Intelligent perception big data collection system of first-view multi-camera

Abstract

The application relates to the technical field of big data collection, data organization and data processing, in particular to an intelligent perception big data collection system of a first-view multi-view camera. The system receives multiple image frames or video clips output by the multi-view camera, constructs a candidate collection unit based on a unified collection time axis, extracts time offset, picture quality, motion continuity, view overlap and view complementary features, generates a multi-view collection state vector, and discriminates collection states such as consistent synchronization, local anomaly, cross-view compensation, continuous repetition and unavailability; then, according to the discrimination result, the system performs complete writing, down-sampling writing, buffer collection or discards record collection control, and writes the collection data, quality labels, view association labels, time window version identifiers and original position identifiers into a data management layer. The application can reduce low-quality, repeated and asynchronous data before big data collection and storage, retain data clips with cross-view compensation value, and keep the collection link traceable.

Description

Technical Field

[0001] This invention relates to the fields of data acquisition, data organization and data processing technology, specifically to an intelligent sensing big data acquisition system using a first-view multi-camera. Background Technology

[0002] First-person view multi-camera intelligent sensing data acquisition systems are typically used to receive multiple image frames or video clips, and use the continuously acquired data as the basic data source for subsequent sensing analysis, sample retrieval, model training, or data playback. Existing conventional solutions primarily organize data based on camera number, acquisition time, and file sequence number. Each camera continuously outputs data at a preset frame rate, and the acquisition program writes the multiple data streams to a buffer according to local timestamps or the order of reception. A background task then generates file packages or data fragments by time period. For big data acquisition scenarios, conventional systems focus on data access bandwidth, write throughput, and file integrity, typically using buffer queues, batch writing, breakpoint resumption, and basic verification to ensure continuous data storage. While these solutions can achieve continuous collection of multi-camera data, the acquisition control logic mainly relies on a fixed frequency and fixed window. Whether the data has cross-camera synchronization relationships, whether there are duplicate views, or whether there is compensable viewpoint information are usually not considered in the writing decision during the acquisition phase.

[0003] In similar existing solutions, multi-view data synchronization is generally achieved through a unified clock, frame number alignment, or receiver timestamp normalization. The acquisition end collects image frames from different cameras within a certain time range, grouping data with time differences less than a preset range into the same time slice and storing them in units of time slices. Some solutions perform brightness, blur, occlusion, or frame loss detection on individual images. When the quality of a certain image is lower than a preset condition, the data of that image is marked or removed. Some solutions compare the similarity of consecutive adjacent frames during the post-acquisition processing stage to reduce completely duplicated image frames. The above processing is mostly monocular quality judgment or timestamp-based synchronization merging, focusing on the sharpness of a single frame, the continuity of the single-channel frame sequence, and the static time difference. It rarely considers the viewpoint overlap relationship between multi-view images, the information supplementation relationship of non-overlapping areas, and the consistent change relationship of motion state within adjacent time windows. Therefore, it is difficult to form a basis for writing to the overall state of multi-view images during acquisition.

[0004] In the big data acquisition and storage stage, conventional techniques typically write raw images, video clips, camera numbers, acquisition times, file paths, and a few quality labels into a database or file index table. When sample screening is needed later, the raw data is read offline to calculate image similarity, sharpness, temporal continuity, or manual annotation results, and cleaning, sampling, or training set construction are performed accordingly. Because the acquisition end does not establish a cross-view spatiotemporal consistency description before data is written, offline processing can only perform remedial screening after a large amount of data has accumulated. Duplicate data, misaligned data, and locally anomalous data have already occupied storage space and entered the subsequent retrieval chain. If a camera in a particular view experiences short-term occlusion, motion blur, or time drift, conventional solutions tend to mark the entire time slice as anomalous or only remove anomalous view data, without considering the perspective compensation capabilities of other view cameras to determine whether the time slice is still necessary to retain. This leads to the mixing of usable complementary information with invalid information.

[0005] Therefore, the main technical problem with existing first-view multi-camera big data acquisition systems lies in the lack of a cross-view spatiotemporal consistency discrimination mechanism at the acquisition end, which focuses on candidate acquisition units. This prevents the simultaneous identification of the correlation between temporal offset, image quality, motion continuity, viewpoint overlap, and viewpoint complementarity before writing to the data management layer. The root cause of this problem is not simply insufficient acquisition frequency or storage capacity, but rather the failure to uniformly express the relationships between multiple image frames in both the temporal and viewpoint dimensions as state data usable for acquisition control. This forces the system to write, downsample, or discard data according to fixed rules. When multi-view data exhibits local anomalies, continuous repetition, or slight temporal drift, the acquisition system cannot distinguish between situations requiring complete retention, downsampling, buffering and reconstruction, and discarding. Consequently, low-quality, repetitive, or asynchronous data enters the intelligent sensing dataset, affecting the stability of subsequent data retrieval, sample extraction, and model training data construction. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent sensing big data acquisition system using a first-view multi-camera, which can solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] The intelligent sensing big data acquisition system for first-person view multi-view cameras includes a multi-view data receiving module, a unified time window construction module, a cross-view feature extraction module, a spatiotemporal consistency discrimination module, an acquisition strategy control module, and an acquisition index writing module.

[0009] The multi-view data receiving module is used to receive multiple image frames or video clips output by a first-view multi-view camera.

[0010] The unified time window construction module is used to organize multiple image frames or video segments into candidate acquisition units according to the acquisition time.

[0011] The cross-view feature extraction module is used to extract time offset features, image quality features, motion continuity features, and cross-view correlation features from candidate acquisition units;

[0012] The spatiotemporal consistency discrimination module is used to generate consistency discrimination results for candidate acquisition units based on the features;

[0013] The acquisition strategy control module is used to determine the control strategy of valid acquisition, downsampling acquisition, buffered acquisition, or discarded acquisition based on the consistency discrimination result.

[0014] The acquisition index writing module is used to associate and write the acquired data, control strategies, and quality labels into the intelligent sensing data acquisition system of the data management layer.

[0015] Preferably, the unified time window construction module includes a timestamp normalization processing submodule and a candidate acquisition unit generation submodule;

[0016] The timestamp normalization processing submodule is used to map the local timestamps carried by the output data of different cameras to a unified acquisition time axis, and generate time-aligned records according to the adjacent frame interval, frame sequence number continuity and cross-camera arrival order.

[0017] The candidate acquisition unit generation submodule is used to aggregate multiple image frames or video segments within the same time boundary into candidate acquisition units based on the unified acquisition time axis, and to add status identifiers to missing frames, duplicate frames and out-of-order frames respectively, so that subsequent consistency judgment takes multi-view data within the same time window as the processing object.

