An AI vision and geomagnetic fusion-based charging parking space occupation identification system

Abstract

The present application relates to the technical field of charging parking space image recognition and vehicle state detection, in particular to a charging parking space occupation recognition system based on AI vision and geomagnetic fusion. The system comprises an image acquisition module, a geomagnetic detection module and a processing module. The processing module extracts state transition events and corresponding time from the geomagnetic sequence, and accordingly intercepts the associated vision sequence. The vehicle is identified through an artificial intelligence vision model, and a cross-frame trajectory is established. The direction and stay relationship of the vehicle crossing the parking space boundary are obtained. The time sequence mutual verification of the vision trajectory and the geomagnetic event is completed through the state transition table, and the conflict data is retained. The present application updates the empty parking space vision baseline and the geomagnetic environment baseline when the parking space is idle, and performs event playback verification on the candidate baseline. The false state switching caused by occlusion, light change and magnetic field drift is reduced, and the continuity and long-term stability of the charging parking space occupation recognition are improved.

Description

Technical Field

[0001] This invention relates to the field of charging parking space image recognition and vehicle status detection technology, specifically a charging parking space occupancy recognition system based on the fusion of AI vision and geomagnetism. Background Technology

[0002] Existing charging parking space occupancy recognition systems typically place fixed cameras above or to the side of the charging area. The image area of ​​each charging parking space is pre-defined by camera calibration, and the system continuously captures images of the parking spaces. The processing equipment performs distortion correction, noise reduction, and brightness compensation on the images. Then, it uses background subtraction, empty parking space template comparison, or vehicle target recognition methods to extract the vehicle area. The system determines whether the parking space is empty or occupied based on the overlap between the vehicle area and the parking space image area. It also suppresses state transitions caused by people passing by and vehicles briefly stopping by by counting multiple consecutive frames. Some systems also place geomagnetic detectors on the parking space ground to obtain visual occupancy results and geomagnetic occupancy results. The final state is then output using a fixed priority, logical AND-OR relationship, or result voting method. The visual background template and geomagnetic environmental baseline are updated independently according to their respective collected data.

[0003] The main drawback of the above-mentioned identification method is that visual observation and geomagnetic observation are only combined in parallel at the final judgment stage. The two do not establish a verifiable temporal correspondence around the same vehicle entry, parking, or departure event. When the visual area changes due to shadows, reflections, personnel obstruction, or camera shake, the system has difficulty using geomagnetic changes to determine whether the visual event is real. When the environmental magnetic field drifts or adjacent vehicles cause geomagnetic disturbances, it is also impossible to use vehicle trajectories to determine whether the geomagnetic event originated from the corresponding parking space. When the two results conflict, fixed priority or voting rules are prone to incorrectly switching the parking space status. Independently updated visual background and geomagnetic baseline may also write contaminated data in the wrong idle state, causing misjudgments to accumulate continuously in subsequent identification processes. Therefore, it is necessary to solve the problem of unstable occupancy status caused by the lack of event-level causal mutual verification between visual and geomagnetic information. Summary of the Invention

[0004] The purpose of this invention is to provide a charging parking space occupancy recognition system based on the fusion of AI vision and geomagnetism, which can solve the problems in the background art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion includes an image acquisition module and a geomagnetic detection module set for the corresponding charging parking space, and a processing module that communicates with the two. The processing module establishes a parking space image area and caches continuous images, extracts geomagnetic state transition events and corresponding times from the geomagnetic sequence of the geomagnetic detection module, and extracts the visual sequence related before and after the corresponding times. The visual sequence is used to identify vehicles and correlate cross-frame trajectories using an AI visual model to obtain the vehicle's crossing direction and stopping relationship relative to the boundary of the parking space image area. The geomagnetic state transition event is used to generate a parking space occupancy status by corresponding the crossing direction and the dwelling relationship with the time sequence of the vehicle entering the space or occupying the space and leaving the space. The visual baseline of the empty parking space and the geomagnetic environment baseline are updated when the geomagnetic sequence and the visual sequence jointly represent that the parking space is vacant.

[0006] Preferably, the processing module performs baseline separation on the geomagnetic sequence according to the geomagnetic environment baseline, determines the disturbance starting point based on the consistency of the vector change direction of continuous sampling relative to the geomagnetic environment baseline, and identifies the continuous segment that deviates from the geomagnetic environment baseline or the recovery segment that returns to the geomagnetic environment baseline from the disturbance starting point. The boundary between the disturbance initiation point and the sustained or recovered segment is taken as the corresponding moment of the geomagnetic state transition event, and the start and end anchor points of the visual sequence are determined respectively. The preceding and following related images covering the start and end anchor points are extracted from the cached continuous images.

[0007] Preferably, the processing module performs vehicle instance segmentation on the preceding and following images, and associates the mask of the same vehicle as the target trajectory based on the regional overlap continuity and motion direction of the vehicle instances in adjacent images. The boundary of the parking space image area is divided into an entrance side and an inner side. The order in which the target trajectory crosses from the entrance side to the inner side or in the opposite direction is recorded. The dwelling relationship is generated based on the spatial inclusion relationship between the instance mask at the end of the target trajectory and the parking space image area.

[0008] Preferably, the processing module stores a state transition table, which is associated with the current confirmed state, the geomagnetic state transition event, the boundary crossing order of the target trajectory, and the candidate state. When the geomagnetic state transition event and the boundary crossing sequence conform to the temporal correspondence recorded in the state transition table, the candidate state is written into the parking space occupancy state. When the two do not conform to the temporal correspondence, a pending confirmation record is established. The pending confirmation record is associated with geomagnetic sequence segments and target trajectories within the same anchor point range and retains the current confirmation state.

[0009] Preferably, the baseline separation uses the most recent geomagnetic segment characterized by parking space vacancy by both geomagnetic and visual sequences as the reference segment, and generates an environmental trajectory covering the geomagnetic state transition event along the geomagnetic change trend before the disturbance initiation point. The vector difference between the current geomagnetic sequence and the environmental trajectory is used to form a vehicle disturbance sequence. The start and end boundaries of the vehicle disturbance sequence are used to correct the corresponding time. After the recovery segment is formed, the continuous change between the recovery segment and the reference segment is used to correct the environmental trajectory.

[0010] Preferably, when the target trajectory is interrupted between adjacent images, the processing module maps the motion direction of the instance mask before the interruption and the appearance position of the instance mask after the interruption to the parking space image area, and constrains the candidate association with the time range from the start of the disturbance to the continuous segment; Merge the preceding and following trajectory segments that conform to the boundary crossing sequence and correspond to the same geomagnetic state transition event to form a target trajectory that includes markers for the occluded sections.

[0011] Preferably, the processing module identifies modal conditions in the record to be confirmed that do not meet the state transition table, expands the visual sequence with the corresponding time of the geomagnetic state transition event adjacent to the modal condition, and reads the geomagnetic sequence after the corresponding time period; The expanded visual sequence and the geomagnetic sequence are sequentially spliced ​​with the record to be confirmed. When the splicing result satisfies the state transition table, the corresponding candidate state is written. When the subsequent geomagnetic state transition event is completed before the modal condition is met, the record to be confirmed is transferred to an abnormal event record.

[0012] Preferably, after the state transition table completes the boundary crossing from the inner side of the parking space image area to the entrance side based on the target trajectory and the geomagnetic sequence enters the recovery segment, updating the current confirmed state from occupied to idle, the processing module determines the common idle period that simultaneously covers the visual sequence and the geomagnetic sequence. The image of the empty parking space visual baseline is updated by extracting the image of the mask of the non-vehicle instance from the common idle period, and the geomagnetic environment baseline is updated with the recovery segment corresponding to the time. The two updated data are bound to the same baseline version identifier.

