Road traffic state detection method and system based on deep learning

Abstract

The invention discloses a road traffic state detection method and system based on deep learning. The method comprises the following steps: S1, acquiring continuous traffic perception data covering a target road area; s2, extracting local traffic state features; s3, constructing a traffic state evolution sequence of the road segments, and generating a time constraint relation and traffic state propagation prior information; s4, obtaining a connection relation between the road segments, and generating a road topological relation; s5, inputting the input tensor into an improved VMama model, introducing a dual propagation guidance mechanism to execute traffic state propagation modeling, and outputting spatial consistency features; s6, generating a consistency evaluation result; and S7, performing correction processing, and outputting a road traffic state detection result of the corresponding road segment. According to the method, combined constraint of the road traffic state on time sequence consistency and space consistency is realized, and the accuracy and robustness of a traffic state detection result in a complex road scene are improved.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and computer vision technology, and in particular to a road traffic condition detection method and system based on deep learning. Background Technology

[0002] With the continuous expansion of urban road traffic and the gradual improvement of intelligent transportation infrastructure, the demand for real-time perception and refined management of road traffic conditions is increasing. Therefore, technologies for the collection, fusion, and analysis of multi-source traffic perception data have received widespread attention. Existing road traffic condition detection methods mainly rely on video surveillance, fixed detectors, or floating car data to determine traffic flow, vehicle speed, and congestion status. However, these methods generally suffer from the following problems in practical applications: The collected traffic perception data comes from diverse sources, including image and video data. It is significantly affected by weather changes, lighting conditions, viewpoint obstruction, and equipment stability. The time synchronization error and spatial positioning deviation between different sensing devices are large, resulting in inconsistent traffic state representations of the same road segment at different times or from different viewpoints. Existing time and space alignment methods are difficult to maintain data continuity and consistency in complex road network environments over the long term. At the same time, road traffic state has significant temporal evolution and spatial propagation characteristics. Traditional analysis methods based on local statistics or short-term features are difficult to characterize the correlation between road segments and cannot effectively reflect the propagation process of traffic state between adjacent road segments. This leads to insufficient ability to identify sudden congestion, state fluctuations, and cascading effects, affecting the stability and accuracy of road traffic state detection results.

[0003] Therefore, how to provide a road traffic condition detection method and system based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a road traffic state detection method and system based on deep learning. This invention fully utilizes continuous traffic perception data processing, road segment modeling, temporal constraint relationship construction, and traffic state propagation modeling techniques to construct a traffic perception time series with road segments as the basic unit. It extracts local traffic state features from each road segment and generates temporal constraint relationships and prior information on traffic state propagation based on the traffic state evolution sequence. Simultaneously, it incorporates road topology relationships and introduces a dual propagation guidance mechanism through an improved VMamba model to model the propagation of traffic states. It integrates temporal constraint relationships and spatial consistency features to generate consistency evaluation results, and corrects traffic state features when consistency conditions are not met, outputting the road traffic state detection results corresponding to each road segment. This invention possesses advantages such as strong temporal consistency of traffic states, high spatial propagation characterization capability, and good stability of detection results, making it suitable for road traffic state detection and analysis in complex road network environments.

[0005] A road traffic state detection method based on deep learning according to an embodiment of the present invention includes the following steps: S1. Acquire continuous traffic perception data covering the target road area, perform time and space alignment processing, and generate traffic perception time series in road segments. S2. Extract local traffic state features of the corresponding road segments based on the traffic perception time series; S3. Construct a traffic state evolution sequence for road segments based on the local traffic state features, generate time constraint relationships, and generate prior information for traffic state propagation based on the time constraint relationships; S4. Obtain the connection relationships between road segments and generate road topology relationships; S5. Encode the local traffic state features, traffic state propagation prior information and road topology relationship to generate an input tensor. Input the input tensor into the improved VMamba model, introduce a dual propagation guidance mechanism to perform traffic state propagation modeling, and output spatial consistency features. S6. Integrate the time constraint relationship with the spatial consistency feature to generate a consistency evaluation result; S7. If the consistency assessment result does not meet the preset consistency conditions, then the local traffic state features are corrected to generate traffic state features and the road traffic state detection results of the corresponding road segments are output.

[0006] Optionally, S1 specifically includes: Collect continuous traffic sensing data covering the target road area, including image data or video data; The continuous traffic sensing data is processed for time synchronization to generate a traffic sensing data sequence with a unified timestamp. Spatial registration processing is performed on the traffic sensing data sequence to establish the correspondence between traffic sensing data coordinates and road coordinates, and to generate a spatially registered traffic sensing data sequence. Obtain road segment information, and perform segment mapping processing on the spatially registered traffic perception data sequence according to the road segment information to generate a traffic perception time series based on road segments.

[0007] Optionally, S2 specifically includes: Based on traffic perception time series, traffic perception data segments corresponding to road segments are extracted from the traffic perception time series. Perform data preprocessing operations on the traffic perception data segments, including size normalization, illumination normalization and noise suppression, to generate preprocessed traffic perception data segments. The preprocessed traffic perception data segments are input into a deep learning feature extraction network, which outputs feature maps corresponding to road segments. The feature map is subjected to feature aggregation processing to generate local traffic state features corresponding to road segments.

[0008] Optionally, S3 specifically includes: The local traffic state features corresponding to road segments are obtained, arranged in a unified time order, and a traffic state evolution sequence of road segments is constructed. The traffic state evolution sequence is segmented into time windows to generate multiple time window sequences. Perform state difference calculation on adjacent time window sequences to generate traffic state change sequences; Based on the traffic state change sequence, perform time-series constraint determination and processing to generate time constraint relationships; The state propagation correlation strength between road segments is calculated based on time constraints to generate prior information on traffic state propagation.

[0009] Optionally, S4 specifically includes: Obtain road segment information for the target road area, wherein the road segment information includes road segment identifiers, coordinates of the starting point of the road segment, and coordinates of the ending point of the road segment; The spatial adjacency relationship between road segments is calculated based on the coordinates of the starting point and ending point of each road segment, thus identifying pairs of road segments that are connected. For road segments with connection relationships, establish directed or undirected connection records to generate a set of road segment connection relationships; The road topology is constructed based on the road segment identifiers and the set of road segment connection relationships.

