A vehicle feature recognition method and system based on spatial computing

Abstract

This invention discloses a vehicle feature recognition method and system based on spatial computing, belonging to the field of vehicle recognition technology. The method includes the following steps: acquiring and standardizing multi-source spatial perception data and vehicle image data in a target vehicle recognition scene; performing spatial occupancy relationship parsing on the standardized spatial perception dataset; detecting and segmenting the vehicle target; marking the occupancy state of the vehicle target; performing spatial occupancy inference processing through an improved PoinTr network to generate a vehicle invisible inference feature set; and performing part matching, confidence correction, and fusion recognition processing on the vehicle visible feature set and the vehicle invisible inference feature set to generate a vehicle fusion feature recognition result. This invention utilizes spatial computing and occupancy inference to achieve occluded vehicle feature completion and fusion recognition, possessing advantages such as strong occupancy adaptability, high recognition accuracy, and high feature completeness.

Description

Technical Field

[0001] This invention relates to the field of vehicle recognition technology, and in particular to a vehicle feature recognition method and system based on spatial computing. Background Technology

[0002] With the development of intelligent transportation, smart security, and autonomous driving technologies, vehicle feature recognition technology based on video surveillance has been widely applied. Existing vehicle recognition methods typically utilize vehicle images captured by cameras to detect and identify license plates, vehicle models, body colors, and other vehicle exterior features, thereby enabling functions such as vehicle retrieval, vehicle tracking, and vehicle management.

[0003] In real-world applications, vehicles are often affected by factors such as pedestrians, guardrails, trees, other vehicles, and camera viewing angles, resulting in parts of the vehicle being obscured or invisible. Existing technologies mostly rely on the visible areas of the vehicle for feature extraction and recognition, lacking the comprehensive analytical capabilities to consider the vehicle's spatial occupancy and occupancy relationships. This makes it difficult to accurately recover the feature information corresponding to the invisible parts of the vehicle, easily leading to problems such as incomplete license plate recognition, biased vehicle model identification, and missing vehicle features, thereby reducing the accuracy and completeness of vehicle feature recognition in complex scenarios.

[0004] Therefore, how to provide a vehicle feature recognition method and system based on spatial computing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a vehicle feature recognition method and system based on spatial computing. This invention utilizes spatial computing and occupancy reasoning to achieve feature completion and fusion recognition of occluded vehicles, and has the advantages of strong occupancy adaptability, high recognition accuracy and high feature completeness.

[0006] A vehicle feature recognition method based on spatial computation according to an embodiment of the present invention includes the following steps: Acquire multi-source spatial perception data and vehicle image data in the target vehicle recognition scenario, and perform standardization processing to generate standardized spatial perception dataset and standardized vehicle image dataset; The standardized spatial perception dataset is parsed to determine the candidate spatial regions of vehicles, the spatial regions of occupants, the visible areas of cameras, and the deep correlation between each spatial region, thereby generating scene spatial occupancy data. Vehicle targets are detected and segmented based on a standardized vehicle image dataset, generating a set of visible vehicle features; Based on the scene space occupancy relationship data and the vehicle visible feature set, the occupancy status of the vehicle target is marked, and a vehicle space occupancy feature set is generated. The improved PoinTr network is used to perform spatial occupancy reasoning on the vehicle spatial occupancy feature set and vehicle structural constraint data to generate a vehicle invisible reasoning feature set. The visible feature set and the invisible inference feature set of the vehicle are subjected to part matching, confidence correction and fusion recognition processing to generate vehicle fusion feature recognition results.

[0007] Optionally, the multi-source spatial perception data includes scene 3D point cloud data, camera pose data, spatial depth data, scene structure data, occlusion spatial location data, and vehicle passage area data. The vehicle image data includes vehicle video frames, vehicle appearance time, vehicle image acquisition viewpoint, vehicle local image region, and vehicle image quality parameters. The standardization processing includes timestamp unification, spatial coordinate alignment, sensor pose correction, data format unification, abnormal data removal, missing data completion, and data scale normalization processing.

[0008] Optionally, the generation of the scene space occupancy relationship data specifically includes: Based on the scene 3D point cloud data and scene structure data in the standardized spatial perception dataset, the target vehicle recognition scene is divided into spatial grids to generate a set of scene spatial units. Based on the spatial location data of the occlusion objects and the scene structure data, fixed occlusion areas and dynamic occlusion areas are identified from the set of scene spatial units to determine the spatial area of ​​the occlusion objects; Based on spatial depth data and vehicle traffic area data, spatial units with vehicle traffic conditions are selected from the scene spatial unit set to determine candidate vehicle spatial areas; Based on camera pose data, spatial depth data, and scene structure data, the shooting coverage of each camera in the target vehicle recognition scene is projected and analyzed to determine the camera's visible area. Spatial depth analysis is performed based on the vehicle candidate space area, the obstruction space area, and the camera visible area to determine the depth correlation between each space area. The candidate space area of ​​the vehicle, the space area of ​​the obstruction, and the visible area of ​​the camera are marked and their coordinates are bound. Spatial association matching is performed in combination with deep correlation to generate scene space occupancy relationship data.

[0009] Optionally, the generation of the vehicle visible feature set specifically includes: Based on the vehicle video frames, vehicle appearance time and vehicle image acquisition angle in the standardized vehicle image dataset, the vehicle video frames under the same acquisition angle are time-series organized to generate a vehicle image frame sequence. Vehicle target detection processing is performed on the vehicle image frame sequence to extract the image region where the vehicle target is located and generate the vehicle target detection region. Foreground segmentation is performed on the vehicle target detection area to separate the vehicle target area from the non-vehicle background area and generate vehicle visible area features. Boundary extraction and contour fitting are performed on the vehicle target area to generate vehicle edge contour features; Perform color distribution statistics and color region segmentation on the body pixels in the vehicle target area to generate vehicle color distribution features; The local structural regions in the target area of ​​the vehicle are identified, and the window area, headlight area, wheel area, license plate area and local component areas of the vehicle body are extracted to generate local structural features of the vehicle. The visible area features of the vehicle, the edge contour features of the vehicle, the color distribution features of the vehicle, and the local structural features of the vehicle are correlated and organized to generate a set of visible features of the vehicle.

