Multi-camera linkage detection method

Abstract

The application relates to the technical field of intelligent monitoring, and discloses a detection method of linkage of multiple cameras, which comprises the following steps: continuously tracking a vehicle in a parking space by using a parking space camera to obtain motion state information; when it is confirmed that the spatial position of the vehicle is separated from a detection area and the motion direction of the vehicle is away from a charging pile, a vehicle driving-off trigger signal is generated; in response to the trigger signal, the parking space camera is linked to perform instance segmentation, whether a charging gun is on the ground is identified based on geometric expansion and intersection logic; license plate information is identified and is bound with the on-the-ground state to generate associated data; alarm information is generated according to the associated data and is reported to a platform. Through the time sequence linkage of the parking space camera and the gun position camera, the vehicle driving-off is determined in combination with spatial and kinematic double constraints, the charging gun state is identified by using a geometric expansion and mask intersection algorithm, and the accurate identification of the illegal behavior and the automatic reservation of the evidence chain under low algorithm power consumption are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, specifically to a detection method involving multiple cameras. Background Technology

[0002] With the increasing prevalence of unmanned operation and maintenance of new energy vehicle charging stations, higher demands are being placed on the real-time monitoring and evidence preservation of violations (such as charging guns not being returned to their positions or vehicles leaving abnormally). However, existing video surveillance solutions face technical bottlenecks in practical applications, making it difficult to balance resource efficiency with the quality of supervision.

[0003] On the one hand, in order to capture fleeting moments of violation, existing systems usually adopt a high-frequency detection mode that operates at all times and through all channels. This keeps the edge computing units under high load for a long time, resulting in huge power consumption and limiting the deployment of high-precision complex models (such as instance segmentation). This leads to low accuracy in recognizing irregularly shaped and easily occluded targets such as charging guns, and frequent false alarms.

[0004] On the other hand, due to the lack of a multi-dimensional spatiotemporal linkage mechanism, vehicle entry and exit monitoring and equipment status monitoring are often logically disconnected, making it impossible for the system to automatically identify the vehicle involved the moment a violation occurs. It is also difficult to generate an evidence chain containing complete information about people, vehicles, places, and events, which brings great difficulties to subsequent responsibility tracing and closed-loop management. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-camera linkage detection method, which solves the problems of high computational resource consumption for all-time algorithm operation in existing charging station monitoring, low accuracy of recognizing the landing status of charging guns in complex scenarios, and difficulty in associating and tracing violations with the identity of the vehicles involved.

[0006] To achieve the above objectives, the present invention provides a detection method involving multiple cameras working together, comprising the following steps:

[0007] Parking space areas are monitored using parking space cameras. First, a sequence of images of empty parking spaces is acquired, and the pixel mean and variance are calculated to construct a background model. During real-time monitoring, the pixel difference between the current input video frame and the background model is calculated to extract the vehicle foreground region. A multi-target tracking algorithm is used to perform connected component analysis and inter-frame correlation on the foreground region to obtain the vehicle's motion state information, including its spatial position and direction of motion, in the image coordinate system. In this process, Kalman filtering is used to construct the motion state equation and smooth the vehicle constraint boxes and motion speeds of consecutive frames.

[0008] The vehicle departure determination is performed based on the motion state information. The coordinates of the four vertices of the vehicle constraint box are obtained. If the coordinates of the four vertices are all outside the preset parking space detection area, it is determined that the vehicle no longer occupies the parking space area in space. At the same time, the current motion speed and motion direction of the vehicle are read, and the azimuth angle of the pre-calibrated charging pile in the image coordinate system is obtained. The minimum angle between the motion direction and the azimuth angle of the charging pile is calculated.

[0009] When the minimum included angle is greater than 90 degrees and the movement speed is greater than the set threshold, the vehicle is determined to meet the movement conditions of moving away from the charging pile. The judgment results of multiple consecutive frames are logically ANDed. When the spatial position departure condition and the movement departure condition are met at the same time, a vehicle departure trigger signal is generated.