[0010] Optionally, the improved VMamba model specifically refers to: The improved VMamba model includes an input embedding structure, a sequence modeling structure, a propagation guidance structure, and a feature output structure. The propagation guidance structure introduces a dual propagation guidance mechanism during the state propagation modeling process, which includes a propagation prior bias mechanism and a calibration constraint mechanism. The input embedding structure performs dimension mapping and serialization encoding on the input tensor to generate a state sequence representation; The sequence modeling structure performs state dependency modeling processing on the state sequence representation to generate intermediate state representations; The propagation guidance structure includes: Based on the intermediate state representation output by the sequence modeling structure, the state propagation association set between road segments is determined by combining the road topology relationship; The initial state propagation association weights between road segments are calculated based on the intermediate state representation and the state propagation association set, and an initial propagation weight matrix is ​​generated. In the propagation prior bias mechanism, a propagation prior bias matrix is ​​constructed based on the propagation prior information of traffic state. The propagation prior bias matrix and the initial propagation weight matrix are consistent in structural dimension. The propagation prior bias matrix and the initial propagation weight matrix are fused to generate the bias propagation weight matrix. In the calibration constraint mechanism, consistency verification is performed on the state propagation correlation in the bias propagation weight matrix based on the time constraint relationship to generate a calibration reference relationship; constraint adjustment processing is performed on the state propagation correlation strength in the bias propagation weight matrix according to the calibration reference relationship to generate a calibration propagation weight matrix. The intermediate state representation is propagated and aggregated based on the calibration propagation weight matrix to generate the intermediate state representation after propagation guidance. The feature output structure performs feature aggregation processing based on the intermediate state representation after propagation guidance, and outputs spatially consistent features.

[0011] Optionally, S5 specifically includes: Obtain local traffic state features, road topology relationships, and prior information on traffic state propagation corresponding to road segments; The local traffic state features are serialized and arranged according to road segmentation and time sequence to generate a feature sequence. Perform structural encoding processing on the road topology relationships to generate a topology encoding sequence; Prior coding processing is performed on the prior information of traffic state propagation to generate a prior coding sequence; Perform dimension alignment processing on the feature sequence, topological coding sequence, and prior coding sequence to generate a coding-aligned sequence; Tensor quantization is performed on the encoded aligned sequence to generate the input tensor; The input tensor is fed into the improved VMamba model, and a dual propagation guidance mechanism is introduced to perform state propagation modeling on the input tensor, outputting space consistency features.

[0012] Optionally, S6 specifically includes: Obtain the temporal constraint relationship and spatial consistency features, match the temporal constraint relationship and spatial consistency features according to the road segment identifier, and generate road segment consistency input pairs; Perform fusion processing on the road segment consistency input pairs to generate road segment consistency fusion features; Calculate road segment consistency scores based on road segment consistency fusion features, and generate consistency score sequences. Consistency assessment results are generated based on the consistency score sequence and preset consistency conditions.

[0013] Optionally, S7 specifically includes: Based on the comparison between the consistency assessment results and the preset consistency conditions, a set of road segment identifiers with a consistency score less than the consistency threshold is determined; the preset consistency conditions include the consistency threshold, which is determined by the statistical results of historical road segment consistency scores. Based on the road segment identifier set, extract the local traffic state features and historical local traffic state features of the corresponding road segments, and construct the historical state sequence of road segments; The correction amount of local traffic state characteristics is calculated based on the historical state sequence of road segments, and a correction vector is generated. The modified vector is fused with local traffic state features to generate traffic state features; The road traffic condition detection results are output based on traffic condition features, corresponding to the road segments.

[0014] Optionally, a deep learning-based road traffic condition detection system includes the following modules: The perception processing module is used to acquire continuous traffic perception data covering the target road area, perform time and space alignment processing, and generate traffic perception time series in terms of road segments. The state extraction module is used to extract local traffic state features of the corresponding road segments based on the traffic perception time series, and construct the traffic state evolution sequence of the road segments. The prior generation module is used to generate time constraint relationships based on the traffic state evolution sequence, and to generate prior information for traffic state propagation based on the time constraint relationships; The propagation modeling module is used to obtain the connection relationship between road segments and generate road topology relationship; it performs encoding processing on the local traffic state features, traffic state propagation prior information and road topology relationship to generate input tensor, inputs the input tensor into the improved VMamba model, introduces a dual propagation guidance mechanism to perform traffic state propagation modeling, and outputs spatial consistency features; The evaluation output module is used to integrate the time constraint relationship and the spatial consistency features to generate a consistency evaluation result; if the consistency evaluation result does not meet the preset consistency conditions, the local traffic state features are corrected to generate traffic state features, and the road traffic state detection results of the corresponding road segments are output.

[0015] The beneficial effects of this invention are: By constructing a traffic perception time series based on road segments and extracting the local traffic state features corresponding to the road segments, and combining the traffic state evolution sequence to generate time constraint relationships, the ability to consistently represent road traffic states in the time dimension is effectively enhanced. By introducing road topology and prior information on traffic state propagation into the traffic state propagation modeling process, a unified input tensor is constructed using an improved VMamba model and state propagation modeling is performed. A dual propagation guidance mechanism is introduced to guide and constrain the state propagation association between road segments, thereby enhancing the ability to characterize the traffic state association between adjacent road segments in the spatial dimension and improving the accuracy and consistency of traffic state propagation modeling in complex road network environments. By integrating temporal constraints and spatial consistency features to generate consistency assessment results, and performing correction processing on local traffic state features when consistency conditions are not met, the stability and reliability of road traffic state detection results are effectively improved. Attached Figure Description

[0016] 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 Here is a flowchart of a deep learning-based road traffic state detection method proposed in this invention; Figure 2 This is a schematic diagram of the structure of the improved VMamba model proposed in this invention; Figure 3 This is a schematic diagram of a road traffic condition detection system based on deep learning proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figure 1 and Figure 2 A deep learning-based road traffic state detection method includes the following steps: S1. Acquire continuous traffic perception data covering the target road area, perform time and space alignment processing, and generate traffic perception time series in road segments. S2. Extract local traffic state features of the corresponding road segments based on the traffic perception time series; S3. Construct a traffic state evolution sequence for road segments based on the local traffic state features, generate time constraint relationships, and generate prior information for traffic state propagation based on the time constraint relationships; S4. Obtain the connection relationships between road segments and generate road topology relationships; S5. Encode the local traffic state features, traffic state propagation prior information and road topology relationship to generate an input tensor. Input the input tensor into the improved VMamba model, introduce a dual propagation guidance mechanism to perform traffic state propagation modeling, and output spatial consistency features. S6. Integrate the time constraint relationship with the spatial consistency feature to generate a consistency evaluation result; S7. If the consistency assessment result does not meet the preset consistency conditions, then the local traffic state features are corrected to generate traffic state features and the road traffic state detection results of the corresponding road segments are output.