[0010] Optionally, the generation of the vehicle space occupancy feature set specifically includes: Based on the vehicle spatial location relationship, visible area relationship and spatial depth relationship in the scene spatial occupancy relationship data, the set of visible vehicle features is mapped to the vehicle candidate spatial region to generate a spatial mapping relationship. Based on the vehicle's visible area features, vehicle edge contour features, and spatial mapping relationships, the visible spatial range of the vehicle target in the vehicle candidate spatial region is marked to generate the vehicle's visible occupied part. Based on the spatial position relationship, spatial depth relationship and occlusion boundary position in the vehicle edge contour features, the vehicle parts covered by the occluded object spatial region in the vehicle candidate spatial region are marked to generate the vehicle occluded part. Based on the visible area relationship, spatial adjacency relationship and local structural features of the vehicle, invisible vehicle parts that do not fall into the camera's visible area and are not covered by the vehicle's visible occupied parts in the candidate vehicle spatial area are marked as invisible, thus generating invisible vehicle occupied parts. Based on the spatial location and depth relationships of the obstruction space, the source of the vehicle's obstructed area is matched with the corresponding obstruction space area to generate the obstruction source and obstruction depth relationship; Based on the relationship between the visible and invisible parts of the vehicle, the occupied parts of the vehicle, the occupied parts of the vehicle, the occupancy source and the occupancy depth, the vehicle parts are matched to the visible feature set to generate the spatial correspondence of vehicle parts. The visible and invisible occupied parts of the vehicle, the occupied parts of the vehicle, the occupied parts of the vehicle, the source of occupancy, the occupancy depth relationship, and the spatial correspondence of vehicle parts are correlated and organized to generate a set of vehicle spatial occupancy features.

[0011] Optionally, the generation of the vehicle invisible inference feature set specifically includes: The vehicle spatial occupancy feature set and vehicle structural constraint data are input into the improved PoinTr network, and point cloud construction and grouping are performed through the vehicle local point cloud grouping unit to generate vehicle local point cloud groups. The improvement of the improved PoinTr network compared with the original PoinTr network is that a vehicle spatial location embedding unit is added between the part point proxy encoding unit and the missing part Transformer decoding unit, and an occlusion encoding unit is added between the vehicle spatial location embedding unit and the missing part Transformer decoding unit. The vehicle local point cloud group is processed by point proxy coding unit, and the visible occupied parts, invisible occupied parts, occupied parts and the part status information corresponding to the occupancy source are integrated to generate vehicle part point proxy data. Based on vehicle contour constraint data, license plate area constraint data, relative position constraint data of vehicle parts and vehicle size constraint data, the vehicle spatial position embedding unit performs spatial position embedding processing on the proxy data of vehicle parts points to generate vehicle spatial position embedding features. The occlusion coding unit encodes the vehicle spatial location embedding features, occlusion source, and occlusion depth relationship to generate occlusion coding features. The missing part decoding features are processed by the missing part Transformer decoding unit to generate missing part decoding features corresponding to the invisible occupied part of the vehicle and the occluded part of the vehicle. The missing parts are filled in by using vehicle occupancy completion units to generate a vehicle occupancy point cloud. The invisible feature output unit performs part matching and feature transformation on the vehicle completion occupied point cloud and the vehicle visible feature set to generate the vehicle invisible inference feature set.

[0012] Optionally, the generation of the vehicle fusion feature recognition result specifically includes: Based on the spatial correspondence of vehicle parts, perform part matching on the visible feature set and the invisible inference feature set of the vehicle to generate a vehicle matching feature set. Based on the confidence level of the visible area and the confidence level of the inference features, the vehicle matching feature set is corrected to generate a vehicle corrected feature set; Based on the vehicle correction feature set, license plate missing inference features and license plate region constraint data, license plate region fusion recognition is performed to generate license plate recognition results; Based on the vehicle correction feature set, vehicle contour completion feature and vehicle component completion feature, vehicle model structure matching is performed to generate vehicle model recognition results; Based on vehicle color distribution characteristics, vehicle color consistency characteristics, and vehicle side inference characteristics, vehicle body color fusion is performed to generate vehicle body color recognition results. Based on the license plate recognition results, vehicle model recognition results, vehicle body color recognition results, and occupancy statistics, the result fields are correlated, recognition confidence is summarized, and completeness is evaluated to generate vehicle fusion feature recognition results.

[0013] A vehicle feature recognition system based on spatial computing according to an embodiment of the present invention includes: The data processing module is used to acquire and standardize multi-source spatial perception data and vehicle image data; The occupancy analysis module is used to analyze the spatial occupancy relationship of the standardized spatial perception dataset and generate scene spatial occupancy relationship data. The visible recognition module is used to detect and segment vehicle targets, and generate a set of visible vehicle features; The occupancy marking module is used to mark the occupancy status of vehicle targets and generate a set of vehicle space occupancy features; The inference completion module is used to perform spatial occupancy inference processing on the vehicle spatial occupancy feature set and vehicle structural constraint data through the improved PoinTr network, and generate a vehicle invisible inference feature set. The fusion recognition module is used to perform part matching, confidence correction and fusion recognition processing on the vehicle's visible feature set and the vehicle's invisible inference feature set to generate vehicle fusion feature recognition results.

[0014] The beneficial effects of this invention are: This invention constructs scene spatial occupancy relationship data, unifying the modeling of vehicle candidate spatial regions, occlusion spatial regions, camera visible regions, and the deep correlations between these spatial regions. It then analyzes and labels the occupancy state of vehicle targets by combining this with a set of visible vehicle features, accurately distinguishing between visible, occluded, and invisible vehicle occupancy areas. Compared to existing technologies that rely solely on two-dimensional images for vehicle recognition, this invention fully utilizes scene spatial structure information and occupancy relationship information to achieve a correspondence analysis between vehicle features and spatial occupancy states. This improves the accuracy of vehicle state description in complex scenes and provides a reliable data foundation for subsequent inference using invisible features.

[0015] This invention utilizes an improved PoinTr network to perform spatial occupancy inference processing on the vehicle spatial occupancy feature set. It adds vehicle spatial position embedding units and occlusion encoding units to the original network structure, enabling the model to simultaneously perceive vehicle structural constraints and spatial occlusion relationships. By incorporating vehicle contour constraint data, license plate region constraint data, relative position constraint data of vehicle parts, and vehicle size constraint data into the feature learning process, the model's ability to express the overall structural consistency of the vehicle is enhanced. Furthermore, by encoding the occlusion source and occlusion depth relationships, the model's inference ability for occluded and missing regions is improved. This allows for the reasonable completion of invisible vehicle areas, obtaining license plate missing inference features, vehicle contour completion features, vehicle part completion features, and vehicle color consistency features, effectively reducing recognition errors caused by missing vehicle features in complex occlusion scenarios.