[0010] After generating the vehicle departure trigger signal, the system retrieves real-time images from the charging gun camera for instance segmentation. The system defaults to keeping the charging gun camera in standby or low frame rate mode, switching to real-time acquisition mode only when a trigger signal is detected. The acquired images are input into a pre-trained instance segmentation model, which outputs the minimum bounding rectangle of the detected charging gun target and the pixel-level segmentation masks of the charging piles, vehicles and personnel identified in the same scene.

[0011] Using the center point of the smallest bounding rectangle of the charging gun target as a reference, the width and height are expanded outward by a preset ratio to obtain the charging gun expansion area; the pixel intersection between the charging gun expansion area and the pixel-level segmentation mask of the charging pile, vehicle and personnel is calculated; if the pixel intersection result is all zero, it is determined that the charging gun is in the ground state; if there are non-zero pixels, it is determined that the charging gun is in the non-ground state.

[0012] The system uses parking space cameras to acquire vehicle identification images and recognize vehicle identification information. It extracts the vehicle constraint box region image corresponding to the identity of the tracked vehicle, obtains the license plate corner coordinates through a key point detection model, calculates the homography matrix mapped to the standard license plate target coordinate system, performs perspective transformation and resampling on the original license plate area to generate a front view image, and inputs it into an optical character recognition model to decode the license plate string information. The recognized vehicle identification information is spatiotemporally bound with the confirmed landing status to construct associated data including the charging gun landing confirmation timestamp, device unique identifier, parking space number, and vehicle identification information.

[0013] Based on the associated data, alarm information is generated. The system maintains a circular buffer for the video stream in memory. When associated data is generated, the buffer is frozen and video frame data before and after the event are stitched together to synthesize the on-site video file. The vehicle identification image and the panoramic view containing the frame of the charging gun are selected as key frame images. The on-site video file and key frame images are written to local non-volatile memory and a media file link is generated. The alarm information and media file link are reported to the management platform, and the charging pile is controlled to suspend service according to the instructions fed back by the platform.

[0014] This invention provides a detection method involving multiple cameras working together. It has the following beneficial effects:

[0015] 1. This invention employs a linkage mechanism between parking space monitoring and camera detection. The instance segmentation algorithm of the camera is only activated when the parking space camera confirms that the vehicle meets the departure conditions of spatial detachment and motion deviation. Compared to parallel analysis of dual-channel video streams throughout the entire time period, this avoids the ineffective operation of high-computing-power algorithms, reduces the resource consumption of computing units, and lowers the overall power consumption of the system.

[0016] 2. This invention utilizes instance segmentation combined with the intersection judgment logic of geometrically expanded regions. By expanding the charging gun detection frame outward and calculating the overlap between the charging gun detection frame and the masks of charging piles, vehicles, and personnel, the physical contact state of the charging gun is determined. This helps solve the problem that relying solely on target detection algorithms cannot effectively distinguish between the charging gun's suspended, connected to vehicle, and grounded states, improving the accuracy of determining the grounded state of irregular shapes and thin cables.

[0017] 3. This invention constructs a four-tuple of violation events containing time, device identifier, parking space number, and license plate information, logically linking physically separate vehicle departure events with charging gun detection results. Combined with video stitching technology using a circular buffer, the system can automatically retain complete video evidence before and after the event, providing data support for remote operation and maintenance and violation tracing of charging stations. Attached Figure Description

[0018] Figure 1 This is a flowchart of the intelligent detection method of the present invention;

[0019] Figure 2 This is a schematic diagram of the ROI (Region of Interest) for parking space detection according to the present invention;

[0020] Figure 3 This is a schematic diagram of the camera monitoring system for the charging pile according to the present invention. Detailed Implementation

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0022] Please refer to the attached Figure 1-3 , this embodiment is applied to a monitoring system including intelligent charging piles. The hardware environment of this system includes a computing unit set in a charging station and an image acquisition device communicatively connected thereto. The image acquisition device includes at least one parking space camera and one gun position camera. The field of view angle of the parking space camera covers the vehicle parking area in front of the charging pile, that is, the parking space detection area, and is used to monitor the entry and exit status and movement trajectory of the vehicle. The field of view angle of the gun position camera covers the gun hanging area of the charging pile and the surrounding ground, and is used to monitor the home position status of the charging gun. The computing unit is configured to receive image data and execute corresponding computer programs to implement detection logic.