[0019] This implementation acquires continuous traffic perception data covering the target road area and performs time and space alignment processing to form a traffic perception time series based on road segments. This achieves a unified representation of traffic information in both time and space dimensions, reducing spatiotemporal deviations between multi-source perception data. Furthermore, based on the traffic perception time series, local traffic state features corresponding to road segments are extracted, transforming the road traffic state from raw perception data into a structured state representation, improving the stability and comparability of traffic state descriptions. Simultaneously, based on the local traffic state features, a traffic state evolution sequence for road segments is constructed, generating time constraints and prior information for traffic state propagation. This provides clear time evolution constraints and propagation direction guidance for the traffic state modeling process, reducing state distortion caused by disordered propagation. Furthermore, by acquiring the connection relationships between road segments and constructing road topology relationships, the road... Spatial structure information is incorporated into the traffic state modeling process, ensuring that traffic state propagation conforms to the characteristics of real road structures. Simultaneously, local traffic state features, prior information on traffic state propagation, and road topology are uniformly encoded and input into an improved VMamba model. A dual propagation guidance mechanism is introduced to constrain and calibrate the traffic state propagation process, enhancing the consistency of traffic state representation in the spatial dimension. Furthermore, consistency assessment results are generated by fusing temporal constraints and spatial consistency features, enabling a quantitative determination of the spatiotemporal consistency of traffic states. When the consistency assessment results do not meet preset consistency conditions, local traffic state features are corrected, allowing them to adaptively adjust based on historical evolution patterns. This results in road traffic state detection results that better reflect real road conditions, improving the overall accuracy and stability of road traffic state detection.

[0020] In this embodiment, S1 specifically refers to: Collect continuous traffic sensing data covering the target road area. The continuous traffic sensing data is acquired by fixed road monitoring equipment or roadside sensing equipment. The continuous traffic sensing data includes image data or video data collected at a fixed sampling frequency. The sampling frequency of the image data is set to 1 frame / second to 5 frames / second, and the frame rate of the video data is set to 15 frames / second to 30 frames / second. The continuous traffic sensing data contains time stamp information and spatial location information. The continuous traffic sensing data is processed for time synchronization. The time synchronization process corrects the time stamp information of different collection times by using a unified time reference. The time stamp information is aligned with milliseconds as the smallest time unit. The time synchronization error is controlled within 50 milliseconds. A traffic sensing data sequence with a unified timestamp is generated. The unified timestamp represents the correspondence of continuous traffic sensing data on the same time axis. Spatial registration processing is performed on the traffic sensing data sequence. Spatial registration is completed by mapping the image pixel coordinates or video frame coordinates in the traffic sensing data to the road spatial coordinate system. The road spatial coordinate system consists of the road centerline coordinates, road boundary coordinates, and road direction vector. The spatial registration process corrects the coordinate mapping relationship according to the installation height, pitch angle, and horizontal angle of the traffic sensing equipment. The spatial registration error is controlled within 0.5 meters, and a spatially registered traffic sensing data sequence is generated. Obtain road segmentation information, which is determined by the basic road data. A road segment is defined as the road interval between two adjacent road nodes. Road nodes are determined by the location of intersections, ramp connection points, or road attribute change points. The length of the road segment is set to be within the range of 20 meters to 100 meters. The spatially registered traffic perception data sequence is segmented and mapped according to road segment information. The segment mapping process is completed by determining whether the spatial coordinates corresponding to the traffic perception data fall within the spatial range of a certain road segment. Traffic perception data that fall within the range of the same road segment at the same timestamp constitute the traffic perception data fragment of that road segment at the corresponding time point. After being arranged in chronological order, a traffic perception time sequence is generated with road segments as the unit.

[0021] In this embodiment, S2 specifically refers to: Based on the traffic perception time series, traffic perception data segments corresponding to road segments are extracted from the traffic perception time series. The traffic perception data segments consist of traffic perception data corresponding to the same road segment under consecutive timestamps. The time length of the traffic perception data segments is set to be within the range of 3 to 10 seconds. The traffic perception data segments are arranged in the order of timestamps to form a set of perception data of road segments within a local time range. Data preprocessing is performed on traffic perception data segments. Size normalization is achieved by uniformly scaling the image or video frames in the traffic perception data to a fixed resolution, which is set to a width of 640 pixels and a height of 360 pixels. Illumination normalization is achieved by linearly stretching the pixel brightness distribution of the traffic perception data so that the pixel grayscale value distribution is mapped to the range of 0 to 255. Noise suppression is achieved by performing pixel-level smoothing on the traffic perception data under adjacent timestamps, with the smoothing window size set to the range of 3 to 5 frames, to generate preprocessed traffic perception data segments. The preprocessed traffic perception data segments are input into a deep learning feature extraction network, which consists of a multi-layer convolutional structure and a non-linear mapping structure. The multi-layer convolutional structure is used to extract spatial texture information and target distribution information from the preprocessed traffic perception data segments layer by layer, and the non-linear mapping structure is used to perform non-linear transformation on the extracted features, outputting feature maps corresponding to road segments. The feature maps represent the traffic state feature distribution of road segments in a local time range in the form of a multi-channel two-dimensional matrix. Feature aggregation is performed on the feature maps. This aggregation is accomplished by statistically converging the feature maps in both spatial and temporal dimensions. Spatial aggregation is achieved by summing and normalizing the spatial location features of the same channel in the feature maps. Temporal aggregation is achieved by weighted averaging of the feature maps corresponding to different timestamps. The weighting coefficients are set to decrease linearly in chronological order, and the weight decay coefficients are set to the range of 0.8 to 0.95. This generates local traffic state features corresponding to the road segments. The local traffic state features represent the traffic state of the road segments within a local time range in the form of fixed-length feature vectors.