[0016] This invention further performs part matching, confidence correction, and fusion recognition processing on the vehicle's visible feature set and the vehicle's invisible inferred feature set. By fusing the vehicle's actual observed features with inferred completion features, it achieves joint optimization of license plate recognition results, vehicle model recognition results, and vehicle body color recognition results, and evaluates the completeness of the recognition results based on the vehicle's occupancy status. Compared to traditional vehicle recognition methods that rely solely on visible areas for judgment, this invention maintains high recognition accuracy and completeness even with partial vehicle occlusion, missing vehicle features, and complex monitoring scenarios, improving vehicle feature recovery capabilities, vehicle feature recognition capabilities, and the system's robustness and practicality in complex scenarios. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a vehicle feature recognition method based on spatial computing proposed in this invention; Figure 2 This is a flowchart illustrating the generation of a vehicle space occupancy feature set in a vehicle feature recognition method based on spatial computing proposed in this invention. Figure 3 This is a flowchart of the spatial occupancy reasoning process for a vehicle feature recognition method based on spatial computation proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figures 1-3A vehicle feature recognition method based on spatial computing includes the following steps: Acquire multi-source spatial perception data and vehicle image data in the target vehicle recognition scenario, and perform standardization processing to generate standardized spatial perception dataset and standardized vehicle image dataset; The standardized spatial perception dataset is parsed to determine the candidate spatial regions of vehicles, the spatial regions of occupants, the visible areas of cameras, and the deep correlation between each spatial region, thereby generating scene spatial occupancy data. Vehicle targets are detected and segmented based on a standardized vehicle image dataset, generating a set of visible vehicle features; Based on the scene space occupancy relationship data and the vehicle visible feature set, the occupancy status of the vehicle target is marked, and a vehicle space occupancy feature set is generated. The improved PoinTr network is used to perform spatial occupancy reasoning on the vehicle spatial occupancy feature set and vehicle structural constraint data to generate a vehicle invisible reasoning feature set. The visible feature set and the invisible inference feature set of the vehicle are subjected to part matching, confidence correction and fusion recognition processing to generate vehicle fusion feature recognition results.

[0020] In this embodiment, the multi-source spatial perception data includes scene 3D point cloud data, camera pose data, spatial depth data, scene structure data, occlusion spatial location data, and vehicle passage area data. The vehicle image data includes vehicle video frames, vehicle appearance time, vehicle image acquisition viewpoint, vehicle local image area, and vehicle image quality parameters. The standardization processing includes timestamp unification, spatial coordinate alignment, sensor pose correction, data format unification, abnormal data removal, missing data completion, and data scale normalization processing.

[0021] In this embodiment, the generation of scene space occupancy relationship data specifically includes: Based on the scene 3D point cloud data and scene structure data in the standardized spatial perception dataset, the target vehicle recognition scene is divided into spatial grids to generate a set of scene spatial units. Based on the spatial location data of the occlusion objects and the scene structure data, fixed occlusion areas and dynamic occlusion areas are identified from the set of scene spatial units to determine the spatial area of ​​the occlusion objects; Based on spatial depth data and vehicle passage area data, spatial units with vehicle passage conditions are selected from the set of scene spatial units to determine candidate vehicle spatial areas. Vehicle passage conditions include: the spatial unit is located within the vehicle passage area; the spatial unit is not occupied by an obstruction; the spatial unit meets the vehicle occupancy size requirements; and there is a connectivity relationship between the spatial unit and adjacent spatial units. Vehicle occupancy size requirements include the minimum vehicle passage width, the minimum vehicle passage height, and the minimum vehicle turning radius. Based on camera pose data, spatial depth data, and scene structure data, the shooting coverage of each camera in the target vehicle recognition scene is projected and analyzed to determine the camera's visible area. Spatial depth analysis is performed on the vehicle candidate space region, the occlusion space region, and the camera's visible area to determine the depth correlation between each space region. Specifically, the overlapping area of ​​the projected areas of the vehicle candidate space region and the occlusion space region in the camera's visible direction is calculated, and the ratio of the overlapping area to the total projected area of ​​the vehicle candidate space region is calculated to obtain the occlusion overlap ratio. The spatial distance between the occlusion space region and the vehicle candidate space region is calculated, and the spatial distance is normalized to obtain a distance attenuation value. It is determined whether the occlusion space region is located between the camera position and the vehicle candidate space region, and front and rear occlusion indicators are generated. The occlusion overlap ratio and distance attenuation are then compared. The occlusion correlation degree is generated by weighted fusion calculation of the reduction value and the front and rear occlusion indicators; the projected area of ​​the vehicle candidate space region falling into the camera's visible area is calculated, and the ratio of the projected area to the total projected area of ​​the vehicle candidate space region is calculated to obtain the visible coverage ratio; the depth distance between the vehicle candidate space region and the camera position is calculated, and the depth distance is normalized to obtain the depth distance attenuation value; the visible coverage ratio, occlusion correlation degree, and depth distance attenuation value are weighted fusion calculation to generate the visible correlation degree; the occlusion correlation degree and the visible correlation degree are weighted fusion calculation to obtain the correlation depth value, and the correlation depth value is sorted according to the correlation depth value to generate the depth correlation relationship; The process involves marking and binding coordinates to the candidate vehicle space region, the occlusion space region, and the camera's visible region. Spatial association matching is then performed using depth relationships to generate scene spatial occupancy data. Specifically: the spatial coordinate difference between the candidate vehicle space region and the occlusion space region is calculated to generate the vehicle's spatial position relationship; the overlap between the candidate vehicle space region and the occlusion space region is calculated to generate the occlusion spatial position relationship; the coverage between the candidate vehicle space region and the camera's visible region is calculated to generate the visible region relationship; the association depth value and sorting results in the depth relationships are extracted to generate the spatial depth relationship; the connectivity between the candidate vehicle space region, the occlusion space region, and the camera's visible region is calculated to generate the spatial adjacency relationship; and the vehicle spatial position relationship, the occlusion spatial position relationship, the visible region relationship, the spatial depth relationship, and the spatial adjacency relationship are integrated to generate scene spatial occupancy data.

[0022] In this embodiment, the generation of the vehicle visible feature set specifically includes: Based on the vehicle video frames, vehicle appearance time and vehicle image acquisition angle in the standardized vehicle image dataset, the vehicle video frames under the same acquisition angle are time-series organized to generate a vehicle image frame sequence. Vehicle target detection processing is performed on the vehicle image frame sequence to extract the image region where the vehicle target is located and generate the vehicle target detection region. Foreground segmentation is performed on the vehicle target detection region to separate the vehicle target region from the non-vehicle background region, generating vehicle visible region features. The vehicle visible region features include the visible region area, visible region location, visible region shape, visible region boundary, and visible region confidence. The visible region confidence is calculated by weighted fusion of the segmentation confidence, boundary integrity, and visible occlusion overlap of the vehicle target region. Specifically, the segmentation confidence is obtained by the average probability of each pixel in the vehicle target region belonging to the vehicle category, the boundary integrity is obtained by the ratio of the actual boundary length of the vehicle target region to the predicted boundary length of the vehicle in the vehicle target detection region, and the visible occlusion overlap is obtained by the proportion of the overlap area between the vehicle target region and the occlusion boundary region to the area of ​​the vehicle target region. The vehicle target region is subjected to boundary extraction and contour fitting to generate vehicle edge contour features. The vehicle edge contour features include the outer contour curve of the vehicle target region, contour key points, contour direction change information and contour continuity information. The contour continuity information includes the connection status between adjacent contour points in the outer contour curve of the vehicle target region, contour break position, contour break length, contour missing ratio and occlusion boundary position. The color distribution statistics and color region division of the vehicle body pixels in the vehicle target area are performed to generate vehicle color distribution features. The vehicle color distribution features include the main color category, color region position, color proportion and color consistency information in the vehicle target area. The local structural regions in the target area of ​​the vehicle are identified, and the window area, headlight area, wheel area, license plate area and local component areas of the vehicle body are extracted to generate local structural features of the vehicle. The visible area features of the vehicle, the edge contour features of the vehicle, the color distribution features of the vehicle, and the local structural features of the vehicle are correlated and organized to generate a set of visible features of the vehicle.