[0018] Preferably, the cross-view feature extraction module includes a view overlap analysis submodule and a view complement analysis submodule;

[0019] The viewpoint overlap analysis submodule is used to extract local visual description features from different target image frames within the same candidate acquisition unit, and generate overlap representations based on the distribution density, matching stability and screen proportion of the cross-target matching region.

[0020] The view complementarity analysis submodule is used to extract edge integrity, texture difference and motion direction difference features from image regions that have not formed a stable match, and combine them with the overlapping representation to generate view complementarity representation, so that the cross-view correlation features simultaneously include repeated field of view information and non-repeated field of view information.

[0021] Preferably, the spatiotemporal consistency discrimination module includes a multidimensional state vector generation submodule and a consistency classification submodule;

[0022] The multidimensional state vector generation submodule is used to encode the time offset feature, image quality feature, motion continuity feature and cross-viewpoint correlation feature into a multi-view acquisition state vector of the candidate acquisition unit;

[0023] The consistency classification submodule is used to identify the candidate acquisition unit's category (consistent synchronization, local anomaly, cross-view compensation, continuous repetition, and unavailable) based on the multi-view acquisition state vector, and outputs the category as the consistency discrimination result to the acquisition strategy control module.

[0024] Preferably, the acquisition strategy control module includes a hierarchical acquisition decision submodule and a multi-view segment association submodule;

[0025] The hierarchical acquisition decision submodule is used to configure different data writing levels for candidate acquisition units according to the consistency discrimination results. The consistent synchronization class corresponds to complete writing, the cross-destination compensation class corresponds to retaining compensable destination data and abnormal destination data index, the continuous repetition class corresponds to key segment writing, and the unusable class corresponds to discarding record writing.

[0026] The multi-view segment association submodule is used to establish viewpoint association identifiers for different viewpoint data retained in the same candidate acquisition unit, and to associate and record redundant data that has not been written with summary identifiers and original location identifiers.

[0027] Preferably, the timestamp normalization processing submodule includes a drift fragment identification submodule and a time window reconstruction submodule;

[0028] The drift segment identification submodule is used to identify target data segments with time drift based on changes in cross-target arrival order, abrupt changes in frame number intervals, and deviations in visual motion trends within adjacent time windows in continuous candidate acquisition units.

[0029] The time window reconstruction submodule is used to reconfigure the time boundaries of the affected candidate acquisition units after identifying the time-drifted target data segment, and to re-aggregate the multiple image frames or video segments before and after the drift into reconstructed candidate acquisition units, so that the candidate acquisition units do not use a single path timestamp as the sole alignment basis.

[0030] Preferably, the viewpoint complementarity analysis submodule includes an occlusion anomaly localization submodule and a compensation value assessment submodule;

[0031] The occlusion anomaly localization submodule is used to compare the sharpness distribution, edge breakage distribution, and local texture loss distribution of different target image frames within the same candidate acquisition unit, and generate anomaly region location records.

[0032] The compensation value assessment submodule is used to perform cross-view correspondence between the abnormal region location record and the non-abnormal region in other target image frames, generate a compensable region representation for abnormal target data, and incorporate the compensable region representation into the viewpoint complementary representation to distinguish between data segments that are locally abnormal but still have retention value and data segments that are unusable overall.

[0033] Preferably, the consistency classification submodule includes a temporal continuity constraint submodule and a category stability correction submodule;

[0034] The temporal continuity constraint submodule is used to sequentially associate the multi-view acquisition state vector of the current candidate acquisition unit with the multi-view acquisition state vector of the adjacent candidate acquisition units to form a temporal discrimination sequence including the current state, the previous state, and the subsequent state.

[0035] The category stability correction submodule is used to perform secondary classification on isolated local anomaly classes, continuous repetition classes and unusable classes according to the time-series discrimination sequence, and output the secondary classification result as the updated consistency discrimination result to the acquisition strategy control module.

[0036] Preferably, the time window reconstruction submodule includes a reconstruction version management submodule and a data acquisition link marking submodule;

[0037] The reconstruction version management submodule is used to configure version identifiers for the pre-reconstruction candidate acquisition units and the reconstruction candidate acquisition units respectively, and to record the source of target data, time boundary change information and frame sequence number mapping information for participating in the reconstruction.

[0038] The acquisition link marking submodule is used to write the version identifier, the time boundary change information and the frame sequence number mapping information into the index record formed by the acquisition index writing module, so that the data fragments reconstructed by the time window retain the correspondence before and after reconstruction in the data management layer.

[0039] Preferably, the category stability correction submodule includes a policy backtracking update submodule and a label consistency maintenance submodule;

[0040] The strategy backtracking update submodule is used to backtrack and update the acquisition control strategy that has been generated but not yet solidified when the secondary classification results of adjacent candidate acquisition units change continuously, and send the updated effective acquisition, downsampling acquisition, cached acquisition or discard acquisition control strategy to the acquisition index writing module.

[0041] The label consistency maintenance submodule is used to synchronously update quality labels, viewpoint association labels, time window version identifiers, and control strategy identifiers, so that multiple types of index labels for the same acquisition segment maintain consistent reference relationships.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. After receiving multi-view data, candidate acquisition units are constructed using a unified time window. Time offset features, image quality features, motion continuity features, and cross-viewpoint correlation features are extracted from these candidate acquisition units. This ensures that multiple image frames or video clips no longer enter the big data acquisition chain solely based on the receiving order or a single path timestamp. The spatiotemporal consistency discrimination process encodes these features into a multi-view acquisition state vector, distinguishing between acquisition states such as consistent synchronization, local anomalies, cross-view compensation, continuous repetition, and unavailability. The acquisition strategy control process generates different processing methods based on this, such as complete writing, downsampling writing, cached acquisition, or discarding records. Through this processing, the acquisition end can complete acquisition control based on multi-view relationships before the data falls into the data management layer, reducing the direct entry of time-displaced, continuous, repetitive, locally blurred, and monocular occlusion data into the intelligent sensing dataset. For candidate acquisition units where one view's data is abnormal but other view data still have compensation areas, the compensable view data and the abnormal view data index can be retained, avoiding the same processing of the entire time slice based on monocular quality results. As a result, the objects into which the collected data is written correspond to the spatiotemporal relationships of the multi-view system, reducing the sources of data noise encountered in subsequent retrieval, sampling, and training data construction, and making the synchronization and perspective relationships between data fragments clearer.