[0013] Preferably, the processing module configures trajectory identifiers for target trajectories in the expanded visual sequence and configures anchor point identifiers for geomagnetic state transition events in the geomagnetic sequence; When multiple target trajectories or multiple geomagnetic state transition events exist within the same extended time period, corresponding records of the trajectory identifier and the anchor point identifier are established according to the order in which the target trajectory crosses the boundary of the parking space image area and the order in which the geomagnetic state transition events occur. The target trajectories and geomagnetic state transition events that form corresponding records are sent to the state transition table, and the data that does not form corresponding records are retained in the abnormal event record.

[0014] Preferably, the processing module combines the updated visual baseline of the empty parking space and the geomagnetic environment baseline into a candidate baseline version. Before replacing the current baseline version, the module calls the candidate baseline version to reprocess the associated visual sequence and geomagnetic sequence after the common idle period. The candidate baseline version is enabled when the reprocessed boundary crossing sequence and the geomagnetic state transition event still satisfy the state transition table; otherwise, the current baseline version is retained, and the candidate baseline version and the corresponding common idle time period are written into the baseline contamination record.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts geomagnetic state transition events and their corresponding times from geomagnetic sequences, and extracts related visual sequences before and after the corresponding times, so that the visual processing range corresponds to the actual time period when the vehicle causes changes in the magnetic field; the artificial intelligence visual model obtains the vehicle's cross-frame trajectory, parking space boundary crossing direction, and dwelling relationship from the related visual sequences, and the state transition table then pairs the visual trajectory with the geomagnetic state transition events according to the time sequence of the vehicle entering to occupying or occupying to leaving; when shadows, people passing by, etc. only produce image changes without corresponding geomagnetic events, or when environmental magnetic field changes do not produce vehicle boundary crossing trajectories, the relevant data enters the pending confirmation record and retains the current confirmation state, thus reducing the situation where a single observation anomaly directly triggers an erroneous state switch and improving the continuity of charging parking space occupancy results under conditions of shading, changes in lighting, and magnetic field disturbances.

[0016] 2. This invention determines a common idle period when both the visual sequence and the geomagnetic sequence characterize the vacancy of a parking space. Images without vehicle instance masks are extracted from the corresponding period to update the visual baseline of the vacant parking space, and the geomagnetic environment baseline is updated with the geomagnetic recovery segment corresponding to the time. The two updated data are bound to the same baseline version. Candidate baseline versions reprocess subsequent associated data before replacing the current version. Candidate data that fails to maintain the correspondence between the visual boundary crossing order and geomagnetic state transition events are written into the baseline contamination record. The geomagnetic environment trajectory can also separate vehicle disturbance sequences from the geomagnetic sequence. The time range defined by geomagnetic events can constrain the merging of trajectory segments before and after occlusion, thereby reducing the continuous impact of erroneous idle sample writing, gradual changes in the environmental magnetic field, and short-term visual occlusion on subsequent recognition, and reducing the computational load of continuously processing video frames without events. Attached Figure Description

[0017] Figure 1 A flowchart for the fusion recognition process of charging parking space occupancy status; Figure 2 This is a graph showing the distance of geomagnetic disturbance and the deviation of the triaxial baseline. Figure 3 A graph showing the relationship between the overlap ratio of parking spaces and the area of ​​the region. Figure 4 This is a flowchart of the process for generating and replaying paired baselines for visual and geomagnetic data. Detailed Implementation

[0018] In one embodiment, please refer to Figure 1 A charging parking space occupancy recognition system based on the fusion of artificial intelligence vision and geomagnetism includes an image acquisition module and a geomagnetic detection module corresponding to the charging parking space, as well as a processing module that communicates with both. The processing module establishes a parking space image region and caches continuous images. It extracts geomagnetic state transition events and corresponding times from the geomagnetic sequence of the geomagnetic detection module. Based on the corresponding times, it extracts the visual sequence that is related to the previous and subsequent times. It uses an artificial intelligence vision model to perform vehicle recognition and cross-frame trajectory association on the visual sequence to obtain the crossing direction and dwelling relationship of the vehicle relative to the boundary of the parking space image region. It generates the parking space occupancy status by corresponding the geomagnetic state transition events with the crossing direction and dwelling relationship according to the time sequence of the vehicle entering to occupying or occupying to leaving. When the geomagnetic sequence and the visual sequence jointly represent the parking space being vacant, it updates the visual baseline of the empty parking space and the geomagnetic environment baseline. This ensures that each state change corresponds to the continuous observation of the same vehicle event in the image domain and the magnetic field domain, rather than being triggered independently by any detection result.

[0019] Specifically, the image acquisition module is installed in a location that can cover the entrance boundary and parking area of ​​the target charging space, while the geomagnetic detection module is set within the influence range of vehicles parked in the same charging space. The processing module receives the frame images generated by the image acquisition module and the multi-axis magnetic field sampling values ​​generated by the geomagnetic detection module, and establishes a parking space identifier for the same charging space. The parking space identifier is simultaneously associated with the image acquisition channel, the geomagnetic acquisition channel, the parking space image area, the visual baseline of the empty parking space, and the geomagnetic environment baseline. The processing module writes the two types of data into independent but time-consistent buffer queues according to the parking space identifier. The image buffer queue stores the frame identifier, acquisition time, and image data, while the geomagnetic buffer queue stores the sampling time, magnetic field vector, and baseline version identifier, so that the same parking space and the same time range can be located when subsequently capturing visual sequences, tracing geomagnetic events, and updating paired baselines.

[0020] In this embodiment, the processing module performs unified clock correction on the image acquisition module and the geomagnetic detection module. Unified clock correction does not require the two types of modules to have the same sampling period. Instead, the processing module records their original timestamps and converts them to a unified time axis. When an image frame is located between adjacent geomagnetic sampling times, the actual acquisition time of the image frame is retained, and the original image is not replaced by interpolated data. When the time corresponding to a geomagnetic event falls between adjacent image frames, the nearest frame before the event time and the nearest frame after the event time are selected from the image buffer queue as anchor frames. The processing module also records the difference between the arrival time and the acquisition time caused by the communication delay, and performs cross-modal association based on the acquisition time to avoid the network transmission order changing the real time order of vehicle events.

[0021] Preferably, the parking space image area is calibrated using the actual empty parking space image. The processing module determines the parking space entrance boundary, side boundaries, and end boundary in the image, and saves the polygon formed by each boundary as the parking space image area. When the camera has a top-down perspective, the mapping relationship between the image coordinates and the parking plane coordinates is established using the parking space corner points on the ground. The vehicle instance mask, vehicle trajectory position, and boundary crossing position are all converted through the same mapping relationship. After the camera position changes due to maintenance, the processing module recalculates the mapping relationship using parking space markings or fixed ground features, and freezes the writing of the empty parking space visual baseline before the mapping relationship is confirmed, so that the background area under the old coordinates will not be mixed with the parking space area under the new coordinates.

[0022] In a preferred embodiment, the processing module generates an image quality record for each frame. The image quality record includes the visibility status of the parking space boundary, the offset status of fixed ground features, and the occlusion status of the parking space area. When the fixed ground features undergo an overall displacement relative to the current mapping relationship, the corresponding image is marked as an image to be verified and the generation of new boundary crossing records is stopped. When partial occlusion covers the entrance boundary but the interior of the parking space is still visible, the vehicle instance and the internal dwell observation are retained and the entrance crossing conditions are written into the record to be verified. When the parking space area is not visible as a whole, only the original frame index and the occlusion interval are saved. After the image is restored, the camera position is checked by the fixed ground features before and after occlusion to see if the position has changed. If the position has not changed, the original mapping relationship continues. If the position has changed, the parking space image area is recalibrated and a new mapping version is established.