[0023] In this embodiment, the generation of the vehicle space occupancy feature set specifically includes: Based on the vehicle spatial location relationship, visible area relationship and spatial depth relationship in the scene spatial occupancy relationship data, the set of visible vehicle features is mapped to the vehicle candidate spatial region to generate a spatial mapping relationship. Based on the vehicle's visible area features, vehicle edge contour features, and spatial mapping relationships, the visible spatial range of the vehicle target within the vehicle candidate space region is marked to generate the vehicle's visible occupied parts. Specifically, this involves: reading the visible area position and visible area boundary from the vehicle's visible area features, and mapping the visible area position and visible area boundary to the vehicle candidate space region according to the spatial mapping relationship; reading the outer contour curve and contour key points from the vehicle's edge contour features, and performing contour constraint correction on the mapped visible area boundary to obtain the vehicle's visible spatial range; configuring visible occupied markers for the spatial units covered by the vehicle's visible spatial range to generate the vehicle's visible occupied parts. Based on the spatial position relationship, spatial depth relationship, and occlusion boundary position in the vehicle edge contour features, occlusion markings are applied to vehicle parts covered by occluded object spatial regions within the vehicle candidate spatial region to generate occluded and occupied vehicle parts. Specifically, this involves: reading the region overlap in the spatial position relationship and the associated depth value in the spatial depth relationship to determine the overlapping region between the vehicle candidate spatial region and the occluded object spatial region; extracting the occlusion boundary position from the vehicle edge contour features and mapping the occlusion boundary position to the overlapping region; configuring occlusion and occupied markers for spatial units located behind the occlusion boundary position and within the overlapping region to generate occluded and occupied vehicle parts. Based on the visible area relationship, spatial adjacency relationship, and local structural features of the vehicle, invisible occupants of the vehicle are marked as not falling within the camera's visible area and not covered by the vehicle's visible occupants in the candidate spatial area. Specifically, the process involves: reading the area coverage information in the visible area relationship and filtering spatial units in the candidate spatial area that do not fall within the camera's visible area; removing spatial units with visible occupant markers and spatial units with occupant obstruction markers; extracting spatial units adjacent to the vehicle's visible occupants based on spatial adjacency relationships; and configuring invisible occupant markers on the remaining spatial units based on the spatial positions of the window area, headlight area, wheel area, license plate area, and body component areas in the vehicle's local structural features, thereby generating invisible occupants of the vehicle. Based on the spatial location and depth relationships of the occupants, the source of the vehicle's occupant is matched with the corresponding occupant space area to generate an occupant source and occupant depth relationship. Specifically: the occupant identifier corresponding to the vehicle's occupant is read, and the occupant space areas that overlap with the vehicle's occupant are extracted; the percentage of overlap between each occupant space area and the vehicle's occupant is calculated, and the occupant space area with the largest overlap percentage is selected as the corresponding occupant space area to generate the occupant source; the associated depth value and sorting result in the spatial depth relationship are extracted, the depth difference between the corresponding occupant space area and the vehicle's occupant is calculated, and a correspondence is established according to the depth difference and associated depth value to generate the occupant depth relationship. Based on the relationships between the visible and invisible parts of the vehicle, the occluded parts, the occluded source, and the occluded depth, vehicle parts matching is performed on the set of visible features to generate spatial correspondences of vehicle parts. These spatial correspondences include the correspondence between visible area features and visible parts, the correspondence between edge contour features and visible parts, the correspondence between edge contour features and occluded parts, the correspondence between color distribution features and visible parts, and the correspondence between local structural features and visible, invisible, and occluded parts. The visible and invisible occupied parts of the vehicle, the occupied parts of the vehicle, the occupied parts of the vehicle, the source of occupancy, the occupancy depth relationship, and the spatial correspondence of vehicle parts are correlated and organized to generate a set of vehicle spatial occupancy features.