[0044] 2. A complete data processing chain is formed, encompassing timestamp normalization, viewpoint complementarity analysis, temporal continuity correction, and acquisition index writing. Timestamp normalization identifies time drift through frame sequence number continuity, cross-view arrival order, and visual motion trends, and reconstructs time windows when necessary, ensuring that candidate acquisition units do not completely depend on a single timestamp boundary. Viewpoint complementarity analysis distinguishes between locally anomalous segments and overall unusable segments through anomaly region location records and compensable region representations, providing clear data basis for retention, downsampling, and discard strategies. Temporal continuity constraints establish sequence associations between the current candidate acquisition unit and adjacent candidate acquisition units, correcting isolated anomaly categories and reducing acquisition strategy jumps caused by occasional fluctuations in a single frame. During acquisition index writing, control strategies, quality labels, viewpoint association labels, time window version identifiers, frame sequence number mapping information, and original location identifiers are stored together, maintaining a traceable relationship between acquisition segments before and after reconstruction, retained multi-view data, and unwritten redundant data. When retrieving data later, representative fragments, compensation fragments, abnormal records, and summary identifiers can be directly located based on the index, reducing the amount of processing required to repeatedly calculate cross-category relationships during the offline stage, and ensuring a consistent data link between cache collection, downsampling writing, and discarding records. Attached Figure Description

[0045] Figure 1 This is the overall data acquisition and control flowchart;

[0046] Figure 2 A flowchart for unified time window construction and time drift reconstruction;

[0047] Figure 3 A flowchart for cross-perspective association and compensation value determination;

[0048] Figure 4 A flowchart for consistency classification, data collection strategy control, and index write-back. Detailed Implementation

[0049] Reference Appendix Figure 1 In one embodiment, the intelligent sensing big data acquisition system for first-view multi-camera systems uses multiple image frames or video clips as acquisition targets. The system includes a multi-view data receiving module, a unified time window construction module, a cross-view feature extraction module, a spatiotemporal consistency discrimination module, an acquisition strategy control module, and an acquisition index writing module. The multi-view data receiving module receives continuous data streams from different cameras and retains the acquisition time, frame number, source camera number, and original data location carried by each data stream. The unified time window construction module does not directly generate files according to the arrival order of a single path, but rather organizes multiple image frames or video clips into candidate acquisition units based on the acquisition time, ensuring that the same candidate acquisition unit contains multi-view data within the same time boundary. The cross-view feature extraction module extracts time offset features, image quality features, motion continuity features, and cross-view correlation features from candidate acquisition units. Time offset features represent the degree of temporal misalignment between different view data; image quality features represent the clarity, edge characteristics, and abnormal area status of each view image; motion continuity features represent the smoothness of image changes within adjacent time windows; and cross-view correlation features represent the overlap and complementary relationships of views between multi-view images. The spatiotemporal consistency discrimination module generates consistency discrimination results based on these features. The acquisition strategy control module determines the control strategy for effective acquisition, downsampling acquisition, buffered acquisition, or discarded acquisition based on the consistency discrimination results. The acquisition index writing module associates the acquired data, control strategy, and quality label into the data management layer. Its operating principle is as follows: the candidate acquisition unit serves as the smallest control object. The system first determines whether multiple data streams are in the same temporal relationship, then determines whether each data stream has usable image quality, and then determines whether there are overlapping or compensable views between different views. The discrimination results are then converted into data and written into control information. Through this organization, the acquisition end forms a control basis including time, quality, motion, and view relationships before the data enters the big data management layer. The advantage of this embodiment is that the acquisition object is transformed from a single frame or a single time file into a multi-view candidate acquisition unit. The acquisition control no longer relies on a fixed frequency and simple file writing rules. It can distinguish different data states at the front end of the acquisition link, such as data that can be fully written, data that needs downsampling, data that needs buffering and correction, and data that needs to be retained and discarded.

[0050] In this embodiment, the multi-view data receiving module establishes a receiving buffer for each input data stream. The receiving buffer identifies image frames or video segments using both frame sequence numbers and acquisition times. For image frames, the receiving buffer records frame-level timestamps; for video segments, the receiving buffer records the segment start time, segment end time, and frame sequence numbers of keyframes within the segment. To avoid the difference in local time bases between different cameras directly affecting the division of candidate acquisition units, the unified time window construction module performs timestamp normalization processing on multi-view data within the same batch. The normalization processing can use the center value of the acquisition time of each camera within the same time window as a reference, and the time offset feature can be calculated according to the following formula:

[0051] ;

[0052] in, Indicates the first The candidate acquisition unit in the first The offset of the target data relative to the normalized reference time. Indicates the first The candidate acquisition unit in the first Data collection time Indicates the first Normalized reference time of each candidate acquisition unit Indicates the first The number of target data points involved in the normalization calculation within each candidate acquisition unit. For example, when the same candidate acquisition unit contains 3 data streams with acquisition times of 1000, 1006, and 1009 respectively, For 1005, the corresponding data for 3 channels The values ​​are 5, 1, and 4, respectively. The system uses this set of offsets as part of the time offset feature and inputs it, along with frame sequence number continuity and cross-destination arrival order, into the subsequent discrimination process.

[0053] In a preferred embodiment, the cross-eye feature extraction module normalizes and encodes image quality features so that data from different eyes can be compared within the same candidate acquisition unit. The image quality features do not use only a single sharpness index, but combine gradient sharpness distribution, edge integrity, and abnormal area coverage into a quality characterization. The quality characterization can be calculated according to the following formula:

[0054] ;

[0055] in, Indicates the first The candidate acquisition unit in the first Image quality characterization of the target data This represents the normalized gradient sharpness. This represents the normalized edge integrity. This indicates the coverage ratio of abnormal areas in the image. , and These represent the weights of the three features mentioned above, and 。 operator This indicates that the less coverage of the anomalous region, the higher the contribution of this factor. For example, when , , , , , hour, This result serves as an image quality feature in the generation of the multi-view acquisition state vector. The values ​​mentioned above are for illustrative purposes only and do not constitute a limitation on the parameters of the acquisition equipment.

[0056] In this embodiment, after the unified time window construction module generates candidate acquisition units, it also adds status identifiers to the candidate acquisition units. These status identifiers are not used as the final acquisition strategy, but rather as the basic input for spatiotemporal consistency judgment. The missing frame identifier indicates that a certain data segment has no corresponding frame or segment within the current time boundary; the duplicate frame identifier indicates that the same data stream has duplicate frame numbers or image content within adjacent time boundaries; the out-of-order frame identifier indicates that the receiving order is inconsistent with the frame number order; and the suspected drift identifier indicates that a certain data segment shows a continuous time offset in consecutive candidate acquisition units. Table 1 shows the specific meanings of the candidate acquisition unit status identifiers. The content in the table is used to illustrate the software processing logic and does not limit the number of cameras, resolution, or acquisition frequency.