[0022] In this embodiment, S3 specifically refers to: Local traffic state features corresponding to road segments are obtained. These local traffic state features are generated by the aforementioned process and stored in the order of timestamps. Based on a unified time reference, the local traffic state features corresponding to the same road segment at consecutive time points are sorted to form a feature sequence arranged in ascending order of time. The traffic state evolution sequence of the road segment is constructed. The traffic state evolution sequence is used to characterize the trajectory of traffic state changes of the road segment within a continuous time range. The traffic state evolution sequence is segmented into time windows. The time window segmentation is completed by sliding the traffic state evolution sequence according to a fixed time length. The time window length is set to be within the range of 5 to 20 seconds, and the sliding step size between adjacent time windows is set to be within the range of 1 to 5 seconds. Each time window sequence contains a set of local traffic state features within the corresponding time range, generating multiple time window sequences arranged in chronological order. State difference calculation is performed on adjacent time window sequences. The state difference calculation is completed by performing a dimension-wise difference operation on the local traffic state features at corresponding positions in adjacent time window sequences. The difference operation generates a state difference vector by subtracting the feature value of the previous time window from the feature value of the later time window. All state difference vectors are arranged in chronological order to form a traffic state change sequence. The traffic state change sequence is used to describe the magnitude and direction of change of the traffic state of road segments between adjacent time windows. The temporal constraint determination process is based on the traffic state change sequence. The temporal constraint determination is completed by statistical analysis of the amplitude of each state difference vector in the traffic state change sequence. The amplitude is obtained by summing the absolute values ​​of the feature differences of each dimension in the state difference vector. Preferably, the state change amplitude threshold is set to 1.5 times the average amplitude of historical state changes. When the state change amplitude is lower than the state change amplitude threshold in multiple consecutive time windows, it is determined that the traffic state change of the road segment meets the temporal continuous stability constraint, and a temporal constraint relationship is generated. The temporal constraint relationship is represented in the form of a correspondence between time window index and stability identifier. The state propagation association strength between road segments is calculated based on time constraints. The state propagation association strength is completed by jointly analyzing the time constraints of different road segments within the same time window. For road segment pairs that simultaneously satisfy the time continuity and stability constraint within the same time window, the similarity between the corresponding traffic state change sequences is calculated. The similarity is determined by the cosine similarity between the corresponding state difference vectors. Preferably, the similarity value ranges from 0 to 1. The state propagation association strength is directly represented by the similarity value. The state propagation association strength of all road segment pairs constitutes the prior information of traffic state propagation. The prior information of traffic state propagation is stored in the form of a mapping between road segment pairs and corresponding association strengths.

[0023] In this embodiment, S4 specifically refers to: Obtain road segment information for the target road area. The road segment information is obtained from basic road data or electronic map data. The road segment is based on the road interval between two adjacent road nodes as a basic unit. The road nodes are determined by the center position of the intersection, the connection point of the ramp, or the point where the road attributes change. The road segment information includes the road segment identifier, the coordinates of the starting point of the road segment, and the coordinates of the ending point of the road segment. The road segment identifier is represented by a unique number. The coordinates of the starting point of the road segment and the coordinates of the ending point of the road segment are represented by a unified road space coordinate system. The spatial adjacency relationship between road segments is calculated based on the coordinates of the starting point and the ending point of each road segment. The spatial adjacency relationship is determined by judging the connection between different road segments in the spatial coordinate system. When the Euclidean distance between the ending point coordinates of one road segment and the starting point coordinates of another road segment is less than a preset adjacency distance threshold, it is determined that there is a spatial adjacency relationship between the corresponding road segments. Preferably, the adjacency distance threshold is set within the range of 1 meter to 3 meters. At the same time, the directional continuity between road segments is judged by combining the road direction vector. When the angle between the road directions is less than 30 degrees, it is determined that the directions are continuous. Based on spatial adjacency, road segment pairs with connection relationships are identified. For road segment pairs that are directionally continuous and spatially adjacent, a directed connection relationship is established. The direction of the directed connection is determined by the driving direction from the start point to the end point of the road segment. For road segment pairs that are spatially adjacent but directionally discontinuous, an undirected connection relationship is established. All road segment pairs with connection relationships form a set of road segment connection relationships. The set of road segment connection relationships is stored in the form of a correspondence between road segment identifier pairs and connection types. The road topology is constructed based on the set of road segment identifiers and road segment connection relationships. The road topology is represented in the form of a graph structure, with road segment identifiers as nodes and road segment connection relationships as edges. Edge attributes include connection type and direction attributes. The road topology is used to describe the spatial connection structure between road segments within the target road area.