[0024] In this embodiment, the generation of the vehicle invisible reasoning feature set specifically includes: The improved PoinTr network inputs vehicle spatial occupancy feature set and vehicle structural constraint data, and generates vehicle local point cloud groups through vehicle local point cloud grouping units. Specifically, the spatial coordinate information corresponding to the visible, invisible, and occluded parts of the vehicle is converted into occupancy point coordinates, and the corresponding vehicle part category information is configured according to the spatial correspondence of vehicle parts. The corresponding occupancy status information is configured to the occupancy points, forming a three-dimensional point cloud data composed of multiple occupancy points as the vehicle occupancy point cloud. Based on vehicle contour constraint data, license plate area constraint data, relative position constraint data of vehicle parts, and vehicle size constraint data, abnormal points and points exceeding the constraint range in the vehicle occupancy point cloud are removed, and the remaining points are subjected to position correction processing. The spatial distance between each point in the vehicle occupancy point cloud is calculated through the vehicle local point cloud grouping unit, and points with a spatial distance less than a preset distance threshold are grouped into the same point cloud group to generate vehicle local point cloud groups. The preset distance threshold is calculated based on the average spatial distance between adjacent vehicle parts in the relative position constraint data of vehicle parts. The improvements of the improved PoinTr network compared to the existing PoinTr network are as follows: First, a vehicle spatial location embedding unit is added between the part point proxy encoding unit and the missing part Transformer decoding unit. This enables the model to have vehicle structure perception capabilities during feature learning, improves the ability to express spatial relationships between different parts of the vehicle, enhances the overall structural constraint capability of the vehicle, reduces part misalignment, contour distortion, and structural discontinuity problems during the completion process, and improves the consistency and rationality between the vehicle completion result and the real vehicle structure. Second, an occlusion encoding unit is added between the vehicle spatial location embedding unit and the missing part Transformer decoding unit. This enables the model to have occlusion relationship perception capabilities during feature learning, enhances the model's ability to express spatial occlusion relationships between the vehicle and occluded objects, improves the ability to identify occluded areas, missing areas, and invisible areas, reduces erroneous and invalid completion in complex occlusion scenarios, and improves vehicle feature recovery capability, vehicle completion accuracy, and the robustness of the model in complex scenarios. The improved PoinTr network's total loss function is a weighted average of the occupancy point cloud reconstruction loss, vehicle contour constraint loss, vehicle part position constraint loss, and occlusion consistency loss. Specifically, the occupancy point cloud reconstruction loss is calculated using the chamfer distance loss function, the vehicle contour constraint loss using the Hausdorff distance loss function, the vehicle part position constraint loss using the mean squared error loss function, and the occlusion consistency loss using the cross-entropy loss function. The improved PoinTr network uses an adaptive moment estimation optimization algorithm for parameter updates, with a learning rate of 0.0001–0.001, a batch size of 16–64, and a training epoch count of 100–500. The improved PoinTr network is considered to have reached convergence when the variation of the total loss function over multiple consecutive training epochs falls below a preset convergence threshold. This preset convergence threshold is calculated based on the average variation of the total loss function over multiple consecutive training epochs. Vehicle contour constraint data includes vehicle outer contour boundary, vehicle contour continuity information, and vehicle contour size range information; license plate area constraint data includes license plate spatial position range, license plate length-to-width ratio range, and license plate installation orientation range; vehicle part relative position constraint data includes the relative positional relationship and spatial spacing information between window area, headlight area, wheel area, license plate area, and body part area; vehicle size constraint data includes vehicle length range, vehicle width range, vehicle height range, wheelbase range, and track width range. The vehicle local point cloud group is processed by a point proxy encoding unit, and the visible occupied parts, invisible occupied parts, occupied occupied parts, and the part status information corresponding to the occupancy source are fused to generate vehicle part point proxy data. Specifically, the mean coordinates of the occupied points in each vehicle local point cloud group are calculated to obtain the center coordinates of the point cloud group; the coordinate difference between each occupied point coordinate and the center coordinates of the point cloud group is calculated to obtain local geometric offset data; the proportion of visible occupied markers, invisible occupied markers, and occupied occupied markers in the vehicle local point cloud group is counted to obtain occupancy status distribution data; the spatial region identifier of the occupancy source corresponding to the occupancy source is configured as occupancy source encoding data; the center coordinates of the point cloud group, local geometric offset data, occupancy status distribution data, and occupancy source encoding data are concatenated to generate point proxy input data; the numerical normalization of each data item in the point proxy input data is performed, and compression encoding and vector conversion are performed according to the field order in the input data to generate vehicle part point proxy data. Based on vehicle contour constraint data, license plate area constraint data, relative position constraint data of vehicle parts, and vehicle size constraint data, a vehicle spatial position embedding unit is used to perform spatial position embedding processing on the vehicle part point proxy data to generate vehicle spatial position embedding features. Specifically: based on the center coordinates of the point cloud group in the vehicle part point proxy data, the distance between the center coordinates of the point cloud group and the outer contour boundary of the vehicle is calculated to generate contour constraint distance data; based on the license plate area constraint data, the distance between the center coordinates of the point cloud group and the spatial position range of the license plate is calculated to generate license plate area distance data; based on the relative position constraint data of vehicle parts, the spatial distance between the center coordinates of the point cloud group and the corresponding center coordinates of the point cloud group of adjacent vehicle parts is calculated to generate part spacing data; based on the vehicle size constraint data, the positional proportion of the center coordinates of the point cloud group in the vehicle length range, vehicle width range, and vehicle height range is calculated to generate size position data; the contour constraint distance data, license plate area distance data, part spacing data, and size position data are configured to the corresponding vehicle part point proxy data and combined and spliced ​​to generate vehicle spatial position embedding features. The occlusion coding unit encodes the vehicle spatial location embedding features, occlusion sources, and occlusion depth relationships to generate occlusion coding features. Specifically, it determines the occlusion object spatial region corresponding to the vehicle spatial location embedding feature based on the occlusion source; extracts the associated depth value and depth ranking result corresponding to the vehicle spatial location embedding feature based on the occlusion depth relationship; calculates the depth difference between each vehicle spatial location embedding feature and the corresponding occlusion object spatial region, and establishes the occlusion order based on the depth ranking result; configures the occlusion object spatial region identifier, associated depth value, depth difference, and occlusion order to the corresponding vehicle spatial location embedding feature; and performs feature concatenation processing on the configured vehicle spatial location embedding features to generate occlusion coding features. The missing part decoding unit performs missing part decoding on the occlusion coding features to generate missing part decoding features corresponding to the invisible vehicle-occupied parts and the vehicle-occupied parts. Specifically, based on the spatial region identifier of the occluding object, the associated depth value, the depth difference, and the occlusion order in the occlusion coding features, the occlusion coding features corresponding to the invisible vehicle-occupied parts and the vehicle-occupied parts are selected; the spatial distance between the center coordinates of the point cloud groups corresponding to each occlusion coding feature is calculated, and the numerical difference between the associated depth values ​​corresponding to each occlusion coding feature is calculated; occlusion coding features with a spatial distance less than the associated distance threshold and an associated depth value difference less than a preset depth threshold are classified into the same group. The association set is used to arrange the occlusion coding features in the same association set according to the order of occlusion. Based on the vehicle spatial location embedding features in the arranged occlusion coding features, spatial location information, part contour information, and part structure information are recovered. The spatial location information, part contour information, and part structure information are sequentially concatenated according to the arrangement result to generate part completion information. The part completion information is mapped to the corresponding invisible vehicle-occupied parts and occupied vehicle-occupied parts to generate missing part decoding features. The association distance threshold is determined based on the average spatial distance between adjacent vehicle parts in the relative position constraint data of vehicle parts. The preset depth threshold is determined based on the average difference of association depth values ​​in the occlusion depth relationship. The missing parts are filled in using vehicle occupancy completion units to generate a vehicle-completed occupancy point cloud. Specifically: the spatial coordinates of each completion point are calculated based on spatial location information; the connection relationships between completion points are determined based on part contour information, and the spatial contour of the completed part is constructed according to these relationships; the vehicle part category and occupancy status corresponding to each completion point are determined based on part structure information; the spatial coordinates, spatial contour, vehicle part category, and occupancy status are configured to the corresponding completion points to generate completed occupancy points; the completed occupancy points are matched with the occupancy points in the vehicle local point cloud group according to the spatial coordinate correspondence; the matched occupancy points are corrected based on vehicle contour constraint data, license plate area constraint data, vehicle part relative position constraint data, and vehicle size constraint data; and the corrected occupancy points are merged with the occupancy points in the vehicle local point cloud group to generate the vehicle-completed occupancy point cloud. The invisible feature output unit performs part matching and feature transformation on the vehicle completion point cloud and the vehicle visible feature set to generate a vehicle invisible inference feature set. Specifically, it establishes the correspondence between the completion points in the vehicle completion point cloud and the vehicle visible area features, vehicle edge contour features, vehicle color distribution features, and vehicle local structural features in the vehicle visible feature set based on the spatial correspondence of vehicle parts; it sorts the completion points corresponding to the same vehicle part according to spatial coordinates and calculates the connection relationship between adjacent completion points to generate completion position contour data; it splices the completion position contour data with the outer contour curve in the vehicle edge contour features to generate vehicle contour completion features; and it associates the vehicle part category corresponding to the completion point with the window area, headlight area, wheel area, license plate area, and body part in the vehicle local structural features. The system performs position matching on component regions to generate vehicle component completion features; it maps the spatial coordinates corresponding to the completion points to the spatial location range of the license plate in the license plate area constraint data to generate license plate missing inference features; it associates the spatial coordinates corresponding to the completion points and the completion location contour data with the color region location and color proportion in the vehicle color distribution features to generate vehicle side inference features and vehicle color consistency features; it calculates the proportion of corresponding parts, contours, and colors between the vehicle completion point cloud and the vehicle visible feature set, and performs weighted summation of each proportion to generate inference feature confidence; it then associates and organizes the license plate missing inference features, vehicle side inference features, vehicle contour completion features, vehicle component completion features, vehicle color consistency features, and inference feature confidence to generate a vehicle invisible inference feature set.