[0057] Table 1 Candidate Acquisition Unit Status Identifier Table

[0058] Normal marking Data from all targets within the current time boundary can be categorized into the same candidate acquisition unit. Retain the original acquisition time, frame number, and source destination number. Missing identifier One or more data streams do not have corresponding frames or segments within the current time boundary. Record the missing source item number and retain other item data. Duplicate identifier The current data has the same frame number or similar content as the data from the same channel in adjacent candidate acquisition units. Generate duplicate association records and proceed to downsampling judgment. Out-of-order identifiers The receiving order is inconsistent with the frame sequence number order. Reorder by frame number and record the original arrival order. Drift suspected marker The offset trend of the same target data in continuous time boundary continues to deviate Entering the drift segment identification and time window reconstruction process

[0059] The advantage of this embodiment is that the candidate acquisition unit carries an interpretable status identifier when it is generated, and subsequent discrimination does not require re-traversing the original receiving log to obtain suspected information on missing, duplicate, out-of-order and drift information, making the discrimination link before data writing more stable.

[0060] Reference Appendix Figure 2Preferably, the timestamp normalization processing submodule maps the local timestamps carried by the output data of different target cameras to a unified acquisition timeline, and generates time alignment records according to the adjacent frame interval, frame sequence number continuity, and cross-target arrival order. The candidate acquisition unit generation submodule uses the unified acquisition timeline as a reference to aggregate multiple image frames or video segments within the same time boundary into candidate acquisition units. If there are missing frames within the same time boundary, the candidate acquisition unit is not directly discarded, but the missing status is written into the time alignment record; if there are duplicate frames, the candidate acquisition unit retains the duplicate source and duplicate object; if there are out-of-order frames, the candidate acquisition unit retains the original reception order and the rearranged frame sequence number order. The above processing enables subsequent consistency judgment to be performed directly based on multi-target data within the same time window, rather than performing remedial matching between multiple discrete files. In this embodiment, the time alignment record includes at least the normalized reference time, source target number, original timestamp, normalized time offset, frame sequence number, arrival order, and status identifier. For video segments, the time alignment record may also include the correspondence between key frames within the segment and the time boundaries of the candidate acquisition units. In this way, the system can organize multi-channel collected data into data units with unified temporal semantics without relying on manual verification. The advantage of this embodiment is that the time window construction process preserves both the original time series and the normalized time series, and the original data source can still be traced back when time window reconstruction occurs subsequently.

[0061] Furthermore, the drift segment identification submodule identifies time-drifted target data segments based on cross-target arrival order changes, abrupt changes in frame number intervals, and visual motion trend deviations within adjacent time windows in continuous candidate acquisition units. Cross-target arrival order changes are used to detect data from a particular path arriving prematurely or delayed for extended periods; abrupt changes in frame number intervals are used to detect missing, duplicated, or discontinuous intermediate frames; and visual motion trend deviations are used to detect situations where timestamps appear close but image changes are out of sync with other targets. After identifying time-drifted segments, the time window reconstruction submodule does not forcibly merge them according to the original time boundaries. Instead, it reconfigures the time boundaries within the affected range and re-aggregates the multiple image frames or video segments before and after the drift into reconstructed candidate acquisition units. Specifically, the time window reconstruction submodule can establish a frame number mapping between drifting target data and non-drifting target data within continuous candidate acquisition units. If the drifting target data is closer to the adjacent time window in terms of visual motion trend, it is removed from the original candidate acquisition unit and merged into a more matching time boundary, while reserving empty slots for the original candidate acquisition unit. This processing ensures that candidate acquisition units do not rely solely on a single path timestamp for alignment. The advantage of this embodiment is that, in the event of slight drift or short-term misalignment, the system can reconstruct the time boundary of the candidate acquisition unit, thereby reducing the direct entry of abnormal synchronization data into the data management layer.

[0062] In a preferred embodiment, motion continuity features are used to assist in identifying temporal drift and abrupt changes in captured segments. The cross-view feature extraction module can extract the visual motion intensity between adjacent time windows for each image frame, for example, by generating a motion scalar through feature point displacement, block matching displacement, or contour position changes between adjacent frames. The motion continuity representation can be calculated according to the following formula:

[0063] ;

[0064] in, Indicates the first The candidate acquisition unit in the first The motion continuity representation of target data, This represents the visual motion scalar of the current candidate acquisition unit. This represents the visual motion scalar of the previous candidate acquisition unit. This represents the upper limit of motion differences used for normalization. For example, when , , hour, When the difference in motion scalars exceeds At that time, the molecules follow Included, A value of 0 indicates a weak continuity of motion between the current time window and the previous time window. This characteristic, when used in conjunction with the time offset feature, can distinguish between rapid changes in a real scene and discontinuous changes caused by time misalignment.

[0065] In this embodiment, the reconstruction version management submodule configures version identifiers for both the pre-reconstruction candidate acquisition unit and the reconstruction candidate acquisition unit, and records the source data, time boundary change information, and frame sequence number mapping information for the data involved in the reconstruction. The version identifier can be generated by combining the acquisition batch identifier, the original sequence number of the candidate acquisition unit, the reconstruction sequence number, and the source destination number. The acquisition link marking submodule writes the version identifier, time boundary change information, and frame sequence number mapping information into the index record formed by the acquisition index writing module, ensuring that the data segments reconstructed through the time window retain the correspondence before and after reconstruction in the data management layer. Specifically, if the first... The second data point from the candidate acquisition units was determined to be more suitable for inclusion in the first... The candidate collection unit was restructured, and the version management submodule was redesigned as the original... The candidate acquisition unit retains the original version identifier, which is the first version after reconstruction. The and the first Each candidate acquisition unit is configured with a new version identifier, and the original and reconstructed ownership of the second data item is recorded in the frame sequence number mapping information. During subsequent data retrieval, the data link before and after reconstruction can be obtained through the version identifier. The advantage of this embodiment is that time window reconstruction does not destroy the original acquisition record, and the reconstructed data still has a traceable data link, facilitating subsequent verification and reproduction of the acquisition process.

[0066] Reference Appendix Figure 3 In another embodiment, the view overlap analysis submodule and the view complement analysis submodule jointly form cross-view correlation features. The view overlap analysis submodule extracts local visual description features from different view image frames within the same candidate acquisition unit. These local visual description features can be generated from corner points, edge intersection regions, texture blocks, or stable contour regions. For any two view images, the view overlap analysis submodule matches the local visual description features and generates an overlap representation based on the distribution density, matching stability, and screen proportion of the cross-view matching region. The overlap representation can be calculated using the following formula:

[0067] ;

[0068] in, Indicates the first The candidate acquisition unit in the first Item and No. The overlap of perspectives between target data is characterized. This represents the set of local visual descriptive features that form a stable match between two images. and They represent the first Item and No. The set of effective local visual descriptive features in the target image that can be used for matching. This indicates the number of elements in the set. For example, when... , , hour, The calculated value is approximately 0.478. This result indicates that there is a certain proportion of overlapping visual field information between the two images, and this representation can be used for subsequent identification of consecutive repeating classes.