[0023] Refer to the attached Figure 1 , the present invention provides a detection method for multi-camera linkage. By the time sequence linkage of parking space monitoring and gun position monitoring, the identification of illegal behaviors is realized, including the following steps:

[0024] Step S100, use the parking space camera to collect real-time image data of the parking space area, construct a background model and extract foreground targets. Continuously track the vehicles in the parking space through a multi-target tracking algorithm, and obtain the position coordinates, movement speed and movement direction information of the vehicles in the image coordinate system.

[0025] Step S200, perform a vehicle departure determination based on the tracking information of the vehicle. Define the region of interest (ROI) of the parking space detection area, determine whether the vehicle has completely left the area, and combine the vehicle's movement speed threshold and the movement angle relative to the charging pile to confirm whether the vehicle is in a departure state.

[0026] Step S300, in response to the determination result of the vehicle departure, trigger the gun position camera to start the charging gun status detection in a linked manner. Perform instance segmentation on the image collected by the gun position camera, obtain the region information of the charging gun, and combine the geometric expansion algorithm and the intersection determination logic to identify whether the charging gun is in a landed state.

[0027] Step S400, during the vehicle departure process, use the parking space camera to synchronously perform license plate recognition. Detect the license plate area of the vehicle and recognize the character information, and temporally and spatially bind the recognized license plate information to the charging gun landed state generated in step S300 to generate associated data including time, location, license plate and event type.

[0028] Step S500: Generate alarm information based on the associated data. The corresponding live video stream and keyframe images are persistently stored locally as media files, and an access link is generated based on the storage path. Finally, a message containing the alarm information and media file link is reported to the management platform via a communication protocol, triggering the corresponding linkage control strategy.

[0029] Please see the appendix Figure 1-3 The following section will explain each step in detail with reference to the specific algorithm principles.

[0030] Real-time image data of the parking space area is collected using a parking space camera, a background model is constructed and the foreground target is extracted, and the vehicle in the parking space is continuously tracked through a multi-target tracking algorithm, specifically including steps S101 to S104.

[0031] S101, Background model initialization. During system startup or in a vehicle-free state, the computing unit controls the parking space camera to continuously acquire data. A sequence of images showing empty parking spaces. (For image coordinates...) For each pixel at that location, calculate the position of the pixel at that location. Pixel mean in frame image With variance This is used to construct a background model that reflects the inherent characteristics of parking spaces. The background model This is used to distinguish between static environments and dynamic targets in subsequent processing. Specific parameter settings for the image acquisition device and image preprocessing (such as denoising and histogram equalization) can be adjusted by those skilled in the art according to the actual lighting environment, and will not be elaborated upon here.

[0032] S102, Real-time Foreground Extraction. During real-time monitoring, the computing unit acquires the current time. Input video frames To separate the moving vehicle from the background, the system calculates the current frame. With background model The pixel differences between them. Set a grayscale difference threshold. In this embodiment The value is 20, and the foreground mask is calculated using the following formula. :

[0033] ;

[0034] in, For a binarized image, when the absolute value of the gray-level difference of a pixel is greater than a threshold... When the signal is clear, the point is marked as a foreground point; otherwise, it is marked as a background point. To eliminate the interference of environmental noise on the detection, the generated primary foreground mask is... Morphological opening operations are performed to remove isolated noise points, resulting in a smoothed vehicle foreground region.

[0035] S103, Multi-target tracking and state parameter output. The foreground region extracted in step S102 is input into the multi-target tracking algorithm for processing. First, connected component analysis is performed on the foreground region, and the minimum bounding rectangle of each independent connected component is extracted as a candidate detection box. Then, the ByteTrack algorithm is used to perform inter-frame correlation on the candidate detection boxes. The computing unit assigns a unique identifier (ID) to each detected vehicle target and outputs the motion state information of the target in the current frame. This motion state information includes the vehicle constraint box. Speed ​​of movement and direction of movement Among them, the vehicle constraint frame Defined as the smallest bounding rectangle of the vehicle foreground region in the image coordinate system, it is expressed as:

[0036] ;

[0037] in, The coordinates of the top left corner of the rectangle. The coordinates are the bottom right corner of the rectangle.

