Vehicle rapid identification method and system for ETC free flow

Abstract

The application discloses a vehicle rapid identification method and system for ETC free flow, relates to the technical field of image recognition, and comprises the following steps: collecting vehicle images to be identified and performing time sequence alignment, extracting a multi-dimensional dynamic feature sequence of the same vehicle at continuous spatial positions, inputting the multi-dimensional dynamic feature sequence and the vehicle image into a lightweight time sequence fusion identification model, performing cross-modal correlation alignment, outputting a fusion identification result of the current vehicle, calculating a space-time consistency score, taking the space-time consistency score as a credibility identifier of the fusion identification result, correcting the credibility identifier by using a time sequence correlation coefficient, and outputting a final vehicle identification result for ETC free flow tolling decision. The technical problem that the prior art cannot simultaneously meet the requirements of identification accuracy, low time delay and deployment flexibility in the ETC free flow scene is solved.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, specifically to a method and system for rapid vehicle identification for ETC free-flow. Background Technology

[0002] With the continuous promotion of ETC network toll collection on national expressways, the free-flow tolling mode has been widely used. In the ETC free-flow tolling scenario, vehicles pass through the toll gate quickly at normal speed without stopping or slowing down, which greatly improves the traffic efficiency of expressways.

[0003] However, traditional ETC vehicle recognition relies on a single RSU device interacting with the OBU to obtain vehicle information. When there are scenarios such as adjacent lane interference, vehicles following too closely, or abnormal OBU signals, problems such as license plate recognition errors and vehicle type mismatches are likely to occur, leading to incorrect charges and affecting the accuracy of highway toll collection.

[0004] Meanwhile, existing vehicle recognition solutions in ETC free-flow scenarios mostly rely on single-frame image recognition, failing to fully utilize the temporal dynamic information of multiple frames in continuous vehicle passing scenarios, making it difficult to guarantee the stability of recognition results. In addition, most recognition models are deployed on cloud servers, resulting in delays in the transmission of recognition results, making it difficult to adapt to the low-latency and fast recognition requirements of edge nodes. Large-volume models are also difficult to run on the limited computing resources at the edge. Summary of the Invention

[0005] This application provides a method and system for rapid vehicle identification in ETC free-flow scenarios, which solves the technical problem that existing technologies cannot simultaneously meet the requirements of identification accuracy, low latency, and deployment flexibility in ETC free-flow scenarios.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows:

[0007] In a first aspect, this application provides a method for rapid vehicle identification for ETC free-flow systems, the method comprising:

[0008] Acquire images of the vehicle to be identified, and perform time-series alignment on the images of the vehicle to be identified;

[0009] Extract the multi-dimensional dynamic feature sequence of the same vehicle at continuous spatial locations. The multi-dimensional dynamic feature sequence includes the geometric contour change features of the vehicle body, the illumination invariance features of the license plate area, and the inter-frame offset features of the local texture of the vehicle.

[0010] The multi-dimensional dynamic feature sequence and the vehicle image are input into a lightweight temporal fusion recognition model deployed on an edge computing node, and cross-modal correlation alignment is performed to obtain a unified multi-modal temporal feature tensor. The lightweight temporal fusion recognition model outputs the fusion recognition result of the current vehicle, which includes vehicle model recognition result, license plate recognition result, and confidence identifier.

[0011] Based on the fusion recognition result, a spatiotemporal consistency score is calculated, and the spatiotemporal consistency score is used as the credibility identifier of the fusion recognition result. The credibility identifier is corrected using the temporal correlation coefficient, and the final vehicle recognition result for ETC free-flow tolling decision is output.

[0012] Secondly, this application provides a vehicle rapid identification system for ETC free-flow, including:

[0013] The information acquisition module is used to acquire images of the vehicle to be identified and to perform time-series alignment on the images of the vehicle to be identified.

[0014] The feature sequence extraction module is used to extract multi-dimensional dynamic feature sequences of the same vehicle in continuous spatial positions. The multi-dimensional dynamic feature sequences include geometric contour change features of the vehicle body, illumination invariance features of the license plate area, and inter-frame offset features of local texture of the vehicle.