[0069] Preferably, the viewpoint complementarity analysis submodule extracts edge integrity, texture difference, and motion direction difference features from image regions that have not formed a stable match, and generates viewpoint complementarity representations by combining overlap representations. Regions that have not formed a stable match are not necessarily unusable regions; they may be non-repeating fields of view formed due to differences in the viewpoints of different cameras. The viewpoint complementarity analysis submodule first removes abnormal regions such as blurred, occluded, and overexposed regions from the unmatched regions, and then extracts contour boundaries, texture direction, and motion direction from the remaining regions. If a certain region... Stable matching cannot be formed in the target image, but in the first... If an image contains complete edges and continuous texture, the region is recorded as a supplementary region. Cross-viewpoint correlation features consist of overlapping and complementary representations, simultaneously representing information from repeated and non-repeating views. Table 2 shows the main fields and their sources in the multi-view acquisition state vector. The fields in the table can be expanded according to the actual software implementation, but the meaning of each field remains consistent.

[0070] Table 2 Multi-view Acquisition Status Vector Field Table

[0071] Time offset field Timestamp normalization processing submodule This indicates the degree of deviation of each item's data from the unified acquisition timeline. Image quality field Image quality feature extraction and processing Indicates sharpness, edge integrity, and coverage of anomalous areas. Motion Continuous Field Visual motion trends of adjacent candidate acquisition units This indicates the continuous relationship between the current time window and adjacent time windows. View overlap field Viewpoint overlap analysis submodule Indicates the proportion of overlapping fields of view and matching stability between different target images. complementary perspective fields Complementary Perspective Analysis Submodule Represents non-repeating supplementable regions between different target images Abnormal transfer field Temporal continuity constraint processing This indicates the persistence or regression status of local anomalies in adjacent candidate acquisition units.

[0072] The advantage of this embodiment is that the cross-viewpoint association features not only reflect the degree of repetition of the field of view, but also express whether the non-repetitive field of view is necessary to retain. The subsequent acquisition strategy can distinguish redundant segments, compensable segments and overall unusable segments.

[0073] Furthermore, the occlusion anomaly localization submodule compares the sharpness distribution, edge breakage distribution, and local texture loss distribution of different target image frames within the same candidate acquisition unit, and generates anomaly region location records. These anomaly region location records can be represented using image region block coordinates, region contour numbers, or local descriptive feature numbers. The compensation value evaluation submodule performs cross-target mapping between the anomaly region location records and non-anomaly regions in other target image frames, generating a compensable region representation for the anomaly target data. The compensable region representation can be calculated using the following formula:

[0074] ;

[0075] in, Indicates the first The candidate acquisition unit in the first Normal area for the first The degree of compensability of the abnormal area. Indicates the first The collection of abnormal regions in the target image. Indicates the first The set of non-abnormal regions in the target image. For example, when the first... The target abnormal region set contains 1200 region units, which can be mapped to the first... When the intersection of the non-abnormal regions is 840 regional units, When the same abnormal region can be covered by data from multiple other destinations, the system can retrieve multiple [data sources]. The higher value in the index is used as a compensable region representation of the abnormal region, and multiple correspondences can also be retained for subsequent index writing.

[0076] In this embodiment, the occlusion anomaly localization submodule does not directly classify all unmatched areas as anomalies. Instead, it performs a cross-judgment based on sharpness distribution, edge breakage distribution, and local texture loss distribution. If an area has concentrated edge breaks, continuous texture loss, and changes in that area within adjacent time windows do not conform to motion continuity, it is marked as an anomaly area. If an area lacks cross-view matching only due to viewpoint differences, but has complete edges and stable texture, it is retained as a non-repeating visual field area. The compensation value assessment submodule determines whether other viewpoint images can cover the main visual content of the anomaly area through cross-viewpoint correspondence. For data segments with local anomalies but compensable areas, the system writes the anomaly viewpoint data index and the compensable viewpoint data index together into the viewpoint association label. For data segments whose anomalies cannot be covered by other viewpoint data, the system inputs them into the unusable class for judgment. The advantage of this embodiment is that local anomaly data is not simply discarded. The system can retain data segments with compensation relationships based on the regional coverage of other viewpoint data, while avoiding the mixing of overall unusable data into the usable sample set.

[0077] Reference Appendix Figure 4 In another embodiment, the spatiotemporal consistency discrimination module includes a multi-dimensional state vector generation submodule and a consistency classification submodule. The multi-dimensional state vector generation submodule encodes time offset features, image quality features, motion continuity features, and cross-viewpoint correlation features into multi-view acquisition state vectors for candidate acquisition units. The consistency classification submodule identifies the candidate acquisition units into four classes based on their multi-view acquisition state vectors: consistent synchronization class, local anomaly class, cross-view compensation class, continuous repetition class, and unusable class, and outputs the class as the consistency discrimination result to the acquisition strategy control module. The multi-view acquisition state vector can be represented in the following form:

[0078] ;

[0079] in, Indicates the first Multi-view acquisition state vector of candidate acquisition units Indicates the first The average time offset of each target in each candidate acquisition unit This indicates the degree of dispersion of the time offset for each item. This represents the aggregated value representing the image quality of each item. The aggregate value representing the continuity of motion of each object. The aggregate value represents the overlap characteristic across different viewpoints. This represents the aggregated value representing the cross-category compensable region. This represents the recorded value indicating an abnormal state transition between adjacent candidate acquisition units. For example, a candidate acquisition unit is calculated to obtain... This indicates that the candidate acquisition unit has low temporal dispersion, high image quality, high motion continuity, certain viewpoint overlap, strong compensable area, and low abnormal transfer records.

[0080] In a preferred embodiment, the consistency classification submodule can generate consistency judgment results using a category scoring method. The category score is not limited to a fixed model and can be implemented using a linear classifier, a tree classifier, or a combination of rules and classifiers. In the disclosed implementation, the category score can be calculated according to the following formula:

[0081] ;

[0082] in, Indicates the first Each candidate acquisition unit belongs to category Category rating, Indicate category The corresponding transpose weight vector, Indicate category The corresponding bias term, This represents a set of categories consisting of consistent synchronization, local anomaly, cross-category compensation, continuous repetition, and unavailable. This represents the transpose of a vector. This represents the dot product of the weight vector and the multi-view acquisition state vector. For example, if the weight vector of a certain category is... The bias term is 0.02. After normalization according to the classifier requirements, the data is used in the dot product, resulting in a score of 0.943 for this category. The system compares the scores of each category and outputs the consistency judgment result of the candidate acquisition units based on the category constraints.