[0023] Furthermore, continuous images are stored using a circular caching method. The processing module maintains a frame index for each parking space, sorted according to the acquisition time. The frame index points to the original image or a copy of the image after uniform scale transformation. When the geomagnetic state has not changed, only the image covering the most recent observation period is retained. After detecting a geomagnetic state transition event, the frame index before and after the corresponding time of the event is locked and a related visual sequence is formed. Images within the locked range are not overwritten by the circular cache before the state transition determination ends. If the duration of the event exceeds the current cache range, the processing module continuously expands the frame index according to the new geomagnetic sampling without repeatedly copying existing images, thereby maintaining the temporal continuity of the visual sequence and ensuring that one event corresponds to only one traceable image data.

[0024] In a preferred embodiment, the artificial intelligence visual model includes a vehicle feature extraction network, a vehicle instance segmentation output terminal, and a cross-frame feature association terminal. The input image is fed into the vehicle feature extraction network after color normalization, local brightness correction, and parking space region cropping. The vehicle feature extraction network retains the vehicle body outline, wheel adjacent area, and vehicle lower edge structure from feature maps of different scales. The vehicle instance segmentation output terminal forms an instance mask for pixels belonging to the same vehicle and provides instance confidence information. The cross-frame feature association terminal does not directly use the single-frame detection box as the occupancy basis, but establishes the target trajectory based on the spatial overlap of instance masks in adjacent frames, the direction of motion continuation, and the consistency of appearance features. The processing module only sends vehicle instances that have motion association with the boundary of the target parking space entrance to the subsequent state transition determination.

[0025] Preferably, the training samples of the artificial intelligence visual model consist of historical images of the corresponding charging parking space and similar parking environment images. The sample annotations at least distinguish between vehicle instances and non-vehicle areas, and retain images where vehicles are partially occluded by charging piles, pillars, neighboring vehicles, or parking space obstructions. During training, the instance labels of the same vehicle in consecutive frames remain consistent, enabling the cross-frame feature association end to learn the appearance relationship between the complete visible state and the partially visible state. New images generated after deployment are not directly written into the training samples. Only vehicle trajectories confirmed by both geomagnetic state transition events and state transition tables form candidate annotation data. The candidate annotation data is used for model updates after offline verification to avoid visual misjudgments and is automatically fed back into subsequent model versions.

[0026] In this embodiment, the artificial intelligence visual model maintains the original temporal order of the associated visual sequences during inference. The processing module uses the pre-event sequence to extract the texture of empty parking spaces and moving targets near the entrance, uses the anchor point proximity sequence to extract the changes in vehicle outlines from the outside to the inside of the parking space, and uses the post-event sequence to determine whether vehicle instances form a stable stop. The model retains the appearance features of the complete image for vehicle instances clipped by the parking space boundary, and then restricts the instance mask to the vicinity of the target parking space to participate in trajectory association. When pedestrians, bicycles, or hand-push devices appear in the image at the same time, the category output separates non-motorized vehicle instances from vehicle instances. The processing module does not send areas with only a local rectangular appearance but lacking continuous vehicle outlines to the boundary crossing judgment. When the model outputs a vehicle instance with low confidence but continuous with the existing trajectory position, it is retained as a trajectory candidate and handed over to subsequent geomagnetic event association for confirmation.

[0027] In this embodiment, the geomagnetic detection module outputs magnetic field components collected along mutually orthogonal directions. The processing module performs timestamp sorting, abnormal missing measurement marking, and baseline normalization on each component. Abnormal missing measurement locations are left empty without using data from adjacent vehicle events to fill in the gaps. The geomagnetic environment baseline is formed by continuous sampling within confirmed idle periods. The baseline stores the center value of each direction, the range of natural fluctuations, and the trend of slow change over time. When the current sample arrives, it is first compared with the current baseline version by vector, and then it is determined whether the change has a continuous directional correlation. Even if a single isolated sample deviates from the baseline, it is only recorded as a transient point. Only when subsequent samples form a continuous disturbance trajectory with it is a geomagnetic state transition event established.

[0028] Furthermore, the processing module performs channel integrity checks on multi-axis geomagnetic sampling. If there are missing measurements in some directions at the same sampling time, the original values ​​of the other directions are retained but the complete vector disturbance is not calculated. When the subsequent supplementary data has the original acquisition time, it is backfilled to the corresponding sampling position and the event continuity is rechecked. If a certain direction remains unchanged for a long time while the other directions change continuously with the movement of the vehicle, the unchanged direction is marked as a channel abnormality and a geomagnetic segment to be confirmed is formed with the available direction. The geomagnetic segment to be confirmed must be established to correspond with the complete visual entry or exit trajectory before it enters the state transition table. After all directions return to normal and the parking spaces are collectively represented as empty, the new common empty geomagnetic segment can participate in the generation of candidate geomagnetic baselines. Sampling during the channel abnormality period will not replace the original geomagnetic environment baseline.

[0029] In a preferred embodiment, the processing module performs baseline separation on the geomagnetic sequence according to the geomagnetic environment baseline, determines the disturbance starting point based on the consistency of the vector change direction of continuous sampling relative to the geomagnetic environment baseline, and identifies the continuous segment that deviates from the geomagnetic environment baseline or the recovery segment that returns to the geomagnetic environment baseline from the disturbance starting point. The processing module takes the boundary between the disturbance starting point and the continuous segment or the recovery segment as the corresponding time of the geomagnetic state transition event, and determines the starting anchor point and the ending anchor point of the visual sequence respectively. Then, it extracts the related images before and after covering the starting anchor point and the ending anchor point from the cached continuous images, so that the driving in event has an idle image before the disturbance, a disturbance formation image and a parking image after the disturbance, and the driving out event has an occupied image before recovery, a recovery process image and an idle image after recovery.

[0030] Specifically, baseline separation uses the idle geomagnetic segment corresponding to the current baseline version as a reference. The processing module calculates the difference direction of the current sampling vector relative to the reference vector. When the difference direction of continuous sampling remains in the same direction or gradually rotates and the difference amplitude increases continuously in adjacent time moments, the earliest sampling position that satisfies the continuity relationship is traced back as the disturbance starting point. When the difference amplitude stops increasing but still deviates from the baseline, a continuous segment is formed. When the difference vector converges towards the baseline and re-enters the natural fluctuation range, a recovery segment is formed. When a sudden change in direction occurs during the disturbance process and the previous and subsequent samples cannot form a continuous vector trajectory, the processing module separates the sudden change sample from the current event and retains it as an anomaly point to prevent instantaneous electromagnetic interference from truncating the real vehicle event.

[0031] Furthermore, the amount of geomagnetic disturbance is calculated according to the following formula: ; in, This represents the dimensionless geomagnetic disturbance at sampling time t. This represents the component of the current geomagnetic vector in the j-direction after baseline normalization. This represents the baseline component of the environmental trajectory in the j-direction at the same moment, where j takes the values ​​x, y, and z in sequence. This represents the sum of the squared differences of the components in the three directions. The square root represents the sum of the three-axis differences into a vector distance. For example, if the current vector value is (0.66, 0.34, 0.58) and the environmental trajectory value is (0.20, 0.14, 0.22), Approximately 0.617. If the upper bound of the natural fluctuation of the corresponding reference segment is 0.09, and the difference vector of subsequent sampling maintains continuous directional correlation, the processing module will include the current moment in the disturbance starting point candidate, thereby retaining the common changes of multiple axes as vehicle disturbances and excluding isolated jump points that only occur on a single axis.

[0032] Please refer to Figure 2 , Figure 2The horizontal axis represents the sampling phases on a unified time axis, with consecutive sampling windows labeled start, entry, disturbance, pause, recovery, and end. The vertical axis represents the triaxial deviation and geomagnetic disturbance distance after baseline separation. The thick black solid line represents the geomagnetic disturbance distance. Its input comes from the current geomagnetic vector components. baseline components of the environmental trajectory at the same time The difference; the dashed line, dotted line and gray solid line correspond to the horizontal, vertical and longitudinal baseline deviations respectively, and the three component curves together constitute the vector input when j takes x, y and z.

