial geomagnetic time series data after the vehicle is stably parked in the parking space. The normalized geomagnetic time series data for the vehicle model includes the following acquisition steps: Acquire initial geomagnetic time series data and match it with a pre-set vehicle type-magnetic field feature database to determine the vehicle type category of the vehicle to be placed; Based on the determined vehicle type, select the corresponding vehicle type normalization factor from the preset normalization factor table; Based on the initial geomagnetic time series data and vehicle model normalization factor, normalized geomagnetic time series data for vehicle models are obtained through normalization methods. The potential abnormal event identification module is used to identify the occurrence of potential abnormal events and generate trigger signals by constructing a multi-dimensional coupled feature vector across parking spaces based on the vehicle type normalized geomagnetic time series data of the first and second parking spaces. The steps for constructing the multidimensional coupled feature vector across parking spaces include: Obtain the normalized geomagnetic vector corresponding to time t from the normalized geomagnetic time series data of the first and second parking spaces; At the same time, the differential signal and cross-correlation coefficient between the normalized geomagnetic vectors of the first parking space and the second parking space, the magnetic field gradient of the signal of the first parking space, and the magnetic field gradient of the signal of the second parking space are concatenated into a multi-dimensional vector, which serves as the multi-dimensional coupling feature vector across parking spaces. The abnormal event data analysis module is used to respond to trigger signals, extract geomagnetic event segments and synchronous video event segments within a preset time window when potential abnormal events occur, perform parallel analysis on geomagnetic event segments and video event segments, and obtain geomagnetic event classification results for geomagnetic event segments, quality assessment results for video event segments, and video motion saliency scores, respectively. The abnormal event determination module is used to adaptively weight and fuse the geomagnetic event classification results and video motion saliency scores based on the quality assessment results of video event segments to obtain the final event determination result. Based on the event determination result, it confirms whether there are any abnormal events such as scratches or collisions in the parking lot. If so, it outputs alarm information. The quality assessment results based on video event segments involve an adaptive weighted fusion of geomagnetic event classification results and video motion saliency scores, including: The image contrast and information entropy of video event segments are obtained, and the video quality score is obtained by weighted summation. Based on the video quality score, the geomagnetic weight coefficient and the video weight coefficient are determined through a preset mapping function; The final event confidence score is calculated using the following weighted fusion formula and compared with the judgment threshold to obtain the final event judgment result: ； in, The final event confidence score is given. The probability of scraping or collision in the geomagnetic event classification results. For video motion saliency score, and These are the geomagnetic weighting coefficient and the video weighting coefficient, respectively, and their sum is 1; The video weight coefficient The determination process is as follows: ； in, The video quality score is calculated by evaluating the average contrast of the target region image within the video clip. and information entropy We obtain the result by weighted summation, i.e. , and To preset weights, Let e ​​be the sigmoid mapping function, and e be the natural constant. Let be the gain coefficient of the function. This serves as a reference inflection point for video quality.

2. The parking lot anomaly detection system based on multi-source data fusion as described in claim 1, characterized in that, The method of identifying potential abnormal events by constructing a multi-dimensional coupled feature vector across parking spaces includes: Real-time calculation of the differential signal, cross-correlation coefficient, and magnetic field gradient between the normalized geomagnetic time series data of the first and second parking spaces; The differential signal, cross-correlation coefficient, and magnetic field gradient are fused into a multi-dimensional coupled feature vector across parking spaces, and the corresponding norm value is obtained. The norm value of the multidimensional coupling feature vector across parking spaces is compared with the normal disturbance threshold. When the norm value exceeds the normal disturbance threshold, it is determined that a potential abnormal event has occurred, and a trigger signal is generated.

3. The parking lot anomaly detection system based on multi-source data fusion as described in claim 1, characterized in that, The analysis of the geomagnetic event segments includes: Segments of geomagnetic events are fed into a pre-trained, lightweight neural network model designed for processing time-series data; The lightweight neural network model outputs the probability values ​​of geomagnetic event segments as normal disturbances, scraping, and collisions, forming the geomagnetic event classification results.

4. The parking lot anomaly detection system based on multi-source data fusion as described in claim 1, characterized in that, The output alarm information includes the type of abnormal event, the time of occurrence, the parking space number, and associated storage of geomagnetic event segments and video event segments.

Images