[0083] In this embodiment, category constraints are used to prevent a single high-scoring feature from misleading the acquisition strategy. If both the time offset field and the time discrete field are low, the image quality field and the motion continuity field are within the available range, and the abnormal transition field is low, then the candidate acquisition unit is identified as the consistent synchronization class. If there is a monocular or local area anomaly, but the cross-view compensable area representation is low, then the candidate acquisition unit enters the local anomaly class. If the anomaly area can be covered by the non-anomaly area of ​​other eyes, and the viewpoint complementarity field meets the retention condition, then the candidate acquisition unit enters the cross-view compensation class. If the viewpoint overlap field is high, the motion continuity field is stable, and the content changes of adjacent time windows are small, then the candidate acquisition unit enters the continuous repetition class. If the time misalignment cannot be corrected by reconstruction, the image quality is low, and there is no compensable area, then the candidate acquisition unit enters the unusable class. The classification results and the state vector are output to the acquisition strategy control module. The advantage of this embodiment is that the classification criteria for candidate acquisition units include multi-view time relationships, image status, and viewpoint relationships, which can reduce misjudgments caused by single quality detection or single timestamp alignment.

[0084] Furthermore, the temporal continuity constraint submodule sequentially associates the multi-view acquisition state vector of the current candidate acquisition unit with the multi-view acquisition state vectors of adjacent candidate acquisition units, forming a temporal discrimination sequence including the current state, the preceding state, and the following state. The category stability correction submodule performs secondary classification on isolated local anomaly classes, continuously repeating classes, and unusable classes based on this temporal discrimination sequence. This secondary classification does not change the original features but performs neighborhood fusion on the category scores, ensuring that instantaneous noise or occasional image fluctuations do not directly cause changes in the acquisition control strategy. The category stability correction can be calculated using the following formula:

[0085] ;

[0086] in, Represents the first after time series continuity correction Each candidate collection unit targets the category Corrected rating , and These represent the previous, current, and next candidate collection units for the category, respectively. Category rating, , and Let represent the fusion coefficients of the preceding, current, and subsequent scores, respectively. For example, if a category has scores of 0.40, 0.80, and 0.50 in the preceding, current, and subsequent candidate collection units, respectively, , , hour, The system updates the consistency judgment result based on the corrected score and sends the updated result to the data acquisition strategy control module.

[0087] In this embodiment, the temporal continuity constraint can also reference the anomaly transition field (\Gamma_{k}). When a candidate acquisition unit is initially identified as unusable, but the preceding and subsequent candidate acquisition units are both consistently synchronized, and the abnormal region in the current candidate acquisition unit can be covered by data from other targets, the category stabilization correction submodule corrects the candidate acquisition unit from unusable to cross-target compensation or local anomaly. If multiple consecutive candidate acquisition units are identified as continuous repetitions, the category stabilization correction submodule retains the repetition state and hands the continuous segment over to the acquisition strategy control module for key segment writing. If a candidate acquisition unit is discontinuous in both time offset and motion continuity, and the preceding and subsequent sequences also drift in the same direction, then the candidate acquisition unit remains unusable or enters the cache acquisition process. The advantage of this embodiment is that the discrimination result can utilize the state correlation between adjacent time windows, reducing the acquisition strategy fluctuations caused by isolated anomalies, and making the acquisition control temporally continuous.

[0088] In another embodiment, the acquisition strategy control module includes a hierarchical acquisition decision submodule and a multi-view segment association submodule. The hierarchical acquisition decision submodule configures different data writing levels for candidate acquisition units based on consistency discrimination results. Consistent synchronization corresponds to full writing, cross-view compensation corresponds to retaining compensable view data and abnormal view data indexes, continuous repetition corresponds to writing key segments, and unusable corresponds to discarding record writing. Locally abnormal classes can enter cached acquisition or partial writing based on the coverage area of ​​the abnormal region and the characteristics of the compensable region. The multi-view segment association submodule establishes viewpoint association identifiers for different view data retained in the same candidate acquisition unit and associates unwritten redundant data with summary identifiers and original location identifiers. For continuous repetition classes, the system can determine whether to write representative segments based on information increments. Information increments can be calculated using the following formula:

[0089] ;

[0090] in, Indicates the first Incremental representation of information from each candidate acquisition unit The aggregate value represents the overlap characteristic across different viewpoints. The aggregate value representing the image quality. This represents the aggregate value representing the compensable region. This represents the value recorded during an abnormal state transition. For example, when... , , , hour, This result indicates that although the current candidate acquisition units have a high degree of visual overlap, they still contain certain non-repetitive and compensable information, and can be included in key segment writing or summary association records based on hierarchical acquisition decisions.

[0091] Table 3 shows the correspondence between different consistency categories and acquisition control strategies. This table is used to illustrate the software control logic; the specific storage paths and database fields can be determined according to the implementation method of the data management layer.

[0092] Table 3. Correspondence between Consistency Categories and Acquisition Control Strategies

[0093] Consistent synchronization class Full write Write multi-view raw data pointers, unified time window identifiers, quality labels, and view-related labels. Local anomaly class Cache collection or partial writing Write the location record of the abnormal area, the data index of the abnormal area, and the identifier to be reviewed. Cross-destination compensation Retain indexes of compensable target data and abnormal target data. Write compensable region representation, anomaly region correspondence, and multi-view segment association identifier. Continuous repeating class Key segment writing Write the representative fragment pointer, redundant fragment digest identifier, and original location identifier. Unusable Class Discard record write Write the discard reason label, original location identifier, and candidate acquisition unit version identifier.

[0094] In this embodiment, the data written by the acquisition index writing module includes not only the original file path, but also control strategies, quality labels, viewpoint association labels, time window version identifiers, and candidate acquisition unit categories. For fully written data, the index record retains multi-view data pointers and unified time window identifiers. For cross-view compensation data, the index record retains both the abnormal view data index and the compensable view data index, enabling subsequent retrieval to locate the relationship between abnormal and compensated regions. For continuously repeating data, the index record retains representative fragment pointers and redundant fragment summary identifiers; unwritten redundant data remains traceable through the original location identifier. For unusable data, the index record does not write complete data, but retains the discard reason label and the original location identifier to avoid subsequent repeated processing of the same useless fragment. The advantage of this embodiment is that the acquisition control strategy and the index record form a correspondence during the writing stage, and subsequent data retrieval, sample extraction, and reconstruction verification can directly utilize the labels and association identifiers generated during acquisition.