[0038] Speed ​​of movement The unit is pixels per frame, representing the displacement rate of the vehicle's centroid. Direction of motion. The angle of the vehicle's motion vector relative to the horizontal axis of the image coordinate system.

[0039] S104, Kalman filtering smoothing of state parameters. Due to changes in ambient lighting or noise from the camera itself, the coordinates and velocity values ​​of the directly output single-frame detection box may fluctuate. To improve the accuracy of subsequent departure determination, the system introduces Kalman filtering to construct a motion state equation, establishing a state vector containing the vehicle's center coordinates, aspect ratio, and rate of change. The system uses the estimated value from the previous moment and the observed value from the current moment to predict the vehicle's optimal position in the next frame. The calculation unit processes the tracking boxes from three consecutive frames. Coordinates and velocity The system performs smoothing processing to output a filtered vehicle state sequence. This smoothing process corrects trajectory deviations caused by brief occlusions or detection jumps, ensuring a continuous and stable output motion trajectory.

[0040] The vehicle departure is determined based on the vehicle tracking information. The vehicle departure behavior is confirmed by using the fixed perspective of the parking space camera and combining multi-dimensional constraints of spatial position and motion state. Specifically, steps S201 to S204 are included.

[0041] S201, Define the parking space detection area. Based on the installation location and monitoring field of view of the parking space camera, define the region of interest in a two-dimensional image coordinate system. The system calibrates the image by selecting the coordinates of four vertices during the configuration phase. The system then generates a closed rectangular polygon mask based on the selected vertices.

[0042] The boundary of this area is aligned with the parking space markings in the physical scene, serving as a spatial reference for determining whether a vehicle is parked within a parking space.

[0043] S202, Perform region location determination. The calculation unit obtains the vehicle constraint box in the current frame. Let the coordinates of the four vertices of the vehicle constraint box be denoted as . The spatial logic condition for determining that a vehicle has completely left the parking space area. All vertices of the vehicle constraint box are located within the region of interest. Externally, mathematically expressed as:

[0044] ;

[0045] in, In mathematical terms, it is a universal quantifier, indicating that a determination cannot be made by considering only a single point, but must be made for each vertex. All must be checked. The universality of the conditions was emphasized, meaning that all four corners of the vehicle must be checked without exception. The mathematical meaning of is the membership symbol in set theory, representing... It is taken from the set that follows. Elements in; The set of vertices representing the vehicle constraint box is defined in the text as follows: This includes the coordinate data of the four corner points of the vehicle's rectangular frame: the top left, top right, bottom left, and bottom right. The mathematical meaning of is the negation form of a symbol, representing a point. The coordinates are not in the area specified later. Within the range. That is, the point is located outside the parking space area.

[0046] That is, by using a point-polygon positional relationship algorithm (such as the ray casting method), it is determined whether each corner point of the vehicle's bounding box falls outside the polygon. Within the pixel range.

[0047] When this condition is met, it is determined that the vehicle no longer occupies the parking space area in terms of spatial location.

[0048] S203, Perform kinematic feature verification. To distinguish between the vehicle's actual departure behavior and operations such as fine-tuning and reversing into a parking space, the system introduces kinematic constraints. The calculation unit reads the vehicle's current speed. and direction of movement And obtain the azimuth angle of the pre-calibrated charging pile in the image coordinate system. Determining the motion conditions for valid vehicle departure. Both the minimum speed threshold constraint and the directional deviation constraint must be satisfied simultaneously. The calculation formula is as follows:

[0049] ;

[0050] in, Indicates the direction of movement relative to the azimuth of the charging pile The smallest included angle between them.

[0051] Considering the periodicity of the angle, the specific calculation logic for this included angle is as follows: First, calculate the difference. ,like Then the included angle is ;like Then the included angle is When the calculated minimum included angle is greater than 90 degrees, the vehicle's trajectory is determined to deviate from the charging pile; simultaneously, the speed is required to... Speeds greater than 5 km / h are used to eliminate interference caused by stationary or slow-moving vehicles.