[0015] The recognition model training module is used to input the multi-dimensional dynamic feature sequence and the vehicle identity information into a lightweight temporal fusion recognition model deployed on an edge computing node, and perform cross-modal correlation alignment to obtain a unified multi-modal temporal feature tensor. The lightweight temporal fusion recognition model outputs the fusion recognition result of the current vehicle, which includes vehicle model recognition result, license plate recognition result, and credibility identifier.

[0016] The identification result output module is used to calculate the spatiotemporal consistency score based on the fusion identification result, and use the spatiotemporal consistency score as the credibility identifier of the fusion identification result. The credibility identifier is corrected using the temporal correlation coefficient, and the final vehicle identification result is output for ETC free-flow tolling decision.

[0017] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0018] This application provides a method and system for rapid vehicle recognition in ETC free-flow tolling scenarios. First, by temporally aligning consecutive vehicle images, it extracts geometric contours, license plate illumination invariance, and multi-dimensional dynamic features of local motion from multiple frames of the same vehicle, addressing the issues of insufficient recognition information and unstable results from a single frame image. Second, a lightweight heterogeneous fusion architecture is used to build a temporal fusion recognition model, adapting to the limited computing resources of edge computing nodes and reducing recognition latency. Third, a credibility indicator is obtained by combining spatiotemporal consistency scoring with temporal correlation and motion physical constraints for dual correction, which measures the reliability of the recognition results and improves recognition accuracy in complex interference scenarios. Finally, through a cloud-edge collaborative difficult example selection and incremental update mechanism, the model is continuously optimized using high-value difficult examples collected from the edge side, balancing deployment flexibility and dynamic improvement of model performance, adapting to the multiple requirements of ETC free-flow tolling scenarios for recognition accuracy, low latency, and deployment scalability.

[0019] Through the above technical solution, this application solves the technical problems of insufficient single-frame recognition information, poor result stability, difficulty in deploying large models at the edge, high recognition latency, and insufficient recognition accuracy in complex interference scenarios under existing ETC free-flow scenarios. It can achieve fast, stable, and accurate ETC vehicle recognition under the limited computing power of edge computing nodes, providing a reliable decision-making basis for highway free-flow toll collection and ensuring toll accuracy and traffic efficiency. At the same time, the incremental optimization mechanism of cloud-edge collaboration allows the model to continuously adapt to constantly changing actual traffic scenarios, possessing good scalability and application prospects. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the vehicle rapid identification method for ETC free-flow provided in the embodiments of this application;

[0022] Figure 2 This is a schematic diagram of the structure of the vehicle rapid identification system for ETC free flow provided in the embodiments of this application;

[0023] Figure 3 This is a schematic diagram of the process for generating a consistency score in the vehicle fast identification method for ETC free flow provided in the embodiments of this application.

[0024] The components represented by each number in the attached diagram are explained below:

[0025] Information acquisition module 11, feature sequence extraction module 12, recognition model training module 13, and recognition result output module 14. Detailed Implementation

[0026] This application provides a method and system for rapid vehicle identification in ETC free-flow scenarios, addressing the technical problem that existing technologies cannot simultaneously meet the requirements of identification accuracy, low latency, and deployment flexibility in ETC free-flow scenarios.

[0027] Example 1, as Figure 1 As shown, this application provides a method for rapid vehicle identification in ETC free-flow mode, including:

[0028] S10: Acquire an image of the vehicle to be identified, and perform time-series alignment on the image of the vehicle to be identified;

[0029] In this embodiment, an image acquisition device deployed on an ETC free-flow gantry acquires images of vehicles to be identified. At a preset time interval, such as 30ms interval, consecutive passing vehicles are captured frame by frame to obtain an image sequence to be identified. Then, the image sequence is timestamped according to the timestamp of the capture device to eliminate the time error of multiple devices capturing images, and ensure that the time sequence of multiple frames of images corresponding to the same vehicle is aligned, providing accurate input for subsequent time sequence feature extraction.

[0030] Specifically, performing time-series alignment on the vehicle image to be identified includes:

[0031] Perform correlation matching on the same vehicle target between adjacent frames;

[0032] When the cross-union ratio of bounding boxes between adjacent frames is greater than a preset association threshold, a consistent mapping of vehicle identities between frames is established.

[0033] When the cross-union ratio of the bounding boxes between adjacent frames is less than or equal to the association threshold, a secondary matching is performed to complete vehicle alignment in low overlap scenarios.