[0025] In this embodiment, the generation of vehicle fusion feature recognition results specifically includes: Based on the spatial correspondence of vehicle parts, the visible feature set and the invisible inference feature set of the vehicle are matched to generate the vehicle matching feature set; specifically, the visible area features, edge contour features, color distribution features, and local structural features of the vehicle are matched with the license plate missing inference features, side body inference features, vehicle contour completion features, vehicle component completion features, and vehicle color consistency features according to the same vehicle parts. Based on the confidence scores of the visible area and the inference features, the vehicle matching feature set is corrected to generate a vehicle correction feature set. Specifically, the confidence scores of the visible area and the inference features corresponding to the same vehicle part are normalized; state weights are configured according to the occupancy states corresponding to the visible, invisible, and occupied parts of the vehicle; the normalized confidence scores of the visible area and the inference features are weighted and fused according to the state weights; and the corresponding features in the vehicle matching feature set are updated based on the weighted fusion calculation results to generate the vehicle correction feature set. Based on the vehicle correction feature set, license plate missing inference features, and license plate region constraint data, license plate region fusion recognition is performed to generate license plate recognition results. Specifically, the visible license plate area in the vehicle correction feature set is aligned with the license plate completion area corresponding to the license plate missing inference features. The aligned license plate area is constrained and filtered according to the license plate spatial location range, license plate length-width ratio range, and license plate installation orientation range. The filtered license plate area is then spliced ​​to generate a complete license plate area. Finally, the characters in the complete license plate area are arranged in order and combined according to the character distribution information to generate the license plate recognition results. Vehicle model structure matching is performed based on vehicle correction feature set, vehicle contour completion feature, and vehicle component completion feature to generate vehicle model recognition results. Specifically, the vehicle edge contour features and vehicle local structural features in the vehicle correction feature set are structurally fused with the vehicle contour completion feature and vehicle component completion feature. Based on the spatial layout relationship between the fused vehicle contour information, window area, headlight area, wheel area, license plate area, and local body component areas, a vehicle structure combination feature is constructed. The vehicle length, vehicle width, vehicle height, wheelbase, and track width corresponding to the vehicle structure combination feature are calculated and matched with vehicle size constraint data. The target vehicle model is determined based on the matching results, and a vehicle model recognition result is generated. The vehicle body color is fused based on vehicle color distribution features, vehicle color consistency features, and vehicle side inference features to generate a vehicle body color recognition result. Specifically, the main color category, color region position, and color proportion in the vehicle color distribution features are matched with the continuous color distribution information in the vehicle color consistency features. The color distribution corresponding to the occluded and invisible areas of the vehicle is completed based on the vehicle side inference features. The color proportion of the color distribution corresponding to the visible, occluded, and invisible areas of the vehicle is statistically analyzed. The target vehicle body color is determined according to the color proportion ranking result, and a vehicle body color recognition result is generated. Based on the license plate recognition results, vehicle model recognition results, vehicle body color recognition results, and occupancy statistics, the results fields are associated, recognition confidence is summarized, and completeness is evaluated to generate vehicle fusion feature recognition results. Specifically, the license plate recognition results, vehicle model recognition results, and vehicle body color recognition results are associated and combined according to the same vehicle target identifier to generate vehicle recognition result data; the number of spatial units corresponding to the visible, invisible, and occluded occupancy parts of the vehicle is counted, and the proportion of each spatial unit to the total number of spatial units in the vehicle is calculated to obtain the proportion of visible, invisible, and occluded occupancy parts of the vehicle; based on the characters corresponding to the license plate recognition results... The confidence scores for symbol recognition, structure matching, and color recognition are weighted and summed to generate a comprehensive recognition confidence score. The percentage of visible vehicle parts is used as a completeness gain, while the percentages of invisible and occluded vehicle parts are used as completeness deductions. A completeness score is calculated by subtracting the sum of the percentages of invisible and occluded vehicle parts from the percentage of visible vehicle parts. A vehicle feature completeness evaluation result is then generated based on this score. Finally, the vehicle recognition result data, comprehensive recognition confidence score, and vehicle feature completeness evaluation result are correlated and organized to generate a vehicle fusion feature recognition result.

[0026] A vehicle feature recognition system based on spatial computing, comprising: The data processing module is used to acquire and standardize multi-source spatial perception data and vehicle image data; The occupancy analysis module is used to analyze the spatial occupancy relationship of the standardized spatial perception dataset and generate scene spatial occupancy relationship data. The visible recognition module is used to detect and segment vehicle targets, and generate a set of visible vehicle features; The occupancy marking module is used to mark the occupancy status of vehicle targets and generate a set of vehicle space occupancy features; The inference completion module is used to perform spatial occupancy inference processing on the vehicle spatial occupancy feature set and vehicle structural constraint data through the improved PoinTr network, and generate a vehicle invisible inference feature set. The fusion recognition module is used to perform part matching, confidence correction and fusion recognition processing on the vehicle's visible feature set and the vehicle's invisible inference feature set to generate vehicle fusion feature recognition results.

[0027] Example 1: To verify the feasibility of this invention in practice, it was applied to a vehicle management scenario in an underground parking lot of a large commercial complex. This underground parking lot has multiple entrances / exits, traffic lanes, and parking areas. Multiple surveillance cameras are deployed inside the parking lot to collect vehicle image data. Simultaneously, laser scanning equipment and spatial perception equipment are configured to acquire scene 3D point cloud data, spatial depth data, camera pose data, and vehicle traffic area data. During peak traffic hours, due to factors such as vehicle queuing, passing, and pillars around parking spaces, parts of vehicles are frequently obscured by other vehicles or fixed structures in the surveillance footage. This results in incomplete display of license plate areas, vehicle side areas, and partial vehicle structural areas, easily leading to incomplete vehicle feature recognition.