[0033] from Figure 2 It can be seen that in the initial stage, the curves fluctuate slightly around the zero value, indicating that the current sampling is basically consistent with the environmental geomagnetic baseline; after entering the stage, the three-axis deviation rises synchronously, the disturbance distance increases rapidly, and a continuous high-value band is formed in the disturbance stage. Figure 2 The peak value of the thick black line is significantly higher than that of the single-directional component, reflecting... The combined effect of the differences in the three directions. In the first event, the rise of the lateral and vertical components is more obvious, while the longitudinal component continues to change within the high-value band; the three-axis combination of the second event is not exactly the same as that of the first event, but it also went through the complete process of entry, continuous perturbation and recovery.

[0034] Preferably, the upper bound of natural fluctuations is calculated from the common idle period corresponding to the current baseline version, and is not used as a fixed value applicable to all parking spaces. The processing module determines the normal fluctuation envelope according to the distribution of geomagnetic disturbances during the idle period, and retains the slow moving trend of the envelope over time. After the candidate for the disturbance starting point appears, the continuity of the vector difference direction and the change of the disturbance amount is checked. When the continuity relationship is established, the sampling that first leaves the normal fluctuation envelope is traced back as the disturbance starting point. When subsequent sampling immediately returns to the envelope and there is no corresponding vehicle trajectory, the candidate is canceled and the transient point is saved. When subsequent sampling forms a continuous segment, the candidate is locked and an anchor point is generated. After the recovery segment is completed, the event boundary is recalculated using the corrected environmental trajectory. When the boundary moves, the frame index range of the associated visual sequence is adjusted synchronously without changing the original acquired data.

[0035] In this embodiment, the starting anchor point is not the same as the single image frame where the disturbance originates. The processing module traces backward from the disturbance origin to the period when the geomagnetic sequence maintains the baseline state and obtains several consecutive images corresponding to that period as the pre-event sequence. It extends backward from the boundary of the continuous segment or the boundary of the recovery segment to the period when the visual trajectory enters a stable state and obtains the corresponding images as the post-event sequence. The pre-event sequence, the anchor point neighbor sequence, and the post-event sequence are merged according to the acquisition time. The repeated frame index is only retained once. When there is a communication gap, a gap mark is written into the sequence without copying adjacent frames, so that the cross-frame trajectory association end can adjust the candidate displacement range according to the actual time interval.

[0036] Furthermore, the entry event is represented by the geomagnetic sequence transitioning from the baseline state to a continuous deviation state, and the exit event is represented by the geomagnetic sequence transitioning from a continuous deviation state to a recovery state. The processing module does not directly infer entry or exit based on the positive or negative direction of the geomagnetic difference vector, because the vehicle chassis material and parking direction will change the direction of the magnetic field components. Instead, it determines the candidate type of the geomagnetic state transition event by combining the current confirmed state before the event occurs. If continuous disturbance occurs when the current confirmed state is idle, an entry candidate is formed. If a recovery process occurs when the current confirmed state is occupied, an exit candidate is formed. The candidate type still needs to be matched with the visual boundary crossing sequence before the parking space occupancy status can be changed.

[0037] In a preferred embodiment, the processing module performs vehicle instance segmentation on the preceding and following images. Based on the regional overlap continuity and movement direction of vehicle instances in adjacent images, the instance masks belonging to the same vehicle are associated as target trajectories. The processing module divides the boundary of the parking space image region into an entrance side and an inner side, records the order in which the target trajectory crosses from the entrance side to the inner side or in the opposite direction, and generates a dwell relationship based on the spatial inclusion relationship between the instance mask at the end of the target trajectory and the parking space image region. The target trajectory thus simultaneously includes the vehicle instance identifier, instance masks of each frame, acquisition time, boundary side, and end spatial state. The state transition determination can read the trajectory formation process and not just the last frame occupancy label.

[0038] Specifically, the vehicle instance segmentation output terminal generates vehicle candidate masks for each frame of image. The processing module first deletes candidate masks located outside the parking space image area and whose movement direction is far from the entrance boundary. Then, it maps the remaining candidate masks to the parking plane. The association cost between candidate masks in adjacent frames is composed of the degree of overlap of the mapping area, the reachability of the center position, and the distance of appearance features. Candidate association needs to satisfy that the vehicle movement direction is continuous in time and will not cross with another existing trajectory. When short-term occlusion causes no candidate mask in the current frame, the trajectory prediction position is retained. The reappearing candidate mask is only connected to the original target trajectory if it is located within the predicted movement range and maintains appearance connection.

[0039] Preferably, the entrance side is formed by the side of the parking space entrance boundary facing the driving lane, and the inner side is formed by the side of the entrance boundary facing the end of the parking space. The processing module performs lateral identification on the representative position of the lower edge of the vehicle instance mask in the target trajectory. When the representative position enters the boundary adjacent area from the entrance side and continues to enter the inner side, it is recorded as entering and crossing. When the representative position enters the boundary adjacent area from the inner side and continues to reach the entrance side, it is recorded as leaving and crossing. Vehicles that only move laterally near the boundary but do not complete the lateral identification change do not form a complete crossing record. When the instance mask at the end of the target trajectory is continuously located in the parking space image area and the motion displacement enters a stable state, a dwell relationship is generated. When the target trajectory continues to drive out of the parking space image area after crossing, no dwell relationship is generated.

[0040] In this embodiment, the processing module stores a state transition table, which associates the current confirmed state, geomagnetic state transition events, the boundary crossing order of the target trajectory, and candidate states. When the geomagnetic state transition event and the boundary crossing order match the temporal correspondence recorded in the state transition table, the processing module writes the candidate state into the parking space occupancy state. When the two do not match the temporal correspondence, a pending confirmation record is created. The pending confirmation record is associated with geomagnetic sequence segments and the target trajectory within the same anchor point range and retains the current confirmed state. Each record in the state transition table stores the allowed previous state, geomagnetic event type, visual trajectory event, dwell relationship requirements, and corresponding subsequent state, thus restricting data without a complete event chain from changing the parking space state at the data structure level.

[0041] Furthermore, when the current confirmed state is idle, the geomagnetic event is a continuous disturbance, the visual trajectory completes the entry and crossing and the end forms a stopping relationship, the state transition table gives the occupancy candidate state. When the current confirmed state is occupied, the geomagnetic event is a recovery after continuous deviation, the visual trajectory completes the departure and crossing and the end leaves the parking space image area, the state transition table gives the idle candidate state. When a vehicle passes near the entrance but does not enter the inner side, the visual trajectory does not meet the entry and crossing conditions. Even if there is a geomagnetic disturbance propagating from an adjacent parking space, it will not be written into the occupancy candidate state. When a vehicle enters the parking space and immediately exits, the processing module forms entry and departure events sequentially according to the same target trajectory and retains the intermediate short-term occupancy record, without splitting the complete round-trip process into unrelated anomalies.

[0042] Specifically, the reading of the state transition table and the writing of the parking space confirmation status use the same event version control. When the processing module starts the judgment, it reads the current state version and binds the geomagnetic anchor point identifier, target trajectory identifier, and candidate status to the event version. When a new visual trajectory or a new geomagnetic state transition event occurs during the judgment, the new data enters the next event container without changing the data set being judged. Before writing the candidate status, the current state version is checked again. If the versions are consistent, the state writing is completed and the event version is incremented. If the version has been updated by an earlier closed event, the current writing is withdrawn and the state transition table is queried again according to the new previous state. This makes the continuous entry, short stay, and departure events on the same parking space connected in real time order, avoiding the parallel processing results from overwriting the already confirmed status due to the order in which they arrive at the processing module.