[0052] S204, Continuous Frame Timing Confirmation. To eliminate false positives caused by detection frame jitter or momentary occlusion, the system establishes a timing confirmation mechanism. The number of consecutive confirmation frames is set. (In this embodiment) The value is 3). The calculation unit is for continuous... The frame determination results are then subjected to a logical AND operation. Only if the frames are consecutive... All frames simultaneously satisfy the spatial logic condition of step S202. and the motion conditions of step S203 At that point, the system finally confirms that the vehicle has left the site and generates a trigger signal to initiate the subsequent charging gun detection process.

[0053] In response to the determination that the vehicle has left, the charging gun position camera is triggered to start the charging gun status detection. Through an algorithm based on instance segmentation and geometric region analysis, it is identified whether the charging gun is in an abnormal landing state, specifically including steps S301 to S305.

[0054] S301, Linkage triggers image acquisition. The computing unit controls the camera to be in standby or low frame rate monitoring mode. When it receives the vehicle departure trigger signal generated in step S204, it immediately retrieves the real-time image stream from the camera.

[0055] S302, Multi-class Target Instance Segmentation. Images captured by the charging station camera are input into a pre-trained YOLOv8-seg instance segmentation model. The YOLOv8-seg instance segmentation model has been pre-trained on a charging station dataset for multiple classes, and the output classes include at least charging guns, charging piles, vehicles, and personnel.

[0056] The computing unit performs inference on the current frame and outputs the minimum bounding rectangle of the detected charging gun target. (including width) and height (and pixel-level segmentation masks for charging piles, vehicles and people identified in the same scene).

[0057] S303, detection region geometric expansion and mask generation. The computational unit is... Using the center point as a reference, the width and height are expanded outwards by a preset proportion (20% in this embodiment) to obtain the expanded area. Expanded width With height The calculation formula is:

[0058] ;

[0059] ;

[0060] Subsequently, the system constructs a blank binary image in memory with the same resolution as the original image, and then extends the region. The corresponding pixel position is set to 1, and the rest are set to 0 to generate the charging gun extended mask. .

[0061] S304, Multi-object mask intersection determination. The calculation unit extracts the masks of other entities output in step S302, and denotes the mask of the charging pile as... The vehicle's mask is The personnel mask is .

[0062] The system calculates the extended mask using pixel-level logical AND operations. The overlap with the aforementioned entity mask set is determined using the following logic:

[0063] If the result of the logical AND operation between the extended mask and any entity mask contains a non-zero pixel, the charging gun is determined to be in a non-grounded state. This state indicates that the charging gun has a physical connection or obstruction relationship with the charging pile, vehicle, or person in its spatial projection. Specifically, the result of the logical AND operation between the extended mask and any entity mask contains a non-zero pixel, i.e., it satisfies:

[0064] ;

[0065] in, This represents the intersection operation (i.e., finding the overlapping part); (Not equal to empty set) indicates that the result of the operation is not empty, that is, there is at least one pixel in the set.

[0066] Conversely, if the result of the logical AND operation is 0 for all pixels, meaning that the extended mask has no intersection with any entity mask, then the charging gun is determined to be in a grounded state.

[0067] S305, Landing status timing confirmation. The computing unit performs continuous... Frame (in this embodiment) The system statistically analyzes the detection results. Only when three consecutive frames are determined to be in a landing state does the system confirm the charging gun landing event and generate a corresponding event flag.

[0068] During the vehicle's departure, license plate recognition and event association are performed simultaneously using parking space cameras, and alarm reporting and linkage are executed based on the association results. The specific implementation steps are divided into S401 to S402 and S501 to S502.

[0069] S401, License plate correction and recognition based on key point detection. At the moment when the vehicle departure determination is confirmed in step S204, the calculation unit extracts the vehicle constraint box corresponding to the tracked vehicle ID from the current frame or the most recent historical buffer frame. Regional image. It should be noted that this embodiment uses license plate recognition as the specific method for obtaining vehicle identification information, but in other embodiments, similar positioning and recognition can also be performed on vehicle body spray-printed codes or affixed QR codes.