[0028] When applying this invention, firstly, multi-source spatial perception data and vehicle image data from a parking lot are acquired, and standardized spatial perception datasets and standardized vehicle image datasets are generated. Then, based on scene 3D point cloud data, spatial depth data, and camera pose data, spatial occupancy relationships between candidate vehicle spatial regions, occupant spatial regions, and camera visible regions are constructed, generating scene spatial occupancy relationship data. Simultaneously, target detection and segmentation processing is performed on vehicle images, extracting vehicle visible area features, vehicle edge contour features, vehicle color distribution features, and vehicle local structural features, generating a vehicle visible feature set. Combining the scene spatial occupancy relationship data, the visible, occupied, and invisible occupant parts of the vehicle target are marked, generating a vehicle spatial occupancy feature set. Further, an improved spatial occupancy inference mechanism is used to analyze the vehicle spatial occupancy feature set, inferring vehicle contour information, vehicle component information, license plate area information, and color distribution information corresponding to missing vehicle regions based on vehicle structural constraints and spatial occupancy relationships, generating a vehicle invisible inference feature set. Finally, the visible feature set and the invisible inference feature set of the vehicle are fused and recognized to obtain the license plate recognition result, vehicle model recognition result, vehicle body color recognition result, and vehicle feature completeness evaluation result.

[0029] In actual operation, parking lot vehicle management data during a National Day holiday was used as the verification object. The invention was continuously run in parking lot entrance and exit areas and densely populated parking areas. Statistical results show that, even with vehicle obstruction, partial missing data, and limited camera viewing angles, the invention can utilize spatial occupancy relationships and vehicle structural constraints to recover key feature information corresponding to invisible vehicle areas. This improves the consistency of vehicle feature integrity evaluation results, enhances the stability of license plate recognition, vehicle model recognition, and vehicle body color recognition results, and reduces recognition omissions and deviations caused by vehicle obstruction. This verifies the feasibility and practical value of the invention in complex parking lot scenarios.

[0030] Table 1. Performance Comparison of the Invention and Traditional Vehicle Feature Recognition Methods

[0031] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0032] In terms of license plate recognition accuracy, the traditional method achieves 91.7%, while the method of this invention reaches 96.3%, an improvement of 4.6 percentage points. The main reason for this result is that this invention constructs the spatial occupancy relationship between the vehicle candidate space region, the occlusion space region, and the camera's visible area, and combines this with vehicle invisible inference features to complete the missing license plate region. This allows license plate information in partially occluded states to still participate in the subsequent recognition process, thus achieving a higher license plate recognition accuracy.

[0033] In terms of vehicle model recognition accuracy and vehicle body color recognition accuracy, traditional methods achieve 88.9% and 90.5% respectively, while the method of this invention reaches 94.1% and 95.0% respectively. Simultaneously, the vehicle feature completeness is improved from 82.6% to 93.4%. This improvement is due to the fact that this invention not only utilizes visible vehicle area features for recognition but also recovers vehicle structural and color information corresponding to invisible areas through vehicle contour completion features, vehicle component completion features, and vehicle side inference features. This enhances the completeness of the overall vehicle features, enabling vehicle model and color recognition to be based on more comprehensive vehicle features.

[0034] In complex occlusion scenarios, traditional methods achieve an accuracy rate of 84.3% for occlusion scene recognition and 80.1% for feature recovery; the method of this invention achieves 92.8% and 91.2%, respectively. Specifically, the accuracy rate for occlusion scene recognition is improved by 8.5 percentage points, and the accuracy rate for feature recovery is improved by 11.1 percentage points. The main reason for this improvement lies in the introduction of an occupancy state analysis mechanism for visible, occluded, and invisible vehicle occupancy areas. This mechanism, combined with spatial occupancy reasoning based on the relationship between occupancy source and depth, accurately identifies the occluded areas of the vehicle and the locations of corresponding missing features, thus enhancing feature recovery capabilities under complex occupancy conditions.

[0035] In terms of vehicle tracking consistency index, the traditional method achieves 89.4%, while the method of this invention reaches 94.6%, an improvement of 5.2 percentage points. Simultaneously, the average processing time per vehicle increases from 118ms to 126ms, an increase of only 8ms. This demonstrates that the present invention effectively improves vehicle feature recovery and vehicle recognition capabilities with only a slight increase in computational overhead. This result is attributed to the present invention's use of spatial computation results to fuse and correct vehicle features, reducing the loss of vehicle identity information and recognition fluctuations caused by occlusion, thus enabling vehicles to maintain more stable feature representation and recognition results during continuous monitoring.

[0036] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A vehicle feature recognition method based on spatial computing, characterized in that, Includes the following steps: Acquire multi-source spatial perception data and vehicle image data in the target vehicle recognition scenario, and perform standardization processing to generate standardized spatial perception dataset and standardized vehicle image dataset; The standardized spatial perception dataset is parsed to determine the candidate spatial regions of vehicles, the spatial regions of occupants, the visible areas of cameras, and the deep correlation between each spatial region, thereby generating scene spatial occupancy data. Vehicle targets are detected and segmented based on a standardized vehicle image dataset, generating a set of visible vehicle features; Based on the scene space occupancy relationship data and the vehicle visible feature set, the occupancy status of the vehicle target is marked, and a vehicle space occupancy feature set is generated. The improved PoinTr network is used to perform spatial occupancy reasoning on the vehicle spatial occupancy feature set and vehicle structural constraint data to generate a vehicle invisible reasoning feature set. The visible feature set and the invisible inference feature set of the vehicle are subjected to part matching, confidence correction and fusion recognition processing to generate vehicle fusion feature recognition results.

2. The vehicle feature recognition method based on spatial computing according to claim 1, characterized in that, The multi-source spatial perception data includes scene 3D point cloud data, camera pose data, spatial depth data, scene structure data, spatial location data of occluders, and vehicle passage area data. The vehicle image data includes vehicle video frames, vehicle appearance time, vehicle image acquisition viewpoint, vehicle local image area, and vehicle image quality parameters. The standardization processing includes timestamp unification, spatial coordinate alignment, sensor pose correction, data format unification, abnormal data removal, missing data completion, and data scale normalization processing.

3. The vehicle feature recognition method based on spatial computing according to claim 1, characterized in that, The generation of the scene space occupancy relationship data specifically includes: Based on the scene 3D point cloud data and scene structure data in the standardized spatial perception dataset, the target vehicle recognition scene is divided into spatial grids to generate a set of scene spatial units. Based on the spatial location data of the occlusion objects and the scene structure data, fixed occlusion areas and dynamic occlusion areas are identified from the set of scene spatial units to determine the spatial area of ​​the occlusion objects; Based on spatial depth data and vehicle traffic area data, spatial units with vehicle traffic conditions are selected from the set of scene spatial units to determine candidate vehicle spatial areas. Based on camera pose data, spatial depth data, and scene structure data, the shooting coverage of each camera in the target vehicle recognition scene is projected and analyzed to determine the camera's visible area. Spatial depth analysis is performed based on the candidate space area of ​​the vehicle, the space area of ​​the obstruction, and the visible area of ​​the camera to determine the depth correlation between each space area. The candidate space area of ​​the vehicle, the space area of ​​the obstruction, and the visible area of ​​the camera are marked and their coordinates are bound. Spatial association matching is performed in combination with deep correlation to generate scene space occupancy relationship data.