[0043] In a preferred embodiment, the record to be confirmed is stored in an event container. The event container contains the parking space identifier, the current confirmation status, the geomagnetic event identifier, the anchor point time period, the visual trajectory identifier, and the unmet state transition conditions. When the visual image is occluded and lacks a complete boundary crossing record, the geomagnetic sequence segment and the previous segment of the formed trajectory are still stored in the same event container. When the geomagnetic field is disturbed and there is no corresponding vehicle in the visual image, the geomagnetic event will not be directly discarded. The processing module sets a data closure flag for the record to be confirmed. The recording ends only when the data is supplemented to make the event chain complete or when subsequent events prove that the current record cannot be closed. Before the recording ends, the output terminal continues to publish the most recently confirmed idle or occupied status and attaches the confirmation flag.

[0044] Specifically, the closure determination of the record to be confirmed is based on the missing conditions in the state transition table. The processing module saves the last supplementable time and acceptable data type for each missing condition. When the visual end dwell relationship is missing, only the subsequent instance mask of the same trajectory identifier is received. When the geomagnetic recovery segment is missing, only the geomagnetic sampling of the same anchor point identifier is received. When a new trajectory or new anchor point appears that is opposite to the current record direction and cannot be classified into the original event in time, the expansion of the original record is stopped. The original record is converted into an abnormal event and a subsequent event container is established by the new data. The already closed record is set with a read-only flag. Subsequent baseline playback can recalculate the visual and geomagnetic intermediate data in it, but cannot directly rewrite the original confirmation time and the original acquisition index.

[0045] In a preferred embodiment, baseline separation uses the most recent geomagnetic segment representing parking space vacancy, jointly characterized by the geomagnetic sequence and visual sequence, as the reference segment. The processing module generates an environmental trajectory covering geomagnetic state transition events along the geomagnetic change trend before the disturbance initiation point. The vector difference between the current geomagnetic sequence and the environmental trajectory forms a vehicle disturbance sequence. The start and end boundaries of the vehicle disturbance sequence are used to correct the corresponding time. After forming the recovery segment, the continuous change between the recovery segment and the reference segment is used to correct the environmental trajectory. The environmental trajectory is used to express the gradually changing background formed by temperature changes, changes in the working status of surrounding fixed equipment, and the movement of distant magnetic objects. The vehicle disturbance sequence is used to express the local continuous changes formed by the entry, stay, and departure of the metal vehicle body above the corresponding parking space.

[0046] Specifically, the processing module obtains the direction of change of each axis of geomagnetic components from the end of the reference segment, and uses the median change of the difference between adjacent samples within the reference segment to form the initial environmental trajectory during the disturbance period. The initial environmental trajectory extends with time but does not follow the rapid shift of the vehicle disturbance sequence. After the geomagnetic event is completed and a recovery segment is formed, the processing module compares the difference between the end of the recovery segment and the initial environmental trajectory at the same time, and smoothly distributes the difference along the event duration interval to correct the environmental trajectory. Then, the corrected environmental trajectory is subtracted from the original geomagnetic sequence to obtain the vehicle disturbance sequence. If the corrected vehicle disturbance sequence still maintains the complete start point, duration interval and recovery boundary, its boundary is used to replace the initial corresponding time. If the event has not yet formed a recovery segment, the initial corresponding time is retained and the environmental trajectory is marked as a state to be corrected.

[0047] In a preferred embodiment, when the target trajectory is interrupted between adjacent images, the processing module maps the motion direction of the instance mask before the interruption and the appearance position of the instance mask after the interruption to the parking space image area, and constrains the candidate association with the time range from the start of the disturbance to the continuous segment. The processing module merges the preceding and following trajectory segments that conform to the boundary crossing sequence and correspond to the same geomagnetic state transition event to form a target trajectory containing occlusion segment markers. The occlusion segment markers record the start and end times of the trajectory when it is invisible, the motion direction before it is invisible, the reappearance position, and the corresponding geomagnetic event identifier. The subsequent state transition determination reads the merged complete trajectory and the occlusion segment, without interpreting the same vehicle before and after the occlusion as two independent vehicles.

[0048] Specifically, the candidate association after trajectory interruption is not determined by a single spatial distance. The processing module calculates the range of the vehicle's movement direction on the parking plane from several visible frames before the interruption, extends this range to the occluded area to form an accessible zone, and then limits the time range corresponding to the vehicle passing through the occluded area by the geomagnetic disturbance start point and the continuous segment boundary. When the appearance position of the instance mask after the interruption is within the accessible zone, the appearance time is within the corresponding time range, and the appearance features are associated with the instance before the interruption, a continuation candidate is formed. When there are multiple continuation candidates, priority is given to retaining the candidate that can complete the crossing from the entrance side to the inner side boundary and form a time sequence corresponding with the same geomagnetic event. The remaining candidates are retained in the original target trajectory without being forcibly merged.

[0049] In a preferred embodiment, the processing module identifies modal conditions in the record to be confirmed that do not meet the state transition table. It expands the visual sequence at the time corresponding to the geomagnetic state transition event adjacent to the modal condition and reads the geomagnetic sequence after the corresponding time period. The processing module splices the expanded visual sequence and geomagnetic sequence with the record to be confirmed in a temporal sequence. When the splicing result meets the state transition table, it writes the corresponding candidate state. When a subsequent geomagnetic state transition event is completed before the modal condition, the record to be confirmed is transferred to an abnormal event record. The expansion operation is performed around the original event container and does not recreate independent events competing with the same geomagnetic event, so that the supplementary image, supplementary trajectory, and supplementary geomagnetic sampling always belong to the original record to be confirmed.

[0050] Furthermore, unmet modal conditions are obtained by comparing each item in the state transition table. When the geomagnetic event is complete but the visual trajectory lacks an end-pointing relationship, the processing module extends the visual sequence backward until the vehicle stops, the vehicle leaves the frame, or a new geomagnetic event occurs. When the visual entry trajectory is complete but the geomagnetic sequence has not yet formed a continuous segment, the processing module continues to read subsequent geomagnetic samples until a continuous segment is formed or the baseline is restored. The extended data and the original data are sorted according to the acquisition time. The same frame identifier and the same geomagnetic sample identifier are only retained once. The trajectory segment re-performs identity association at the splicing boundary. The geomagnetic segment re-checks the vector direction continuity at the splicing boundary. If the necessary conditions are still missing after splicing and a new reverse event has occurred, the original record is terminated and an abnormal event type is written.

[0051] Furthermore, the spatial relationship between the vehicle instance mask and the parking space image region is calculated according to the following formula: ; in, This represents the spatial overlap ratio between the vehicle instance in frame u and the target parking space. Let C represent the area of ​​the vehicle instance mask mapped onto the parking plane in frame u, and let C represent the area of ​​the target parking space image region mapped onto the parking plane. The intersection symbol represents the area shared by two regions, and the union symbol represents the area after merging the two regions. The absolute value symbol represents the pixel area or mapped area of ​​the corresponding region. For example, when the vehicle instance region has 5400 mapped pixels, the parking space region has 9600 mapped pixels, the shared area is 5000 mapped pixels, and the merged area is 10000 mapped pixels, then... If the value is set to 0.50, and the value of subsequent frames increases continuously and the vehicle's representative position moves from the entrance side to the interior side, the processing module will write this change together with the entry and crossing record into the dwell relationship, thereby distinguishing between the short-term local overlap formed along the entrance and the continuous spatial inclusion formed after entering the parking space.

[0052] Please refer to Figure 3 , Figure 3 The horizontal axis represents the sampling stage on a unified time axis, the left side of the vertical axis represents the area mapped onto the parking plane, and the right side of the vertical axis represents the spatial overlap ratio. The thick black solid line represents the spatial overlap ratio. The black square dashed lines represent the area of ​​the vehicle zone, the gray circular dotted lines represent the intersection area of ​​the vehicle zone and the target parking space zone, and the gray solid lines represent the union area of ​​the two zones. Each line is labeled with... Figure 3 The curves with the same name in the middle are all corresponding.