[0095] Furthermore, the strategy backtracking update submodule is used to backtrack and update the acquisition control strategies that have been generated but not yet solidified when the secondary classification results of adjacent candidate acquisition units change continuously. It then sends the updated effective acquisition, downsampling acquisition, cached acquisition, or discard acquisition control strategies to the acquisition index writing module. The backtracking update is limited to data segments that have not yet been solidified; data that has already been solidified is not directly overwritten, but the change relationship is saved through newly added index version records. Specifically, when a candidate acquisition unit is initially identified as a local anomaly and enters cached acquisition, but subsequent adjacent candidate acquisition units prove that the anomaly is a short-term occlusion and has been covered by data from other targets, the strategy backtracking update submodule can update the cached acquisition to the retention strategy corresponding to the cross-target compensation class. When multiple consecutive candidate acquisition units gradually change from a consistent synchronization class to a continuous repetition class, the strategy backtracking update submodule can change the repetitive segments that have not yet been solidified into key segments and generate redundant summary identifiers. The label consistency maintenance submodule synchronously updates quality labels, viewpoint association labels, time window version identifiers, and control strategy identifiers, ensuring that multiple index labels for the same acquisition segment maintain consistent reference relationships. The advantage of this embodiment is that the collection strategy can use the neighborhood discrimination results to perform backtracking correction within a limited range, avoiding index link conflicts caused by inconsistencies between early discrimination results and subsequent stable categories.

[0096] In a preferred embodiment, the label consistency maintenance submodule uses a unified reference key to manage multiple types of labels for the same acquisition segment. The unified reference key can be formed by combining the candidate acquisition unit version identifier, source destination number, frame sequence range, and control strategy identifier. When the quality label is updated, the viewpoint-related label references the same unified reference key; when the time window version identifier changes, the control strategy identifier is synchronously generated with a new version; when the redundant segment summary identifier is generated, the original location identifier still points to the original data location retained by the acquisition and receiving module. For data reconstructed after the time window, the label consistency maintenance submodule retains both the pre-reconstruction and post-reconstruction versions in the index record and connects them using the same acquisition link identifier. For cross-destination compensation data, the abnormal destination data index and the compensable destination data index each have independent source destination numbers, but share the same candidate acquisition unit version identifier and viewpoint-related label. For continuous repetitive data, the representative segment pointer and the redundant segment summary identifier share the same repetition cluster identifier, allowing the return of the summarized redundant segment location when retrieving the representative segment subsequently. The advantage of this embodiment is that the quality label, viewpoint-related label, time window version identifier, and control strategy identifier maintain a common source reference, reducing misconnections and broken links between index records.

[0097] In this embodiment, the intelligent sensing big data acquisition system can also logically connect the acquisition index writing module with the data management layer's sharded storage. After receiving the index records output by the acquisition index writing module, the data management layer can establish a sharded directory according to the candidate acquisition unit version identifier, and establish a secondary index according to the source directory number and control strategy identifier. Completely written multi-view data is aggregated and stored using a unified time window identifier; cross-view compensation data is aggregated and stored using the correspondence between abnormal regions and the representation of compensable regions; continuous repetitive data is aggregated and stored using representative fragment pointers and redundant summary identifiers; and unusable data is stored separately as discarded records. For subsequent data retrieval, the requesting end can retrieve data using time range, source directory number, quality label, viewpoint association label, or time window version identifier. If the requesting end needs the original acquisition link, the data management layer returns the original frame or fragment position according to the original position identifier; if the requesting end needs reconstructed multi-view consistent data, the data management layer returns the reconstructed candidate acquisition unit according to the reconstructed version identifier. The advantage of this embodiment is that the acquisition index maintains a consistent organizational relationship with the data shards of the big data management layer, enabling the discrimination results generated during the acquisition phase to be directly used for subsequent data retrieval and sample construction.

[0098] In another embodiment, the system can perform representative segment selection for candidate acquisition units in a continuous repeating category. The representative segment is not determined by the sharpness of a single image path, but rather by a combination of information increment representation, time offset representation, image quality representation, and viewpoint complementarity representation. If multiple consecutive candidate acquisition units maintain high values ​​in the viewpoint overlap field and the motion continuity field changes smoothly, the system does not write each candidate acquisition unit completely, but instead selects the candidate acquisition unit with the higher information increment within the repeating cluster as the representative segment. The data selected as the representative segment is completely written to the data management layer, while the remaining candidate acquisition units generate redundant segment summary identifiers and original location identifiers. The summary identifier may include the repeating cluster identifier, candidate acquisition unit version identifier, source directory number, quality aggregation value, and time boundary. The original location identifier is retained in the acquisition and reception cache or persistent log, enabling the location of incompletely written data when it is necessary to recover the original continuous data later. This embodiment, in conjunction with the hierarchical acquisition decision submodule, ensures that continuous repeating data no longer occupies the main index of the data management layer in exactly the same way. The advantage of this embodiment is that the system preserves the link relationship of continuous data while reducing the number of repeating segments entering the main data set. During subsequent retrieval, representative segments can be returned first, and redundant segments can be traced back as needed.

[0099] Preferably, cached acquisition is used to process data segments that cannot be immediately determined. Data segments acquired in the cache include data with suspected drift but requiring reconstruction within a time window, data with local anomalies requiring confirmation in subsequent adjacent time windows, and data whose category scores are close to multiple category boundaries. During cached acquisition, the acquisition strategy control module temporarily stores candidate acquisition units in the pending cache and generates temporary control strategy identifiers for them. After the temporal continuity constraint submodule obtains the multi-view acquisition state vectors of subsequent candidate acquisition units, the category stability correction submodule performs secondary classification on the candidate acquisition units in the pending cache. If the secondary classification result is consistent with the original temporary control strategy, the acquisition index writing module solidifies the temporary control strategy into the formal control strategy; if the secondary classification result changes, the strategy backtracking update submodule generates an update record, and the label consistency maintenance submodule synchronously updates the relevant labels. Cached acquisition does not change the original data reception order, and the pending cache only changes the solidification time of the acquisition control strategy. The advantage of this embodiment is that the system can handle short-term uncertain states without prematurely discarding data and generate more stable acquisition control results after obtaining information from adjacent time windows.

[0100] In this embodiment, the overall software processing chain can be implemented in the acquisition terminal program, edge processing program, or data management service. Data is transferred between functional modules via candidate acquisition unit identifiers. The multi-view data receiving module outputs raw frame records, the unified time window construction module outputs candidate acquisition unit records, the cross-view feature extraction module outputs feature records, the spatiotemporal consistency discrimination module outputs category records, the acquisition strategy control module outputs control strategy records, and the acquisition index writing module outputs index records that can be written to the data management layer. All of these records contain candidate acquisition unit version identifiers, enabling data from different processing stages to be correlated. In the processing chain, if a stage discovers a state requiring correction in a preceding stage, such as time drift, category change, or strategy backtracking, the system does not overwrite the original record but generates a new version record while retaining the mapping relationship. Through this record organization method, the system forms a traceable data processing chain while completing large-scale data acquisition. The advantage of this embodiment is that the inputs and outputs of each module have clear data boundaries, and time window construction, feature extraction, discrimination and classification, strategy control, and index writing can be combined to form a complete acquisition control closed loop, ensuring that spatiotemporal consistency processing is completed before multi-view data is written.