[0024] In this embodiment, the improved VMamba model is specifically as follows: The improved VMamba model includes an input embedding structure, a sequence modeling structure, a propagation guidance structure, and a feature output structure. The propagation guidance structure introduces a dual propagation guidance mechanism in the traffic state propagation modeling process. The dual propagation guidance mechanism consists of a propagation prior bias mechanism and a calibration constraint mechanism, which are used to guide and constrain the state propagation relationship between road segments. The input embedding structure performs dimension mapping and serialization encoding on the input tensor. Dimension mapping maps the feature dimensions from different sources in the input tensor to a fixed dimension space through linear transformation. Preferably, the mapped feature dimensions are set to the range of 64 to 256. Serialization encoding is completed by arranging the mapped features according to the road segment identifier and time order to generate a state sequence representation. The state sequence representation represents the traffic state of the road segment in the form of a multi-dimensional feature sequence arranged in time order. The sequence modeling structure performs state dependency modeling on the state sequence representation. State dependency modeling is completed by modeling the feature correlation between adjacent time positions in the state sequence representation. The state dependency relationship is determined by the numerical change relationship of the features of the preceding and following states in the sequence, generating intermediate state representations. The intermediate state representations are used to characterize the evolution characteristics of road segment traffic states in the time dimension. The propagation guidance structure is based on the intermediate state representation output by the sequence modeling structure. It combines the road topology relationship to determine the state propagation association set between road segments. The state propagation association set consists of road segment pairs that have a connection relationship in the road topology relationship. Each road segment pair corresponds to a potential state propagation relationship. The initial state propagation association weights between road segments are calculated based on the intermediate state representations and the state propagation association set. The initial state propagation association weights are obtained by measuring the similarity of the intermediate state representations of the interrelated road segments. The similarity measure is obtained by normalizing the inner product result between the intermediate state representation vectors. All initial state propagation association weights constitute the initial propagation weight matrix. Each element in the initial propagation weight matrix represents the basic propagation strength between a pair of road segments. In the propagation prior bias mechanism, a propagation prior bias matrix is ​​constructed based on the traffic state propagation prior information. The propagation prior bias matrix is ​​directly mapped from the state propagation correlation strength in the aforementioned traffic state propagation prior information. The propagation prior bias matrix and the initial propagation weight matrix are consistent in the road segment index dimension and structural dimension. The propagation prior bias matrix and the initial propagation weight matrix are weighted and fused. The fusion method is to add the corresponding elements and then normalize them to generate the bias propagation weight matrix. The bias propagation weight matrix is ​​used to enhance the state correlation relationship that conforms to the historical propagation law during the propagation process. In the calibration constraint mechanism, consistency verification is performed on the state propagation associations in the bias propagation weight matrix based on the time constraint relationship. The consistency verification is completed by judging whether the road segments participating in the propagation simultaneously satisfy the time constraint relationship within the corresponding time window. When the road segment pairs are all marked as satisfying the time continuous stability constraint within the same time window, the corresponding state propagation association is determined to be a consistent association, and a calibration reference relationship is generated. The strength of the state propagation associations in the bias propagation weight matrix is ​​constrained and adjusted according to the calibration reference relationship. Preferably, the strength of the propagation associations that do not meet the consistency condition is multiplied by an attenuation coefficient, which is set to the range of 0.3 to 0.7, and the calibration propagation weight matrix is ​​generated. The intermediate state representation is propagated and aggregated based on the calibration propagation weight matrix. The propagation aggregation is completed by weighted summation of the intermediate state representations of the interrelated road segments according to the calibration propagation weight, generating the intermediate state representation after propagation guidance. The intermediate state representation after propagation guidance contains both temporal evolution information and spatial propagation information. The feature output structure performs feature aggregation processing based on the intermediate state representation after propagation guidance. Feature aggregation is completed by converging the intermediate state representation after propagation guidance in the time dimension. Preferably, a weighted average method is used to generate a fixed-length feature vector and output spatial consistency features. Spatial consistency features are used to characterize the comprehensive consistency of road segment traffic states in the time and space dimensions.

[0025] In this embodiment, the improved VMamba model is based on the VMamba model because road traffic state detection scenarios simultaneously possess significant temporal evolution and spatial propagation characteristics. The VMamba model focuses on modeling sequential state dependencies, but it fails to adequately characterize the state propagation constraints between road segments caused by topological structure and traffic evolution patterns, making it difficult to fully reflect the spatial consistency changes in traffic states between adjacent road segments. Therefore, while maintaining the sequential modeling advantages of the VMamba model, this embodiment introduces a propagation guidance structure at the model structure level, and sets up a dual propagation guidance mechanism within this structure. By introducing prior information on traffic state propagation into the construction process of state propagation association weights in the form of propagation prior biases, the state propagation modeling process is constrained by real traffic evolution patterns. This improvement method combines temporal constraints with consistency calibration of propagation correlation strength to suppress abnormal propagation paths that do not conform to the logic of temporal evolution, thereby achieving synergistic constraints between temporal constraints, road topology relationships, and state sequence modeling results. This improvement is not a simple adjustment of existing model parameters, but rather introduces a structural guidance and constraint mechanism in the state propagation modeling stage. This enables the model to have controllable propagation capabilities and consistency verification capabilities in both spatial and temporal dimensions, forming a dedicated modeling structure for road traffic state detection scenarios. It has the technical feature of explicitly embedding prior information on traffic state propagation and temporal constraints into the deep sequence modeling process, breaking through the modeling method of VMamba model that only relies on implicit state evolution. It forms a new technical effect in traffic state spatial consistency modeling and abnormal state suppression, and has clear novelty and creativity.

[0026] In this embodiment, S5 specifically refers to: The local traffic state features, road topology relationships, and prior information on traffic state propagation corresponding to road segments are obtained. The local traffic state features exist in the form of multi-dimensional numerical vectors associated with road segment identifiers and timestamps. The road topology relationships exist in the form of a set of connection relationships composed of road segment identifier pairs. The prior information on traffic state propagation exists in the form of a numerical set composed of the state propagation association strength between road segments. The local traffic state features are serialized and arranged according to the road segment identifier and time order. The serialization and arrangement are completed by splicing the local traffic state features of the same road segment under consecutive timestamps in ascending order of time to generate a feature sequence. Each time position in the feature sequence corresponds to the traffic state description of a road segment at that time point. Preferably, the dimension of a single feature vector is set to 64 to 256. The road topology relationship is processed by structural encoding. Structural encoding is completed by mapping the road segment connection relationship into a numerical form. Specifically, a connection vector with a length equal to the total number of road segments is constructed for each road segment. The position of the connection vector that has a connection relationship with the road segment is assigned a value of 1, and the position that has no connection relationship is assigned a value of 0. All connection vectors are arranged in the order of road segment identifiers to generate a topology encoding sequence. Prior coding is performed on the prior information of traffic state propagation. Prior coding is completed by mapping the state propagation association strength in the prior information of traffic state propagation into a numerical vector. Each element in the numerical vector corresponds to a pair of propagation association strengths between road segments. The propagation association strength is calculated by the aforementioned time constraint relationship. Preferably, the propagation association strength is normalized to the range of 0 to 1. All numerical vectors are arranged in the order of road segment identification to generate a prior coding sequence. The feature sequence, topological coding sequence and prior coding sequence are dimensionally aligned. Dimensional alignment is achieved by uniformly expanding or pruning the three types of sequences in the time dimension and feature dimension, so that the three types of sequences correspond to the same road segment at each time position and have the same feature length. Preferably, the sequences with insufficient dimensions are filled with zero values, and the sequences with excessive dimensions are truncated to generate the coded aligned sequence. Tensor quantization is performed on the coded alignment sequence. Tensor quantization is completed by stacking the coded alignment sequence according to the time dimension, road segment dimension and feature dimension to generate a three-dimensional numerical structure. In the three-dimensional numerical structure, the first dimension represents the time order, the second dimension represents the road segment identification order, and the third dimension represents the fused feature dimension to generate an input tensor. Preferably, the time length of the input tensor is set to 5 to 30. The input tensor is fed into the improved VMamba model. The improved VMamba model jointly models the temporal relationship in the input tensor and the propagation relationship between road segments based on the input embedding structure, sequence modeling structure and propagation guidance structure. The propagation guidance structure introduces a dual propagation guidance mechanism, which guides the direction of state propagation through the propagation prior bias mechanism and constrains the propagation relationship that does not meet the temporal consistency through the calibration constraint mechanism. Finally, the feature output structure outputs spatial consistency features, which represent the consistency results of the traffic state of road segments in the temporal and spatial dimensions in the form of numerical vectors.