[0053] from Figure 3 The two events show that during the entry phase, the vehicle area, intersection area, and spatial overlap ratio all increase simultaneously. During the disturbance or stationary phase, these values ​​fluctuate continuously within a higher range. During the recovery phase, the intersection area decreases and eventually returns to a near-empty level. If a vehicle only brushes past the parking space boundary, the intersection may briefly show a non-zero value, indicating insufficient stationary relationship; if the vehicle is obscured... The boundaries may jitter in adjacent frames, but the trajectory can still be preserved in consecutive frames.

[0054] In this embodiment, the dwell relationship is jointly formed by the change in spatial overlap ratio, the side of the vehicle's representative position, and the motion state of the target trajectory end. The processing module does not directly confirm dwell using the single-frame spatial overlap ratio. When a vehicle passes diagonally from the driving lane through the parking space entrance, the spatial overlap ratio may increase briefly, but the representative position does not continuously enter the inner side and the trajectory end still maintains the departure direction. Therefore, only a passing event is formed. When the vehicle reverses into the charging parking space, the vehicle's rear instance crosses the entrance boundary first. The processing module retains the driving direction based on the continuous migration of the representative position from the entrance side to the inner side, and does not replace the movement direction with the vehicle's front orientation. When the instance mask changes area due to local occlusion after the vehicle is parked, as long as the representative position remains on the inner side and the displacement of the target trajectory end remains stable, the dwell relationship continues to be associated with the original vehicle instance. The geomagnetic continuous segment provides independent observation that the vehicle is still within the detection range.

[0055] In a preferred embodiment, please refer to Figure 4 After the state transition table completes the boundary crossing from the inner side of the parking space image area to the entrance side based on the target trajectory and the geomagnetic sequence enters the recovery segment, updating the current confirmed state from occupied to idle, the processing module determines the common idle period that simultaneously covers the visual sequence and the geomagnetic sequence. The processing module extracts images without vehicle instance masks from the common idle period to update the visual baseline of the empty parking space, and updates the geomagnetic environment baseline with the recovery segment corresponding to the time. Then, the two updated data are bound to the same baseline version identifier. The starting point of the common idle period is taken as the later moment after the visual trajectory completes the departure and the geomagnetic recovery enters the stable range, and the ending point is taken as the earlier moment before the next geomagnetic disturbance or the next vehicle trajectory approaches the parking space entrance.

[0056] Specifically, the visual baseline for empty parking spaces does not replace the original baseline with a single frame image from the common idle period. The processing module aligns the images that do not have vehicle instance masks and whose parking space boundary mapping relationship remains stable during the common idle period according to their positions. It establishes the temporal distribution of brightness and color for each pixel position. Outliers formed by short-term shadows, light sweeps, and local passing of people are not written into the stable background. Stable parking space markings, ground textures, and edges of fixed charging facilities are written into the candidate visual baseline. The candidate visual baseline retains the source time period and camera mapping version. When the camera mapping version changes, a new candidate visual baseline is generated only under the new mapping relationship. The old coordinate image is not directly accumulated into the new background.

[0057] Preferably, when there are fallen leaves, puddles reflecting light, or temporary small objects during the common idle period, although the artificial intelligence vision model does not generate a vehicle instance mask, the processing module still compares the spatial variation range of the candidate background area with the current empty parking space visual baseline. When the variation area remains fixed in multiple images but has no corresponding geomagnetic change, it is written as a background candidate difference into the auxiliary mask of the candidate visual baseline without being immediately merged into the stable background. When the difference area disappears in subsequent common idle periods, the auxiliary mask is deleted. When the difference area continues to exist and does not change the vehicle trajectory and boundary crossing judgment after the candidate baseline event playback, it is then merged into the stable background, so that non-vehicle objects will not directly change the long-term visual reference of the parking space area due to a single idle confirmation.

[0058] Furthermore, the geomagnetic environment baseline is generated from the recovery segment corresponding to the common idle period and the stable sampling after recovery. The processing module checks whether the 3-axis geomagnetic components continuously revert to the same environmental trajectory and writes the slowly varying components that do not belong to the vehicle disturbance sequence into the candidate geomagnetic baseline. The candidate geomagnetic baseline and the candidate visual baseline use the same baseline version identifier. At the same time, the common idle period, parking space identifier and the state transition record used during generation are saved. When reading the baseline for subsequent occupancy recognition, the same version of the visual baseline and geomagnetic baseline must be read in pairs. It is forbidden to temporarily combine the two types of baselines generated by different idle events to avoid misalignment between the on-site state corresponding to the image background and the on-site state corresponding to the geomagnetic background.

[0059] In a preferred embodiment, the processing module configures trajectory identifiers for target trajectories in the extended visual sequence and anchor point identifiers for geomagnetic state transition events in the geomagnetic sequence. When multiple target trajectories or multiple geomagnetic state transition events exist within the same extended time period, the processing module establishes corresponding records for trajectory identifiers and anchor point identifiers according to the order in which the target trajectories cross the boundary of the parking space image area and the order in which the geomagnetic state transition events occur. The target trajectories and geomagnetic state transition events that form corresponding records are sent to the state transition table, and the data that does not form corresponding records are retained in the abnormal event record. The corresponding record also saves the overlapping interval of the target trajectory and the anchor point time period, the movement direction of the target trajectory relative to the entrance boundary, and the current confirmation status before association.

[0060] Specifically, when multiple vehicles pass through the same charging station entrance consecutively, each target trajectory maintains an independent trajectory identifier. The processing module arranges visual events in the order in which each trajectory begins to approach the entrance boundary, completes boundary crossing, and forms an end-stop relationship. The geomagnetic side arranges anchor point events in the order in which the disturbance start point, the continuous segment boundary, and the recovery segment boundary occur. Only when the time intervals overlap and the event direction can form a trajectory identifier and anchor point identifier that allow migration with the current confirmed state are corresponding records established. When the preceding vehicle passes without stopping and the following vehicle enters and parks, the preceding vehicle's trajectory is retained as a non-stop visual event, and the following vehicle's trajectory is established in correspondence with the subsequent continuous geomagnetic disturbance, avoiding the incorrect binding of the preceding vehicle's trajectory to the following vehicle's geomagnetic event based solely on the closest time.

[0061] In this embodiment, when multiple geomagnetic state transition events occur consecutively within a short period of time, the processing module first determines whether they belong to the internal fluctuations of the same vehicle event or to separate events based on the vector continuity of the vehicle disturbance sequence. Local fluctuations belonging to the same continuous vector trajectory are retained under the same anchor point identifier. Changes where the directional relationship is broken and the intermediate part recovers to the environmental trajectory are established with new anchor point identifiers. When there is a complete trajectory on the visual side but no corresponding anchor point, the trajectory data is written into the visual anomaly type. When there is a complete anchor point on the geomagnetic side but no vehicle trajectory, the geomagnetic data is written into the magnetic field anomaly type. The anomaly data retains the original sampling and baseline version for subsequent confirmation record expansion and baseline contamination inspection.

[0062] Furthermore, the abnormal event records maintain a read-only association with the original data according to the parking space identifier and event time. When a new common idle period is formed or the environmental trajectory is corrected, the processing module rereads the magnetic field abnormality records that are still within the verifiable range, uses the corrected geomagnetic environment baseline to re-separate the vehicle disturbance sequence, and checks whether the visual trajectory in the same period has obtained a complete boundary crossing sequence due to occlusion continuation. Data that can be entered into the state transition table after verification is returned to the corresponding record to be confirmed. Data that still cannot form a correspondence between trajectory identifier and anchor point identifier continues to retain the abnormal type. Data that has participated in the baseline pollution determination is given a pollution source identifier so that the same abnormality will not be repeatedly written into different data sets as normal idle samples or normal vehicle events.