[0034] In this embodiment of the application, the same vehicle target between adjacent frames is associated and matched. Specifically, the intersection-union ratio of the vehicle bounding boxes detected in different frames is first calculated.

[0035] When the intersection-union ratio of the same vehicle candidate box in adjacent frames is greater than the preset association threshold, such as the preset threshold being 0.3, the identity consistency mapping of the vehicle in different frames is directly established to complete the initial alignment.

[0036] When the cross-union ratio of the bounding boxes of vehicles in adjacent frames is less than or equal to the preset association threshold due to high traffic speed and large capture interval, a second matching is performed by combining the similarity of the vehicle's appearance features and the string similarity of the license plate recognition results. After successful matching, an inter-frame identity consistency mapping is also established to complete vehicle alignment in low overlap vehicle following scenarios, avoid the same vehicle being mismatched as different targets, and ensure the accuracy of subsequent temporal feature extraction.

[0037] S20: Extract the multi-dimensional dynamic feature sequence of the same vehicle in continuous spatial positions. The multi-dimensional dynamic feature sequence includes the geometric contour change features of the vehicle body, the illumination invariance features of the license plate area, and the inter-frame offset features of the local texture of the vehicle.

[0038] In this embodiment, after completing the inter-frame temporal alignment of the same vehicle, the geometric contour change features of the vehicle body, the illumination invariant features of the license plate area, and the inter-frame offset features of the vehicle's local texture are extracted from the aligned multi-frame images. Specifically, the geometric contour change features of the vehicle body are used to characterize the pitch attitude change and relative distance change of the vehicle during free-flowing traffic; the illumination invariant features of the license plate area are used to eliminate the interference of sudden changes in illumination in scenes such as backlighting and shade, ensuring the stability of the license plate feature extraction; and the inter-frame offset features of the vehicle's local texture are used to characterize the nonlinear motion offset law of the vehicle surface texture between frames.

[0039] Specifically, this involves extracting multi-dimensional dynamic feature sequences of the same vehicle at continuous spatial locations, including:

[0040] Extract the minimum bounding rectangle of the vehicle from multiple consecutive frames of vehicle images, calculate the aspect ratio change rate and area change rate of the minimum bounding rectangle between adjacent frames, concatenate the aspect ratio change rate and the area change rate along the time axis to form the geometric deformation trajectory feature vector of the vehicle, and use the geometric deformation trajectory feature vector of the vehicle as the geometric contour change feature of the vehicle body.

[0041] Each frame of the passing vehicle image is decomposed into an illumination component and a reflection component. The reflection component is extracted as a license plate candidate region after illumination normalization. Orientation gradient histogram features and local binary pattern features are extracted from the license plate candidate region. The orientation gradient histogram features and the local binary pattern features are serially fused to obtain the illumination-invariant features of the license plate region.

[0042] The optical flow field between adjacent frames is calculated, pixel-level motion vectors are extracted within the vehicle bounding box, the optical flow field is divided into blocks for statistical analysis, the mean and variance of the optical flow vectors in each block are calculated to form local motion feature vectors, and the local motion feature vectors are used as inter-frame offset features of the vehicle's local texture.

[0043] In this embodiment, firstly, the minimum bounding rectangle of the vehicle is extracted from multiple consecutive frames of vehicle passing images. The overall bounding range of the vehicle is located by contour detection. Then, the aspect ratio change rate and area change rate of the minimum bounding rectangle between two adjacent frames are calculated sequentially. All the obtained changes are concatenated along the time axis to construct a unique vehicle geometric deformation trajectory feature vector. This vector can fully reflect the contour deformation law caused by changes in shooting angle and relative distance during vehicle passage, and serves as the vehicle body geometric contour change feature required by this solution.

[0044] For example, if an SUV passes through three capture points in succession, and the aspect ratio change rates between adjacent frames are 0.08 and 0.12, and the area change rates are 0.15 and 0.21, respectively, the geometric contour change feature vector corresponding to the vehicle can be obtained by concatenating the four values ​​in chronological order.

[0045] Secondly, the Retinex decomposition algorithm is used to decompose the license plate candidate region in each frame of the vehicle passing image into illumination component and reflection component. The reflection component, which is not affected by illumination, is extracted as the license plate region after illumination normalization. Then, the directional gradient histogram feature and local binary pattern feature are extracted on the processed license plate region respectively. The two types of features are fused in series to obtain the illumination-invariant features of the license plate region that can resist illumination interference.