[0027] In this embodiment, S6 specifically refers to: The temporal constraint relationship and spatial consistency feature are obtained. The temporal constraint relationship exists in the form of a set of temporal constraint values ​​associated with the road segment identifier, and the spatial consistency feature exists in the form of a multi-dimensional numerical vector associated with the road segment identifier. The temporal constraint relationship and spatial consistency feature are matched one-to-one according to the road segment identifier. Each set of data after matching includes the constraint description in the time dimension and the consistency feature description in the spatial dimension of the same road segment, generating road segment consistency input pairs. The road segment consistency input pairs are fused. The fusion process is completed by concatenating the time constraint values ​​in the time constraint relationship with the multi-dimensional numerical vectors in the spatial consistency features in the feature dimension. After concatenation, a joint feature representation containing time consistency information and spatial consistency information is formed, and the road segment consistency fusion feature is generated. Preferably, the time constraint values ​​are normalized to the range of 0 to 1 before fusion. The road segment consistency score is calculated based on the road segment consistency fusion feature. The consistency score is obtained by weighted summation of the values ​​of each dimension in the consistency fusion feature. The weights are used to characterize the relative influence of time constraint information and spatial consistency information in the consistency assessment. Preferably, the weights corresponding to time constraint information are set to 0.4 to 0.6, and the weights corresponding to spatial consistency information are set to 0.4 to 0.6. The weighted summation result is used as the consistency score of the corresponding road segment, generating a consistency score sequence. Each value in the consistency score sequence corresponds to the comprehensive consistency level of a road segment within the current time window. Consistency assessment results are generated based on the consistency score sequence and preset consistency conditions. The preset consistency conditions include a consistency threshold, which is used to distinguish whether the traffic status of road segments meets the temporal and spatial consistency requirements. Preferably, the consistency threshold is set to 0.6 to 0.8. When the consistency score is greater than or equal to the consistency threshold, the corresponding road segment is determined to meet the consistency conditions. When the consistency score is less than the consistency threshold, the corresponding road segment is determined not to meet the consistency conditions, thereby forming a unified consistency assessment result that includes road segment identification and consistency determination results.

[0028] In this embodiment, S7 specifically refers to: The consistency assessment results are compared with the preset consistency conditions, which include a consistency threshold. The consistency threshold is obtained by statistical analysis of the road segment consistency score sequence in the historical period. The statistical method includes calculating the mean and standard deviation of the historical consistency scores, and subtracting the standard deviation coefficient from the mean to obtain the consistency threshold. Preferably, the standard deviation coefficient is set to 0.5 to 1.0, and the consistency threshold is limited to the range of 0.6 to 0.8. Based on the judgment result that the consistency score is less than the consistency threshold, the set of road segment identifiers that do not meet the consistency conditions is determined. Based on the road segment identifier set, the local traffic state features of the corresponding road segment within the current time window are extracted, and the local traffic state features of the corresponding road segment within the historical time window are extracted simultaneously. The historical time windows are arranged in chronological order to form a historical local traffic state feature set. Preferably, the number of historical time windows is set to 3 to 10, and the interval between adjacent time windows is consistent with the sampling interval of the aforementioned traffic perception time series. Based on the current local traffic state features and the historical local traffic state feature set, a historical state sequence of road segments is constructed. The historical state sequence exists in the form of a multi-dimensional numerical vector arranged in chronological order. The correction amount of local traffic state features is calculated based on the historical state sequence of road segments. The correction amount is obtained by statistically analyzing the changing trends between adjacent state feature vectors in the historical state sequence. The changing trends include the direction of numerical increase or decrease and the magnitude of change. The weighted average change result of the historical state sequence is used as the correction benchmark. Preferably, the historical state features that are closer to the current time are assigned a larger weight. The weight range is set to 0.1 to 0.5. Based on the correction benchmark, a correction vector with the same dimension as the current local traffic state features is generated. The correction vector and local traffic state features are fused. The fusion method includes weighted superposition of the correction vector and local traffic state features in corresponding dimensions. The weighting ratio is used to control the correction intensity. Preferably, the weight of the correction vector is set to 0.2 to 0.4, and the weight of the local traffic state features is set to 0.6 to 0.8. The fusion result forms the traffic state features adjusted by consistency constraints. Based on traffic state characteristics, the output road traffic state detection results corresponding to road segments are provided. The road traffic state detection results include traffic state categories and state value descriptions that correspond one-to-one with the road segment identifiers. The traffic state categories are used to characterize the operational state level of the road segment, and the traffic state value descriptions are used to characterize the comprehensive traffic state level of the road segment within the current time window.

[0029] refer to Figure 3 In this embodiment, a road traffic condition detection system based on deep learning includes the following modules: The perception processing module is used to acquire continuous traffic perception data covering the target road area, perform time and space alignment processing, and generate traffic perception time series in terms of road segments. The state extraction module is used to extract local traffic state features of the corresponding road segments based on the traffic perception time series, and construct the traffic state evolution sequence of the road segments. The prior generation module is used to generate time constraint relationships based on the traffic state evolution sequence, and to generate prior information for traffic state propagation based on the time constraint relationships; The propagation modeling module is used to obtain the connection relationship between road segments and generate road topology relationship; it performs encoding processing on the local traffic state features, traffic state propagation prior information and road topology relationship to generate input tensor, inputs the input tensor into the improved VMamba model, introduces a dual propagation guidance mechanism to perform traffic state propagation modeling, and outputs spatial consistency features; The evaluation output module is used to integrate the time constraint relationship and the spatial consistency features to generate a consistency evaluation result; if the consistency evaluation result does not meet the preset consistency conditions, the local traffic state features are corrected to generate traffic state features, and the road traffic state detection results of the corresponding road segments are output.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a traffic operation monitoring platform. A ring expressway and three urban arterial roads were selected as the target road area. This area was divided into 64 road segments based on intersection spacing and ramp entrances / exits. Road segment identifiers and coordinates were provided by the existing GIS road network. Traditional platform solutions employ a traffic state recognition method based on single-frame features. Common problems include spatial registration deviations caused by timestamp drift and perspective changes in traffic perception data; a lack of consistency constraints between adjacent road segments in local traffic state features; and broken propagation chains when traffic congestion spreads from upstream to downstream, resulting in jumps in congestion segment locations, numerous false alarms, and delayed warnings, especially noticeable at night and in rainy weather.