[0063] In a preferred embodiment, the processing module combines the updated visual baseline of the empty parking space and the geomagnetic environment baseline to form a candidate baseline version. Before replacing the current baseline version, the module calls the candidate baseline version to reprocess the associated visual sequence and geomagnetic sequence after the common idle period. If the boundary crossing sequence and geomagnetic state transition event obtained by reprocessing still satisfy the state transition table, the candidate baseline version is enabled. Otherwise, the current baseline version is retained, and the candidate baseline version and the corresponding common idle period are written into the baseline contamination record. During reprocessing, the AI ​​visual model version, camera mapping version, and state transition table version remain unchanged, so that the changes in the processing results only come from the candidate baseline data.

[0064] Furthermore, the reprocessing of candidate baseline versions adopts an event playback method. The processing module reads several related events that have been confirmed after the candidate baseline is formed, re-executes parking space background alignment and vehicle candidate area verification using the candidate visual baseline, and re-executes geomagnetic state transition event extraction using the candidate geomagnetic baseline. Then, the visual boundary crossing order, dwelling relationship and geomagnetic event order formed by the playback are sent to the same state transition table. When the playback result is consistent with the original confirmed record in terms of event direction and state before and after, the candidate version is set as the current version and the original version is retained as a reversible version. When there are missing events, changes in event direction or state chain breaks, the current version is not replaced and the common idle time period, image sample and geomagnetic fragment that caused the difference are written into the baseline contamination record.

[0065] Preferably, the baseline version storage adopts a front-to-back version association structure. Each candidate version saves the source version identifier, common idle time period, and replay event set. After a candidate version is activated, the source version is not deleted immediately. When the processing module finds that the visual boundary crossing sequence and the geomagnetic state transition event continue to occur in subsequent closing events, it reads the event record after the current version is activated and compares it with the source version for replay. If the source version can restore the original state chain but the current version cannot, the current version is stopped from being used for new events and reverts to the source version. At the same time, the current version is added to the baseline contamination record. After regaining the common idle time period, a new candidate version is generated. The rollback process only replaces the empty parking space visual baseline and the geomagnetic environment baseline, without changing the already saved original image, original geomagnetic sequence, and artificial intelligence visual model version.

[0066] In one implementation, when a person passes by, a charging cable swings, or a vehicle's headlights sweep across, causing a change in the foreground of the parking space image area, the AI ​​visual model may generate a short-term non-vehicle area or an incomplete vehicle candidate. However, the geomagnetic sequence does not show a continuous disturbance starting point corresponding to the entry of the vehicle's metal body, and the state transition table cannot form a complete temporal correspondence from entry to occupancy. The processing module retains the relevant visual data as an abnormal record without a corresponding geomagnetic event and maintains an idle confirmation state. If a real vehicle subsequently enters, a new geomagnetic anchor point will re-capture the associated visual sequence and establish an independent target trajectory. The aforementioned visual anomalies will not occupy new trajectory markers or change the visual baseline of the empty parking space.

[0067] In another implementation, when the output of the geomagnetic detection module shifts due to vehicles passing by adjacent parking spaces, changes in the working status of underground metal facilities, or slow drift of the environmental magnetic field, the processing module first separates the slowly varying components through the environmental trajectory, then checks whether the remaining vehicle disturbance sequence has a continuous starting point and a continuous segment. At the same time, it calls the visual sequence within the anchor point time period to check whether the vehicle has completed crossing the entrance boundary. When the geomagnetic event is complete but there is no corresponding vehicle trajectory, a magnetic field anomaly record is established without changing the current confirmation status. After the subsequent recovery segment arrives, the recovery segment is used to correct the environmental trajectory. Only the common idle time period formed by the visual sequence and the geomagnetic sequence jointly confirming the idle time period is eligible for baseline update. Abnormal offset time periods are not written into the paired baseline version.

[0068] In one occlusion implementation, the vehicle is blocked by the charging pile pillar during its entry, and the target trajectory is interrupted near the entrance boundary. The geomagnetic detection module forms a disturbance start point and enters a continuous segment at the same time. The processing module locks the visual frames before and after the disturbance start point and extends the associated visual sequence. When the vehicle reappears after the occlusion area, the reappearance position is mapped to the parking plane and a reachable zone is formed with the movement direction before the interruption. The duration of invisibility is then limited by the start point of the geomagnetic event to the boundary of the continuous segment. The preceding and following trajectory segments that meet the conditions are merged and the entry and crossing sequence is restored. When the end of the target trajectory is located inside the parking space image area and a dwell relationship is formed, the state transition table writes the geomagnetic continuous disturbance and the visual entry trajectory into the same occupancy event.

[0069] In one implementation of the vehicle departure process, after the vehicle starts, only part of the vehicle body can be observed due to obstruction by neighboring vehicles. The geomagnetic sequence has entered the recovery phase from the continuous deviation state. The processing module interprets the current geomagnetic event as a departure candidate based on the occupancy confirmation state, and reads the occupancy image forward and the vacancy image backward from the corresponding time of the recovery phase. The target trajectory has a movement direction towards the entrance boundary before obstruction. When it reappears on the entrance side after obstruction and leaves the parking space image area, the departure crossing record is completed. The state transition table updates the confirmation state to vacancy. Subsequently, a common vacancy period is generated by the time period when the visual vehicle instance disappears and the geomagnetic recovery stabilizes. The visual baseline of the vacant parking space and the geomagnetic environment baseline are generated as candidate versions from the same time period.

[0070] In one system startup implementation, the processing module loads the most recently verified paired baseline version, camera mapping version, artificial intelligence visual model version, and unclosed pending confirmation records. After the image acquisition module and geomagnetic detection module recover the data, they first perform a unified time axis conversion. The processing module restores the parking space confirmation status based on the vehicle instance in the current image and the state of the current geomagnetic sequence relative to the geomagnetic environment baseline. If both observations indicate that the parking space is idle at startup, the system enters idle observation. If both observations indicate that the parking space is occupied and the historical record shows that the parking space is occupied, the system continues to observe the parking space. If the two observations conflict or the historical pending confirmation record has not been closed, the system maintains the historical confirmation status and expands the original event container. The system does not create a new state chain lacking the cause of entry or exit due to device restart.

[0071] In one communication anomaly implementation, when the image acquisition channel is temporarily interrupted while the geomagnetic acquisition channel is continuous, the processing module saves the geomagnetic state transition event and anchor point identifier and establishes a pending confirmation record for missing visual conditions. After the image channel is restored, it is determined whether the preceding and following related visual sequences can be completed based on the acquisition time of the restored frame and the buffer interruption boundary. If they can be completed, vehicle instance segmentation and trajectory association continue to be performed. If they cannot be completed, the most recently confirmed state is retained and the event is transferred to the anomaly record. When the geomagnetic acquisition channel is interrupted while the image channel is continuous, the corresponding processing method is adopted. The missing time period does not participate in the update of the geomagnetic environment baseline or the visual baseline of the empty parking space. The pairing of baseline candidates is only reopened after both channels are restored and a common idle time period is re-established.

[0072] In one storage recovery implementation, the processing module saves cached index checkpoints according to the event container boundaries. The checkpoints record the most recently confirmed state, the pending confirmation records that are being expanded, the current baseline version, and the end times of each acquisition queue that have been persisted. When recovering from the checkpoints after the processing process ends abnormally, the candidate states that have not yet completed the closure determination are not written into the confirmed state. Instead, the original image index and geomagnetic sampling index after the checkpoint are reread. Data that may be repeatedly arrived is deduplicated according to the frame identifier or sampling identifier, and then the anchor point extraction, target trajectory association, and state transition determination are continued. When the data file corresponding to the checkpoint is damaged, the most recently confirmed state is retained and a data gap anomaly record is generated. The gap coverage period does not participate in the generation of any baseline candidate version.