Claims

1. A first-person view multi-view camera intelligent sensing big data acquisition system, characterized in that, It includes a multi-view data receiving module, a unified time window construction module, a cross-view feature extraction module, a spatiotemporal consistency discrimination module, a data acquisition strategy control module, and a data acquisition index writing module; The multi-view data receiving module is used to receive multiple image frames or video clips output by a first-view multi-view camera. The unified time window construction module is used to organize multiple image frames or video segments into candidate acquisition units according to the acquisition time. The cross-view feature extraction module is used to extract time offset features, image quality features, motion continuity features, and cross-view correlation features from candidate acquisition units; The spatiotemporal consistency discrimination module is used to generate consistency discrimination results for candidate acquisition units based on the features; The acquisition strategy control module is used to determine the control strategy of valid acquisition, downsampling acquisition, buffered acquisition, or discarded acquisition based on the consistency discrimination result. The acquisition index writing module is used to associate and write the acquired data, control strategies, and quality labels into the intelligent sensing data acquisition system of the data management layer.

2. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 1, characterized in that, The unified time window construction module includes a timestamp normalization processing submodule and a candidate acquisition unit generation submodule; The timestamp normalization processing submodule is used to map the local timestamps carried by the output data of different cameras to a unified acquisition time axis, and generate time-aligned records according to the adjacent frame interval, frame sequence number continuity and cross-camera arrival order. The candidate acquisition unit generation submodule is used to aggregate multiple image frames or video segments within the same time boundary into candidate acquisition units based on the unified acquisition time axis, and to add status identifiers to missing frames, duplicate frames and out-of-order frames respectively, so that subsequent consistency judgment takes multi-view data within the same time window as the processing object.

3. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 1, characterized in that, The cross-view feature extraction module includes a view overlap analysis submodule and a view complement analysis submodule; The viewpoint overlap analysis submodule is used to extract local visual description features from different target image frames within the same candidate acquisition unit, and generate overlap representations based on the distribution density, matching stability and screen proportion of the cross-target matching region. The view complementarity analysis submodule is used to extract edge integrity, texture difference and motion direction difference features from image regions that have not formed a stable match, and combine them with the overlapping representation to generate view complementarity representation, so that the cross-view correlation features simultaneously include repeated field of view information and non-repeated field of view information.

4. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 1, characterized in that, The spatiotemporal consistency discrimination module includes a multi-dimensional state vector generation submodule and a consistency classification submodule; The multidimensional state vector generation submodule is used to encode the time offset feature, image quality feature, motion continuity feature and cross-viewpoint correlation feature into a multi-view acquisition state vector of the candidate acquisition unit; The consistency classification submodule is used to identify the candidate acquisition unit's category (consistent synchronization, local anomaly, cross-view compensation, continuous repetition, and unavailable) based on the multi-view acquisition state vector, and outputs the category as the consistency discrimination result to the acquisition strategy control module.

5. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 1, characterized in that, The acquisition strategy control module includes a hierarchical acquisition decision submodule and a multi-view segment association submodule; The hierarchical acquisition decision submodule is used to configure different data writing levels for candidate acquisition units according to the consistency discrimination results. The consistent synchronization class corresponds to complete writing, the cross-destination compensation class corresponds to retaining compensable destination data and abnormal destination data index, the continuous repetition class corresponds to key segment writing, and the unusable class corresponds to discarding record writing. The multi-view segment association submodule is used to establish viewpoint association identifiers for different viewpoint data retained in the same candidate acquisition unit, and to associate and record redundant data that has not been written with summary identifiers and original location identifiers.

6. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 1, characterized in that, The timestamp normalization processing submodule includes a drift fragment identification submodule and a time window reconstruction submodule; The drift segment identification submodule is used to identify target data segments with time drift based on changes in cross-target arrival order, abrupt changes in frame number intervals, and deviations in visual motion trends within adjacent time windows in continuous candidate acquisition units. The time window reconstruction submodule is used to reconfigure the time boundaries of the affected candidate acquisition units after identifying the time-drifted target data segment, and to re-aggregate the multiple image frames or video segments before and after the drift into reconstructed candidate acquisition units, so that the candidate acquisition units do not use a single path timestamp as the sole alignment basis.

7. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 3, characterized in that, The viewpoint complementarity analysis submodule includes an occlusion anomaly localization submodule and a compensation value assessment submodule; The occlusion anomaly localization submodule is used to compare the sharpness distribution, edge breakage distribution, and local texture loss distribution of different target image frames within the same candidate acquisition unit, and generate anomaly region location records. The compensation value assessment submodule is used to perform cross-view correspondence between the abnormal region location record and the non-abnormal region in other target image frames, generate a compensable region representation for abnormal target data, and incorporate the compensable region representation into the viewpoint complementary representation to distinguish between data segments that are locally abnormal but still have retention value and data segments that are unusable overall.

8. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 4, characterized in that, The consistency classification submodule includes a temporal continuity constraint submodule and a category stability correction submodule; The temporal continuity constraint submodule is used to sequentially associate the multi-view acquisition state vector of the current candidate acquisition unit with the multi-view acquisition state vector of the adjacent candidate acquisition units to form a temporal discrimination sequence including the current state, the previous state, and the subsequent state. The category stability correction submodule is used to perform secondary classification on isolated local anomaly classes, continuous repetition classes and unusable classes according to the time-series discrimination sequence, and output the secondary classification result as the updated consistency discrimination result to the acquisition strategy control module.

9. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 6, characterized in that, The time window reconstruction submodule includes a reconstruction version management submodule and a data acquisition link marking submodule; The reconstruction version management submodule is used to configure version identifiers for the pre-reconstruction candidate acquisition units and the reconstruction candidate acquisition units respectively, and to record the source of target data, time boundary change information and frame sequence number mapping information for participating in the reconstruction. The acquisition link marking submodule is used to write the version identifier, the time boundary change information and the frame sequence number mapping information into the index record formed by the acquisition index writing module, so that the data fragments reconstructed by the time window retain the correspondence before and after reconstruction in the data management layer.

10. The intelligent sensing big data acquisition system for a first-view multi-camera according to claim 8, characterized in that, The category stability correction submodule includes a policy backtracking update submodule and a label consistency maintenance submodule; The strategy backtracking update submodule is used to backtrack and update the acquisition control strategy that has been generated but not yet solidified when the secondary classification results of adjacent candidate acquisition units change continuously, and send the updated effective acquisition, downsampling acquisition, cached acquisition or discard acquisition control strategy to the acquisition index writing module. The label consistency maintenance submodule is used to synchronously update quality labels, viewpoint association labels, time window version identifiers, and control strategy identifiers, so that multiple types of index labels for the same acquisition segment maintain consistent reference relationships.

Images