4. The vehicle feature recognition method based on spatial computing according to claim 1, characterized in that, The generation of the vehicle visible feature set specifically includes: Based on the vehicle video frames, vehicle appearance time and vehicle image acquisition angle in the standardized vehicle image dataset, the vehicle video frames under the same acquisition angle are time-series organized to generate a vehicle image frame sequence. Vehicle target detection processing is performed on the vehicle image frame sequence to extract the image region where the vehicle target is located and generate the vehicle target detection region. Foreground segmentation is performed on the vehicle target detection area to separate the vehicle target area from the non-vehicle background area and generate vehicle visible area features. Boundary extraction and contour fitting are performed on the vehicle target area to generate vehicle edge contour features; Perform color distribution statistics and color region segmentation on the body pixels in the vehicle target area to generate vehicle color distribution features; The local structural regions in the target area of ​​the vehicle are identified, and the window area, headlight area, wheel area, license plate area and local component areas of the vehicle body are extracted to generate local structural features of the vehicle. The visible area features of the vehicle, the edge contour features of the vehicle, the color distribution features of the vehicle, and the local structural features of the vehicle are correlated and organized to generate a set of visible features of the vehicle.

5. The vehicle feature recognition method based on spatial computing according to claim 1, characterized in that, The generation of the vehicle space occupancy feature set specifically includes: Based on the vehicle spatial location relationship, visible area relationship and spatial depth relationship in the scene spatial occupancy relationship data, the set of visible vehicle features is mapped to the vehicle candidate spatial region to generate a spatial mapping relationship. Based on the vehicle's visible area features, vehicle edge contour features, and spatial mapping relationships, the visible spatial range of the vehicle target in the vehicle candidate spatial region is marked to generate the vehicle's visible occupied part. Based on the spatial position relationship, spatial depth relationship and occlusion boundary position in the vehicle edge contour features, the vehicle parts covered by the occluded object spatial region in the vehicle candidate spatial region are marked to generate the vehicle occluded part. Based on the visible area relationship, spatial adjacency relationship and local structural features of the vehicle, invisible vehicle parts that do not fall into the camera's visible area and are not covered by the vehicle's visible occupied parts in the candidate vehicle spatial area are marked as invisible, thus generating invisible vehicle occupied parts. Based on the spatial location and depth relationships of the obstruction space, the source of the vehicle's obstructed area is matched with the corresponding obstruction space area to generate the obstruction source and obstruction depth relationship; Based on the relationship between the visible and invisible parts of the vehicle, the occupied parts of the vehicle, the occupied parts of the vehicle, the occupancy source and the occupancy depth, the vehicle parts are matched to the visible feature set to generate the spatial correspondence of vehicle parts. The visible and invisible occupied parts of the vehicle, the occupied parts of the vehicle, the occupied parts of the vehicle, the source of occupancy, the occupancy depth relationship, and the spatial correspondence of vehicle parts are correlated and organized to generate a set of vehicle spatial occupancy features.

6. The vehicle feature recognition method based on spatial computing according to claim 1, characterized in that, The generation of the vehicle invisible reasoning feature set specifically includes: The vehicle spatial occupancy feature set and vehicle structural constraint data are input into the improved PoinTr network, and point cloud construction and grouping are performed through the vehicle local point cloud grouping unit to generate vehicle local point cloud groups. The improvement of the improved PoinTr network compared with the original PoinTr network is that a vehicle spatial location embedding unit is added between the part point proxy encoding unit and the missing part Transformer decoding unit, and an occlusion encoding unit is added between the vehicle spatial location embedding unit and the missing part Transformer decoding unit. The vehicle local point cloud group is processed by point proxy coding unit, and the visible occupied parts, invisible occupied parts, occupied parts and the part status information corresponding to the occupancy source are integrated to generate vehicle part point proxy data. Based on vehicle contour constraint data, license plate area constraint data, relative position constraint data of vehicle parts and vehicle size constraint data, the vehicle spatial position embedding unit performs spatial position embedding processing on the proxy data of vehicle parts points to generate vehicle spatial position embedding features. The occlusion coding unit encodes the vehicle spatial location embedding features, occlusion source, and occlusion depth relationship to generate occlusion coding features. The missing part decoding features are processed by the missing part Transformer decoding unit to generate missing part decoding features corresponding to the invisible occupied part of the vehicle and the occluded part of the vehicle. The missing parts are filled in by using vehicle occupancy completion units to generate a vehicle occupancy point cloud. The invisible feature output unit performs part matching and feature transformation on the vehicle completion occupied point cloud and the vehicle visible feature set to generate the vehicle invisible inference feature set.

7. The vehicle feature recognition method based on spatial computing according to claim 1, characterized in that, The generation of the vehicle fusion feature recognition result specifically includes: Based on the spatial correspondence of vehicle parts, perform part matching on the visible feature set and the invisible inference feature set of the vehicle to generate a vehicle matching feature set. Based on the confidence level of the visible area and the confidence level of the inference features, the vehicle matching feature set is corrected to generate a vehicle corrected feature set; Based on the vehicle correction feature set, license plate missing inference features and license plate region constraint data, license plate region fusion recognition is performed to generate license plate recognition results; Based on the vehicle correction feature set, vehicle contour completion feature and vehicle component completion feature, vehicle model structure matching is performed to generate vehicle model recognition results; Based on vehicle color distribution characteristics, vehicle color consistency characteristics, and vehicle side inference characteristics, vehicle body color fusion is performed to generate vehicle body color recognition results. Based on the license plate recognition results, vehicle model recognition results, vehicle body color recognition results, and occupancy statistics, the result fields are correlated, recognition confidence is summarized, and completeness is evaluated to generate vehicle fusion feature recognition results.

8. A vehicle feature recognition system based on spatial computing, comprising executing the vehicle feature recognition method based on spatial computing as described in any one of claims 1 to 7, characterized in that, include: The data processing module is used to acquire and standardize multi-source spatial perception data and vehicle image data; The occupancy analysis module is used to analyze the spatial occupancy relationship of the standardized spatial perception dataset and generate scene spatial occupancy relationship data. The visible recognition module is used to detect and segment vehicle targets, and generate a set of visible vehicle features; The occupancy marking module is used to mark the occupancy status of vehicle targets and generate a set of vehicle space occupancy features; The inference completion module is used to perform spatial occupancy inference processing on the vehicle spatial occupancy feature set and vehicle structural constraint data through the improved PoinTr network, and generate a vehicle invisible inference feature set. The fusion recognition module is used to perform part matching, confidence correction and fusion recognition processing on the vehicle's visible feature set and the vehicle's invisible inference feature set to generate vehicle fusion feature recognition results.

Images