[0046] Specifically, the Retinex decomposition algorithm is based on the human visual system, assuming that an image can be decomposed into an illuminance component of incident light and a reflection component of the object's inherent reflectivity. The reflection component corresponds to the object's inherent features and is unaffected by changes in illumination, perfectly meeting the feature extraction requirements for license plate areas in ETC free-flow scenarios, thus separating the inherent texture features of the license plate area from overall illumination interference in uneven lighting conditions. The extraction of histogram of oriented gradients (HOR) and local binary pattern features involves calculating the oriented gradient distribution unit by unit in the license plate candidate region using a preset step size and cell size to obtain the HOR feature. Then, local binary pattern features are encoded according to a fixed radius and neighborhood comparison rules. Finally, the HOR and local binary pattern features are concatenated dimensionally to complete the fusion.

[0047] Finally, the Lucas-Kanade algorithm is used to calculate the dense optical flow field of the vehicle region between adjacent frames. Pixel-level motion vectors are extracted within the bounding box obtained by vehicle detection. Then, the entire optical flow field is divided into blocks according to a preset size. The mean and variance of all optical flow vectors in each block are calculated and spliced ​​to obtain the local motion feature vector of the corresponding block. This feature vector can characterize the inter-frame offset pattern of the local texture of the vehicle.

[0048] Specifically, the Lucas-Kanade algorithm is based on the local smoothness assumption, which assumes that the optical flow is constant in the neighborhood of a pixel. By minimizing the sum of squared grayscale errors of all pixels in the neighborhood, the optical flow vector of the current pixel is obtained. This algorithm has low computational complexity and few parameters, and only requires less computing power to estimate the inter-frame motion offset. It is suitable for the computing power limitations of edge computing nodes in ETC free-flow scenarios and can quickly output inter-frame motion information that meets the accuracy requirements.

[0049] By concatenating the above three types of features along the temporal dimension, a multi-dimensional dynamic feature sequence of the same vehicle is obtained, which serves as the input for the subsequent recognition model.

[0050] S30: The multi-dimensional dynamic feature sequence and the vehicle image are input into a lightweight temporal fusion recognition model deployed on an edge computing node, and cross-modal correlation alignment is performed to obtain a unified multi-modal temporal feature tensor. The lightweight temporal fusion recognition model outputs the fusion recognition result of the current vehicle, which includes vehicle model recognition result, license plate recognition result, and confidence identifier.

[0051] In this embodiment, the multi-dimensional dynamic feature sequence and the original vehicle image are respectively input into different branches of the lightweight temporal fusion recognition model. The image branch uses a lightweight convolutional neural network to extract single-frame static visual features, while the temporal branch uses a temporal convolutional network to perform temporal modeling on the multi-dimensional dynamic feature sequence. Then, the features output by the two branches are aligned across modal correlations to finally obtain a unified multi-modal temporal feature tensor.

[0052] Multimodal temporal feature tensors are fed into the vehicle type classification head and license plate recognition head respectively, outputting preliminary vehicle type recognition results and license plate recognition results. At the same time, the initial confidence level output by the model is retained as the initial calculation benchmark for the credibility. Adapting to the computing power of edge computing nodes, the computational amount of the entire inference process is only about 30% of that of a conventional large convolutional model. It can complete fast inference locally without uploading to the cloud for processing, avoiding the decrease in traffic efficiency caused by network latency.

[0053] The lightweight temporal fusion recognition model is constructed using a heterogeneous fusion architecture of temporal convolutional networks and lightweight convolutional neural networks.

[0054] The lightweight convolutional neural network uses MobileNetV3 as its backbone network to extract spatial features of a single frame image.

[0055] The temporal convolutional network employs a dilated convolutional structure to perform temporal modeling of spatial features across multiple consecutive frames along the time dimension.

[0056] The receptive field length of the temporal convolutional network is matched with the number of input frames, and the dilation rate of the dilated convolution increases exponentially with the increase of the number of network layers.

[0057] In this embodiment, the lightweight convolutional neural network first uses MobileNetV3 as the backbone network and replaces conventional convolution with depthwise separable convolution. While ensuring feature extraction capabilities, it significantly reduces the number of model parameters and computational load, enabling rapid extraction of spatial visual features from a single frame of vehicle image.