[0073] Furthermore, the parking space status record output by the processing module includes at least the parking space identifier, current confirmed status, status formation time, associated trajectory identifier, associated anchor point identifier, baseline version used, and pending confirmation flag. When the current confirmed status changes, the state migration table record that triggers the migration and the corresponding visual sequence and geomagnetic sequence index are saved. When the current confirmed status does not change but there is an abnormal event, only the abnormal record and pending confirmation flag are updated. The output interface sends a distinction marker between the confirmed status and the pending confirmation status to the charging parking space management terminal, so that the upper-level device can distinguish between the occupancy result that has completed the cross-modal event closure and the observation result that is still waiting for supplementary data, without directly interpreting the transient conflict as a new idle or occupancy status.

[0074] In this embodiment, starting from the moment the vehicle approaches the charging parking space, the geomagnetic detection module forms a continuous geomagnetic sequence, the image acquisition module forms a continuous image with a unified timestamp, the processing module establishes a state transition event and locks the corresponding visual frame when the geomagnetic difference vector forms a continuous directional correlation, the artificial intelligence visual model generates a vehicle instance mask and target trajectory from the locked sequence, after the target trajectory completes the crossing from the entrance side to the interior side and forms a dwell relationship, it is sent to the state transition table together with the geomagnetic continuous segment, the state transition table is written with the occupancy confirmation state, when the vehicle leaves, it is written with the idle confirmation state according to the reverse event chain, the common observation period after the idle period generates paired candidate baselines, the candidate baselines become the current version after the state chain is kept consistent by the historical event playback, thus constituting a closed-loop processing process of geomagnetic event limiting visual analysis time, visual trajectory interpreting geomagnetic event source, state transition constraint result writing, and common idle data calibration of subsequent identification baselines.

Claims

1. A charging parking space occupancy recognition system based on the fusion of AI vision and geomagnetism, characterized in that, It includes an image acquisition module and a geomagnetic detection module for the corresponding charging parking space, as well as a processing module that communicates with both. The processing module establishes a parking space image area and caches continuous images, extracts geomagnetic state transition events and corresponding times from the geomagnetic sequence of the geomagnetic detection module, and extracts the related visual sequence before and after the corresponding times. The visual sequence is used to identify vehicles and correlate cross-frame trajectories using an AI visual model to obtain the vehicle's crossing direction and stopping relationship relative to the boundary of the parking space image area. The geomagnetic state transition event is used to generate a parking space occupancy status by corresponding the crossing direction and the dwelling relationship with the time sequence of the vehicle entering the space or occupying the space and leaving the space. The visual baseline of the empty parking space and the geomagnetic environment baseline are updated when the geomagnetic sequence and the visual sequence jointly represent that the parking space is vacant.

2. The charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 1, characterized in that, The processing module performs baseline separation on the geomagnetic sequence according to the geomagnetic environment baseline, determines the disturbance starting point based on the consistency of the vector change direction of continuous sampling relative to the geomagnetic environment baseline, and identifies the continuous segment that deviates from the geomagnetic environment baseline or the recovery segment that returns to the geomagnetic environment baseline from the disturbance starting point. The boundary between the disturbance initiation point and the sustained or recovered segment is taken as the corresponding moment of the geomagnetic state transition event, and the start and end anchor points of the visual sequence are determined respectively. The preceding and following related images covering the start and end anchor points are extracted from the cached continuous images.

3. The charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 2, characterized in that, The processing module performs vehicle instance segmentation on the preceding and following images, and associates the instances of the same vehicle with the target trajectory based on the regional overlap continuity and movement direction of the vehicle instances in adjacent images. The boundary of the parking space image area is divided into an entrance side and an inner side. The order in which the target trajectory crosses from the entrance side to the inner side or in the opposite direction is recorded. The dwelling relationship is generated based on the spatial inclusion relationship between the instance mask at the end of the target trajectory and the parking space image area.

4. The charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 1, characterized in that, The processing module stores a state transition table, which is associated with the current confirmed state, the geomagnetic state transition event, the boundary crossing order of the target trajectory, and the candidate state. When the geomagnetic state transition event and the boundary crossing sequence conform to the temporal correspondence recorded in the state transition table, the candidate state is written into the parking space occupancy state. When the two do not conform to the temporal correspondence, a pending confirmation record is established. The pending confirmation record is associated with geomagnetic sequence segments and target trajectories within the same anchor point range and retains the current confirmation state.

5. A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 2, characterized in that, The baseline separation uses the most recent geomagnetic segment that is jointly characterized by the geomagnetic sequence and the visual sequence as the reference segment, and generates an environmental trajectory covering the geomagnetic state transition event along the geomagnetic change trend before the start of the disturbance. The vector difference between the current geomagnetic sequence and the environmental trajectory is used to form a vehicle disturbance sequence. The start and end boundaries of the vehicle disturbance sequence are used to correct the corresponding time. After the recovery segment is formed, the continuous change between the recovery segment and the reference segment is used to correct the environmental trajectory.

6. A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 3, characterized in that, When the target trajectory is interrupted between adjacent images, the processing module maps the motion direction of the instance mask before the interruption and the appearance position of the instance mask after the interruption to the parking space image area, and constrains the candidate association with the time range from the start of the disturbance to the continuous segment. Merge the preceding and following trajectory segments that conform to the boundary crossing sequence and correspond to the same geomagnetic state transition event to form a target trajectory that includes markers for the occluded sections.

7. A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 4, characterized in that, The processing module identifies modal conditions in the record to be confirmed that do not meet the state transition table, expands the visual sequence with the corresponding time of the geomagnetic state transition event adjacent to the modal condition, and reads the geomagnetic sequence after the corresponding time period. The expanded visual sequence and the geomagnetic sequence are sequentially spliced ​​with the record to be confirmed. When the splicing result satisfies the state transition table, the corresponding candidate state is written. When the subsequent geomagnetic state transition event is completed before the modal condition is met, the record to be confirmed is transferred to an abnormal event record.

8. A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 7, characterized in that, After the state transition table completes the boundary crossing from the inside side of the parking space image area to the entrance side based on the target trajectory and the geomagnetic sequence enters the recovery segment, updating the current confirmed state from occupied to idle, the processing module determines the common idle period that simultaneously covers the visual sequence and the geomagnetic sequence. The image of the empty parking space visual baseline is updated by extracting the image of the mask of the non-vehicle instance from the common idle period, and the geomagnetic environment baseline is updated with the recovery segment corresponding to the time. The two updated data are bound to the same baseline version identifier.

9. A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 7, characterized in that, The processing module configures trajectory identifiers for target trajectories in the expanded visual sequence and configures anchor point identifiers for geomagnetic state transition events in the geomagnetic sequence. When multiple target trajectories or multiple geomagnetic state transition events exist within the same extended time period, corresponding records of the trajectory identifier and the anchor point identifier are established according to the order in which the target trajectory crosses the boundary of the parking space image area and the order in which the geomagnetic state transition events occur. The target trajectories and geomagnetic state transition events that form corresponding records are sent to the state transition table, and the data that does not form corresponding records are retained in the abnormal event record.

10. A charging parking space occupancy recognition system based on AI vision and geomagnetic fusion as described in claim 9, characterized in that, The processing module combines the updated visual baseline of the empty parking space and the geomagnetic environment baseline into a candidate baseline version. Before replacing the current baseline version, it calls the candidate baseline version to reprocess the associated visual sequence and geomagnetic sequence after the common idle period. The candidate baseline version is enabled when the reprocessed boundary crossing sequence and the geomagnetic state transition event still satisfy the state transition table; otherwise, the current baseline version is retained, and the candidate baseline version and the corresponding common idle time period are written into the baseline contamination record.

Images