[0031] In this scenario, the platform integrates video data from fixed video checkpoints and high-point panoramic cameras to form continuous traffic perception data. It then performs time synchronization and spatial registration to generate a traffic perception time series based on road segments. The traffic perception time series extracts traffic perception data fragments by road segment, performs size normalization, illumination normalization, and noise suppression, and inputs these fragments into a deep learning feature extraction network to obtain feature maps. Feature aggregation yields the local traffic state features corresponding to each road segment. These local traffic state features are arranged in a unified temporal order to construct a traffic state evolution sequence. Traffic state change sequences are obtained by differentiating the states of adjacent time windows. Temporal constraints are established based on these sequences, and the propagation correlation strength between road segments is calculated using these constraints to obtain prior information about traffic state propagation. The starting and ending coordinates of each road segment are used to calculate the spatial adjacency between road segments, establishing a set of road segment connection relationships and forming a road topology. Local traffic state features, prior information on traffic state propagation, and road topology relationships are serialized, structurally encoded, and aligned with prior codes and dimensions to generate an input tensor. This tensor is then fed into an improved VMamba model, where spatial consistency features are obtained through a dual propagation guidance mechanism. Temporal constraints and spatial consistency features generate a consistency score sequence, which is compared with preset consistency conditions to output a consistency evaluation result. Based on the historical state sequence of road segments, a correction vector is calculated and fused with local traffic state features to generate traffic state features, outputting the road traffic state detection result.

[0032] To ensure data comparability, this comparison used the same road segment, the same camera group, and the same operating time. The statistical caliber used a mixed sample of weekdays and weekends, with a total sample size of approximately 2160 hours of video stream. The comparison objects were the platform's traditional solution and the solution of this invention. Manual verification used a combination of sampling verification and event playback, and the verified events covered morning and evening rush hours, accident-related road closures, construction-related road closures, and scenarios with reduced visibility due to rainfall.

[0033] Table 1. Comparison of Road Traffic Condition Detection Results

[0034] Table 1 reflects the stable returns of the proposed solution during three months of continuous operation. The core improvements are concentrated in the accuracy of road traffic condition detection results, congestion false alarm rate, congestion missed alarm rate, and average warning lead time. The accuracy of the traditional solution fluctuates between 85.7% and 87.4%, with false alarm and missed alarm rates increasing significantly during peak hours and inclement weather. The average warning lead time is only 2.3 to 2.9 minutes, and the average positioning deviation reaches 1.3 to 1.6 road segments, which can easily cause positioning jumps in congested sections. This invention constructs an input tensor by incorporating time constraints and prior information about traffic conditions, and outputs spatial consistency features from an improved VMamba model. A consistency evaluation result is then formed from the consistency scoring sequence, making the congestion propagation chain more coherent under road topological constraints. The positioning error is reduced to 0.6 to 0.7 road segments. Simultaneously, the false alarm rate is reduced to 3.8% to 4.5%, the false alarm rate to 3.4% to 3.9%, and the average warning lead time is increased to 7.2 to 8.4 minutes. This demonstrates improved identification capabilities in the early stages of congestion formation, meeting the platform's requirements for early intervention and stable dissemination.

[0035] Table 2 Comparison of Consistency Assessment and Correction Effects for Typical Road Segments

[0036] Table 2 focuses on the consistency assessment results and the effects of the corrective processing, selecting road segments such as R12 and R25 that are easily affected by ramp merging and signal timing. Traditional schemes show that the mean consistency score for these road segments is often below 0.60 and the standard deviation is large, indicating frequent jumps in state output within adjacent time windows, making it difficult to reliably establish the consistency assessment. Under the same time window length and number of historical time windows, the scheme of this invention improves the mean consistency score to 0.71 to 0.76 and reduces the standard deviation to 0.07 to 0.09, indicating that the road segment consistency fusion characteristics after integrating spatial consistency features and temporal constraints are more stable. The number of times the consistency assessment results did not meet the preset consistency conditions decreased from 4.7 to 6.1 times / day to 1.2 to 1.9 times / day, the frequency of triggering correction vectors decreased and they were more focused on real abnormal segments; the consistency score improved by 0.10 to 0.15 after correction, and the number of false state judgments decreased by 1.5 to 2.1 times / day, which reflects the effectiveness of the fusion of correction vectors and local traffic state features to generate traffic state features. It can suppress unreasonable propagation paths and repair local states when congestion propagation is incomplete or there are perceptual fluctuations, thereby improving the credibility and continuity of road traffic state detection results.

[0037] The above description is only a preferred embodiment 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 road traffic state detection method based on deep learning, characterized in that, Includes the following steps: S1. Acquire continuous traffic perception data covering the target road area, perform time and space alignment processing, and generate traffic perception time series in road segments. S2. Extract local traffic state features of the corresponding road segments based on the traffic perception time series; S3. Construct a traffic state evolution sequence for road segments based on the local traffic state features, generate time constraint relationships, and generate prior information for traffic state propagation based on the time constraint relationships; S4. Obtain the connection relationships between road segments and generate road topology relationships; S5. Encode the local traffic state features, traffic state propagation prior information and road topology relationship to generate an input tensor. Input the input tensor into the improved VMamba model, introduce a dual propagation guidance mechanism to perform traffic state propagation modeling, and output spatial consistency features. S6. Integrate the time constraint relationship with the spatial consistency feature to generate a consistency evaluation result; S7. If the consistency assessment result does not meet the preset consistency conditions, then the local traffic state features are corrected to generate traffic state features and the road traffic state detection results of the corresponding road segments are output.

2. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S1 is: Collect continuous traffic sensing data covering the target road area, including image data or video data; The continuous traffic sensing data is processed for time synchronization to generate a traffic sensing data sequence with a unified timestamp. Spatial registration processing is performed on the traffic sensing data sequence to establish the correspondence between traffic sensing data coordinates and road coordinates, and to generate a spatially registered traffic sensing data sequence. Obtain road segment information, and perform segment mapping processing on the spatially registered traffic perception data sequence according to the road segment information to generate a traffic perception time series based on road segments.

3. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S2 is: Based on traffic perception time series, traffic perception data segments corresponding to road segments are extracted from the traffic perception time series. Perform data preprocessing operations on the traffic perception data segments, including size normalization, illumination normalization and noise suppression, to generate preprocessed traffic perception data segments. The preprocessed traffic perception data segments are input into a deep learning feature extraction network, which outputs feature maps corresponding to road segments. The feature map is subjected to feature aggregation processing to generate local traffic state features corresponding to road segments.

4. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S3 is: The local traffic state features corresponding to road segments are obtained, arranged in a unified time order, and a traffic state evolution sequence of road segments is constructed. The traffic state evolution sequence is segmented into time windows to generate multiple time window sequences. Perform state difference calculation on adjacent time window sequences to generate traffic state change sequences; Based on the traffic state change sequence, perform time-series constraint determination and processing to generate time constraint relationships; The state propagation correlation strength between road segments is calculated based on time constraints to generate prior information on traffic state propagation.

5. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S4 is: Obtain road segment information for the target road area, wherein the road segment information includes road segment identifiers, coordinates of the starting point of the road segment, and coordinates of the ending point of the road segment; The spatial adjacency relationship between road segments is calculated based on the coordinates of the starting point and ending point of each road segment, thus identifying pairs of road segments that are connected. For road segments with connection relationships, establish directed or undirected connection records to generate a set of road segment connection relationships; The road topology is constructed based on the road segment identifiers and the set of road segment connection relationships.

6. The road traffic state detection method based on deep learning according to claim 1, characterized in that, The improved VMamba model is specifically as follows: The improved VMamba model includes an input embedding structure, a sequence modeling structure, a propagation guidance structure, and a feature output structure. The propagation guidance structure introduces a dual propagation guidance mechanism during the state propagation modeling process, which includes a propagation prior bias mechanism and a calibration constraint mechanism. The input embedding structure performs dimension mapping and serialization encoding on the input tensor to generate a state sequence representation; The sequence modeling structure performs state dependency modeling processing on the state sequence representation to generate intermediate state representations; The propagation guidance structure includes: Based on the intermediate state representation output by the sequence modeling structure, the state propagation association set between road segments is determined by combining the road topology relationship; The initial state propagation association weights between road segments are calculated based on the intermediate state representation and the state propagation association set, and an initial propagation weight matrix is ​​generated. In the propagation prior bias mechanism, a propagation prior bias matrix is ​​constructed based on the propagation prior information of traffic state. The propagation prior bias matrix and the initial propagation weight matrix are consistent in structural dimension. The propagation prior bias matrix and the initial propagation weight matrix are fused to generate the bias propagation weight matrix. In the calibration constraint mechanism, consistency verification is performed on the state propagation correlation in the bias propagation weight matrix based on the time constraint relationship to generate a calibration reference relationship; constraint adjustment processing is performed on the state propagation correlation strength in the bias propagation weight matrix according to the calibration reference relationship to generate a calibration propagation weight matrix. The intermediate state representation is propagated and aggregated based on the calibration propagation weight matrix to generate the intermediate state representation after propagation guidance. The feature output structure performs feature aggregation processing based on the intermediate state representation after propagation guidance, and outputs spatially consistent features.

7. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S5 is: Obtain local traffic state features, road topology relationships, and prior information on traffic state propagation corresponding to road segments; The local traffic state features are serialized and arranged according to road segmentation and time sequence to generate a feature sequence. Perform structural encoding processing on the road topology relationships to generate a topology encoding sequence; Prior coding processing is performed on the prior information of traffic state propagation to generate a prior coding sequence; Perform dimension alignment processing on the feature sequence, topological coding sequence, and prior coding sequence to generate a coding-aligned sequence; Tensor quantization is performed on the encoded aligned sequence to generate the input tensor; The input tensor is fed into the improved VMamba model, and a dual propagation guidance mechanism is introduced to perform state propagation modeling on the input tensor, outputting space consistency features.

8. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S6 is: Obtain the temporal constraint relationship and spatial consistency features, match the temporal constraint relationship and spatial consistency features according to the road segment identifier, and generate road segment consistency input pairs; Perform fusion processing on the road segment consistency input pairs to generate road segment consistency fusion features; Calculate road segment consistency scores based on road segment consistency fusion features, and generate consistency score sequences. Consistency assessment results are generated based on the consistency score sequence and preset consistency conditions.

9. The road traffic state detection method based on deep learning according to claim 1, characterized in that, Specifically, S7 is: Based on the comparison between the consistency assessment results and the preset consistency conditions, a set of road segment identifiers with a consistency score less than the consistency threshold is determined; the preset consistency conditions include the consistency threshold, which is determined by the statistical results of historical road segment consistency scores. Based on the road segment identifier set, extract the local traffic state features and historical local traffic state features of the corresponding road segments, and construct the historical state sequence of road segments; The correction amount of local traffic state characteristics is calculated based on the historical state sequence of road segments, and a correction vector is generated. The modified vector is fused with local traffic state features to generate traffic state features; The road traffic condition detection results are output based on traffic condition features, corresponding to the road segments.

10. A road traffic state detection system based on deep learning, comprising executing the road traffic state detection method based on deep learning as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The perception processing module is used to acquire continuous traffic perception data covering the target road area, perform time and space alignment processing, and generate traffic perception time series in terms of road segments. The state extraction module is used to extract local traffic state features of the corresponding road segments based on the traffic perception time series, and construct the traffic state evolution sequence of the road segments. The prior generation module is used to generate time constraint relationships based on the traffic state evolution sequence, and to generate prior information for traffic state propagation based on the time constraint relationships; The propagation modeling module is used to obtain the connection relationship between road segments and generate road topology relationship; it performs encoding processing on the local traffic state features, traffic state propagation prior information and road topology relationship to generate input tensor, inputs the input tensor into the improved VMamba model, introduces a dual propagation guidance mechanism to perform traffic state propagation modeling, and outputs spatial consistency features; The evaluation output module is used to integrate the time constraint relationship and the spatial consistency features to generate a consistency evaluation result; if the consistency evaluation result does not meet the preset consistency conditions, the local traffic state features are corrected to generate traffic state features, and the road traffic state detection results of the corresponding road segments are output.

Images