[0070] First, the image of this region is input into the YOLOv8-POSE keypoint detection model. This model is trained to output the coordinates of the four corner points of the license plate. .

[0071] Subsequently, to eliminate perspective distortion caused by the camera's overhead view, the system performs a perspective transformation. The aspect ratio of standard markings (such as standard license plates) is set to... (e.g., 440:140), construct the corresponding target rectangle coordinate system. Calculate the homography matrix that maps the four detected corner points to the target rectangle, and use this matrix to resample the original identification area to generate a license plate image from a frontal view (hereinafter referred to as "vehicle identification close-up image").

[0072] Finally, the corrected close-up image of the vehicle identification is input into the PPOCR optical character recognition model, which decodes and outputs the vehicle identification information (i.e., license plate string information). If the confidence level of a single identification is lower than a preset threshold (e.g., 0.85), the system will backtrack to the most recent record in the historical tracking queue. The above steps are repeated for each frame image, and a voting mechanism is used to select the string with the highest frequency as the final recognition result. At the same time, the system temporarily stores the close-up image of the vehicle logo corresponding to the frame with the highest recognition confidence (or the one selected by voting) in memory as the best close-up evidence for subsequent steps.

[0073] S402, Multi-dimensional spatiotemporal data association and binding. The computing unit constructs structured event data packets in memory. The system uses the vehicle tracking ID as an anchor point to uniquely match the charging gun landing status confirmed in step S305 with the vehicle identification information identified in step S401.

[0074] Specifically, the system generates a four-tuple of violation events. :

[0075] ;

[0076] in, The timestamp confirming the charging gun landed; It serves as a unique hardware identifier for charging pile equipment; Logical numbering for parking spaces; This refers to the vehicle identification information associated with the vehicle.

[0077] This step merges the physically separate charging gun detection events and vehicle departure events into a single violation record in terms of data logic.

[0078] S501, Evidence chain retention based on a circular buffer. In this embodiment, media files refer to digital audiovisual materials and key image evidence that record the entire process of a violation, specifically including but not limited to synthesized on-site video files, close-up images of vehicle identification, and panoramic images including the frame of a charging gun. A media file link refers to a unique address identifier or Uniform Resource Locator used to locate and access the aforementioned media file in a network or local file system. It should be noted that the storage action is a necessary prerequisite step for generating a media file link.

[0079] To provide traceable on-site evidence, the computing unit creates a video stream circular buffer in the memory and continuously overwrites the latest video frame data.

[0080] In response to the generation of an event the system immediately freezes the buffer and continues to record backward for a preset duration. The system extracts the video data for seconds before the moment of the event occurrence (e.g., 5 seconds) from the buffer, and splices the real-time video data for seconds (e.g., 5 seconds) after the event occurrence to synthesize a complete on-site video file with a total duration of seconds.

[0081] Meanwhile, the system saves the close-up view of the vehicle identification temporarily stored in step S401 and the panoramic view containing the landed charging gun frame in step S305 as key frame pictures. The above files are written into the local non-volatile memory, and corresponding file access URIs are generated.

[0082] The system writes the above on-site video file and key frame pictures (collectively referred to as "media files") into the local non-volatile memory. After the writing is completed, the system generates a media file link in the HTTP or RTSP format according to the file system path mapping.

[0083] S502, Protocol Encapsulation and Multi-level Linkage Reporting. The computing unit constructs an alarm message based on the associated data generated in step S402. Using the MQTT Internet of Things communication protocol, the alarm message is encapsulated into a message payload in the JSON format.

[0084] The message payload specifically includes: the quadruple data of the violation event, the description of the event type, and the media file link generated in step S501. It should be emphasized that only the link is reported instead of the file body, which can greatly reduce the pressure on the Internet of Things transmission bandwidth. [[ID=2,7]]

[0085] The system publishes this message to a preset topic on the management platform. After receiving the reported information, the platform side performs the following linkage operations:

[0086] Work Order Generation: Automatically generate an operation and maintenance dispatch record according to .