[0058] Secondly, the temporal convolutional network adopts a one-dimensional temporal convolutional structure with dilated convolution. Dilated convolution can expand the receptive field of convolution without increasing the number of parameters, and can capture temporal dependencies over a longer time span. Moreover, the dilation rate increases exponentially with the number of network layers, which matches the temporal length requirement of multi-frame input in this method and can effectively model the dynamic correlation between consecutive vehicle passing frames.

[0059] For example, the specific steps for building and training a lightweight temporal fusion recognition model based on a convolutional neural network are as follows:

[0060] First, data preparation: a dataset of labeled ETC free-flowing vehicle image sequences was collected based on historical data. The dataset contains labeled samples with different lighting conditions, different traffic speeds, and different vehicle types. The training set and validation set were divided in an 8:2 ratio.

[0061] Secondly, for model construction, the parameters of the MobileNetV3 backbone network pre-trained on ImageNet are initialized, and the parameters of the model branches are randomly initialized. The number of nodes in the input layer is equal to the dimension of the input features. For example, if there are four features, namely vehicle image, geometric contour change features of the vehicle body, illumination invariance features of the license plate area, and inter-frame offset features of local vehicle texture, then the input layer contains 4 nodes. 1-3 hidden layers are set, and the number of nodes in each layer is adjusted through experiments, such as 64, 32, etc. The activation function is ReLU. The number of nodes in the output layer corresponds to the number of vehicle type categories, the dimension of license plate character encoding, and 1 confidence output node. For example, if the recognition time requires 3 nodes, the output layer generally does not use an activation function and directly outputs continuous values.

[0062] Next, for model training, the Adam algorithm was used with an initial learning rate of 0.01 and a batch size of 32. The cross-entropy loss function was used for multi-task joint training. During training, the learning rate was adjusted using a cosine annealing strategy. The model performance was verified every 5 rounds of training, and the model weights with the highest recognition accuracy on the validation set were retained. The trained model was then processed using quantization compression technology and deployed to edge computing nodes to complete the model's deployment.

[0063] Furthermore, after completing feature extraction and temporal modeling, cross-modal correlation alignment is performed on the features output from the two branches, mapping the features of different modalities to a unified feature space, and concatenating them to obtain a multimodal temporal feature tensor. Finally, the tensor is connected to different output heads to obtain preliminary fusion recognition results.

[0064] Specifically, cross-modal correlation alignment is performed to obtain a unified multimodal temporal feature tensor, including:

[0065] Calculate the temporal correlation coefficient between the rate of change of geometric contour and texture offset of the same vehicle in consecutive frames, and screen out the feature channels that are consistent with the trend of geometric deformation and texture motion.

[0066] Spatial attention weighting of geometric contour features is applied by utilizing the illumination-invariant features of the license plate region.

[0067] After dynamically aligning the geometric contour variation features and texture offset features in the time dimension, they are concatenated with the illumination-invariant features to form a unified multimodal temporal feature tensor.

[0068] The mean of the temporal correlation coefficient is calculated over the effective frame sequence, and this mean is used as the cross-modal matching quality index.

[0069] In this embodiment, firstly, the Pearson correlation coefficient between the geometric contour change rate and the corresponding texture offset in consecutive frames of the same vehicle is calculated to obtain the temporal correlation coefficient of each feature channel. If the temporal correlation coefficient is greater than a preset screening threshold, the feature of that channel is retained. If the temporal correlation coefficient is less than the preset screening threshold, it is determined that the channel has feature noise, and the channel is removed. Only effective channels with consistent motion trends of the two types of features are retained to avoid irrelevant noise interfering with the subsequent recognition accuracy. The preset screening threshold is used to filter invalid feature fluctuations. In practical applications, it can be adjusted according to the features of the dataset, for example, it can be set to 0.6.

[0070] Specifically, the Pearson correlation coefficient is used to measure the degree of linear correlation between two variables. Its value ranges from [-1, 1]. The larger the absolute value, the stronger the correlation. This scheme uses it to determine whether the deformation of the vehicle body contour and the local texture movement conform to the consistent law of the overall vehicle movement, thereby screening out reliable feature channels.