[0087] Notification Push: Send an alarm to relevant management personnel through the SMS gateway or the APP push interface.

[0088] Hardware Linkage: If the system is configured with hardware control permissions, the platform issues a stop service instruction to the corresponding charging pile controller through OCPP (Open Charge Point Protocol) or other industrial control protocols to suspend the charging permission of the pile until the on-site maintenance personnel complete the reset of the charging gun and close the alarm work order, so as to achieve closed-loop control.

Claims

1. A detection method involving multiple cameras, characterized in that, Includes the following steps: The parking space area is monitored using a parking space camera, and the vehicles in the parking space are continuously tracked using a multi-target tracking algorithm to obtain the motion state information of the vehicles in the image coordinate system, including their spatial position and direction of motion. Based on the motion state information, a vehicle departure determination is performed. When it is confirmed that the spatial position of the vehicle has left the preset parking space detection area, and the motion speed is greater than the set minimum speed threshold and the motion direction meets the condition of moving away from the charging pile, a vehicle departure trigger signal is generated. In response to the vehicle leaving trigger signal, the system retrieves and analyzes real-time images from the charging gun camera to determine whether the charging gun is in a grounded state. The vehicle identification image of the vehicle is obtained using a parking space camera and the vehicle identification information is identified. The vehicle identification information is then spatiotemporally bound to the confirmed landing status to generate associated data. An alarm message is generated based on the associated data, and the alarm message and media file link are reported to the management platform; wherein, the media file link is used to access the media file corresponding to the associated data.

2. The detection method using multiple cameras in linkage according to claim 1, characterized in that, The steps of monitoring the parking area using a parking space camera and continuously tracking vehicles within the parking space using a multi-target tracking algorithm specifically include: The parking space camera is controlled to acquire a sequence of images of empty parking spaces, and the pixel mean and variance are calculated to construct a background model. During real-time monitoring, the pixel difference between the current input video frame and the background model is calculated. Pixels with an absolute grayscale difference greater than a set grayscale difference threshold are marked as foreground points. Morphological opening operations are performed on the generated foreground mask to obtain the vehicle foreground region. Connectivity analysis is performed on the foreground region of the vehicle to extract the minimum bounding rectangle of the independent connected regions as candidate detection boxes. Inter-frame correlation is performed on the candidate detection boxes to assign a unique identifier to each vehicle target to distinguish different vehicle targets. The vehicle constraint box and motion speed of the vehicle target in the current frame are output, where the motion speed represents the displacement rate of the vehicle centroid. A motion state equation is constructed using Kalman filtering, and the vehicle constraint boxes and motion speeds of multiple consecutive frames are smoothed to output the filtered motion state information.

3. The detection method using multiple cameras in linkage according to claim 2, characterized in that, The step of determining vehicle departure based on the motion state information specifically includes: Obtain the coordinates of the four vertices of the vehicle constraint box of the vehicle, and determine whether the coordinates of the four vertices are all located outside the parking space detection area. If so, determine that the vehicle no longer occupies the parking space area in the spatial position. Read the current speed and direction of movement of the vehicle, and obtain the azimuth angle of the pre-calibrated charging pile in the image coordinate system; Calculate the minimum angle between the direction of movement and the azimuth angle of the charging pile in the image coordinate system. When the minimum angle is greater than 90 degrees and the speed of movement is greater than the minimum speed threshold, it is determined that the vehicle meets the movement conditions for effective departure. A logical AND operation is performed on the determination results of a consecutive preset number of frames. Only when multiple consecutive frames of images simultaneously meet the conditions that the parking space area is no longer occupied at the spatial location and the motion condition of effective departure, the vehicle departure state is confirmed to be established, and the vehicle departure trigger signal is generated.

4. The detection method using multiple cameras in linkage according to claim 1, characterized in that, The step of analyzing real-time images from the camera at the charging gun location to identify whether the charging gun is in a grounded state specifically includes: The image captured by the gun-position camera is input into a pre-trained instance segmentation model, which outputs the minimum bounding rectangle of the detected charging gun target and the pixel-level segmentation mask of the charging pile, vehicle and personnel identified in the same scene. Using the center point of the smallest bounding rectangle of the charging gun target as a reference, the width and height are expanded outward by a preset ratio to obtain the charging gun expansion area; Calculate the pixel intersection between the extended area of ​​the charging gun and the pixel-level segmentation mask of the charging pile, vehicle and personnel; Based on the calculation results of the pixel intersection, it is determined whether the charging gun is in a grounded state.

5. The detection method using multiple cameras in linkage according to claim 4, characterized in that, The step of identifying whether the charging gun is in a grounded state also includes: If there are non-zero pixels in the result of the pixel intersection, it is determined that the charging gun is in a non-grounded state, indicating that the charging gun has a physical connection or occlusion relationship with the charging pile, vehicle or person in the spatial projection. If the result of the pixel intersection is all zero, that is, the extended area of ​​the charging gun has no intersection with the pixel-level segmentation mask of all entities, then the charging gun is determined to be in the ground state. The detection results of a preset number of consecutive frames are statistically analyzed, and the charging gun landing event is confirmed only when multiple consecutive frames are determined to be in a landing state.

6. The detection method using multiple cameras in linkage according to claim 2, characterized in that, The vehicle identification information specifically refers to license plate string information. The step of acquiring the vehicle identification image and recognizing the vehicle identification information using a parking space camera specifically includes: Extract the vehicle constraint box region image corresponding to the identity identifier of the tracked vehicle, input it into the key point detection model, and output the coordinates of the four corner points of the license plate; Set the aspect ratio of the standard license plate and construct the corresponding target rectangular coordinate system. Calculate the homography matrix that maps the detected coordinates of the four corner points to the target rectangular coordinate system. The original license plate area is subjected to perspective transformation and resampling using the homography matrix to generate a license plate image from a frontal view. The license plate image is input into the optical character recognition model for decoding and output of the license plate string information. If the confidence level of a single recognition is lower than a preset threshold, the recognition steps are repeated by tracing back multiple frames of images in the historical tracking queue, and a voting mechanism is used to select the final recognition result.

7. The detection method using multiple cameras in linkage according to claim 2, characterized in that, The steps for generating associated data specifically include: Extract and utilize the vehicle tracking identifier as an anchor point, and uniquely match the confirmed landing status with the identified vehicle identifier information; Construct a violation event quadruple that includes a timestamp confirming the charging gun's landing, a unique hardware identifier for the charging pile device, a logical parking space number, and the vehicle identifier information of the associated vehicle, and use the violation event quadruple as the associated data.

8. The multi-camera linkage detection method according to claim 4, characterized in that, Also includes: A circular buffer for the video stream is created in memory and continuously overwritten with the latest video frame data. In response to the generation of the associated data, the circular buffer is immediately frozen, the video frame data before the time of the event is extracted and the real-time video frame data after the event is spliced ​​together to synthesize a complete on-site video file; Select the vehicle identification image used to generate the vehicle identification information in the associated data as a key frame image, and save the key frame image and a panoramic view containing the minimum bounding rectangle of the charging gun target. The on-site video file and the keyframe image are written to the local non-volatile memory, and a corresponding media file link is generated.

9. The detection method using multiple cameras in linkage according to claim 1, characterized in that, It also includes the following steps: The message payload containing the associated data and the media file link is published to the management platform using the Internet of Things communication protocol, thereby triggering the management platform to generate an operation and maintenance dispatch record based on the associated data; Receive a stop service command from the management platform in response to the alarm information; In response to the stop service command, a control signal is sent to the corresponding charging pile controller to suspend the charging privileges of the charging pile.

10. The multi-camera linkage detection method according to claim 1, characterized in that, The step of responding to the vehicle departure trigger signal and retrieving and analyzing real-time images from the camera includes the following triggering logic: By default, the camera is controlled to be in standby mode or low frame rate monitoring mode. The camera is switched to real-time acquisition mode only when a vehicle departure trigger signal is detected. The real-time image stream captured by the gun-position camera after switching modes is obtained for subsequent analysis.

Images