[0071] Secondly, the L2 norm of the illumination-invariant features of the license plate region is used as the spatial attention weight to weight the geometric contour feature channels, thereby strengthening the feature response of the license plate region and suppressing the pseudo-changes in the contour caused by illumination changes. For example, in a tree-shaded scene with strong illumination, after illumination normalization, the L2 norm of the features of the license plate region is much higher than that of the background interference region. After weighting, the response of the license plate features is significantly enhanced, which allows the model to pay more attention to the effective features of the license plate region.

[0072] Then, the geometric contour change features and local texture offset features are aligned in the time dimension by using the dynamic time warping algorithm to eliminate the temporal offset caused by the different extraction methods of the two types of features. After alignment, the two types of features are concatenated with the illumination-invariant features according to the dimension to obtain a unified multimodal temporal feature tensor.

[0073] Among them, the dynamic time warping algorithm can normalize the temporal offset between different feature sequences by nonlinearly stretching and matching the time dimension of two sequences of different lengths, so that the feature sequences are in sync in the time dimension and ensure the alignment accuracy of multi-feature splicing.

[0074] Finally, the calculated temporal correlation coefficients are averaged over all the filtered valid frame sequences. The resulting average is the quality index of the current cross-modal matching, which is used to evaluate the feature consistency and motion authenticity of the currently collected data.

[0075] S40: Based on the fusion recognition result, calculate the spatiotemporal consistency score, and use the spatiotemporal consistency score as the credibility identifier of the fusion recognition result. Correct the credibility identifier using the temporal correlation coefficient, and output the final vehicle recognition result for ETC free-flow tolling decision.

[0076] In this embodiment of the application, after the fusion recognition result is output, the score is calculated based on the rule that the recognition result of the same vehicle during the vehicle passage should conform to the spatiotemporal consistency. That is, when the same vehicle passes through adjacent capture points continuously, the vehicle type and license plate results should not change abruptly, and the position and speed changes of the recognition result should conform to the motion law of the vehicle passing at a constant speed.

[0077] Specifically, such as Figure 3 As shown, based on the fusion recognition result, a spatiotemporal consistency score is calculated, and the spatiotemporal consistency score is used as the credibility identifier of the fusion recognition result. The credibility identifier is corrected using the temporal correlation coefficient, including:

[0078] Obtain N vehicle model recognition results and N license plate recognition results corresponding to N consecutive frames of vehicle passing images;

[0079] The proportion of the most frequent vehicle type in the N vehicle type recognition results is used as the vehicle type consistency score.

[0080] Calculate the average string similarity between each pair of the N license plate recognition results, and use it as the license plate consistency score;

[0081] The cross-modal matching quality index is normalized and used as a prior weight for data quality.

[0082] Multiply the average of the vehicle model consistency score and the license plate consistency score by the prior weight of the data quality to obtain the spatio-temporal consistency score, and use this spatio-temporal consistency score as the credibility identifier of the current recognition result;

[0083] When the temporal correlation coefficient is lower than the preset credibility threshold, proportionally reduce the output value of the current credibility identifier. When the temporal correlation coefficient is higher than or equal to the preset credibility threshold, keep the original spatio-temporal consistency score as the credibility identifier output.

[0084] In the embodiments of the present application, first, obtain N recognition results corresponding to N consecutive vehicle passing images. Since multiple frames of images captured continuously correspond to the same passing vehicle, the correct recognition results should be consistent in multiple frames. Therefore, count the proportion corresponding to the vehicle model with the highest frequency of occurrence in all vehicle model recognition results, and directly use this proportion as the vehicle model consistency score. The higher the proportion, the more stable the vehicle model recognition result and the higher the credibility. For example, among 5 consecutive vehicle passing images, 4 frames recognize the vehicle model as SUV and 1 frame recognizes it as a sedan. Then the occurrence proportion of the highest frequency vehicle model SUV is 0.8, and the corresponding vehicle model consistency score is 0.8.

[0085] Second, for all N license plate recognition results, calculate the edit distance between every two license plate strings in sequence, and then convert it to a string similarity based on the edit distance. Take the average of the similarities of all pairwise combinations as the license plate consistency score. The higher the average similarity, the more stable the license plate recognition result and the higher the credibility. For example, among 5 consecutive recognition results, 4 frames recognize the license plate as "Beijing A12345" and 1 frame as "Beijing A12346". Calculate the average string similarity as 0.98 after calculating the edit distance pairwise, and the corresponding license plate consistency score is 0.98.

[0086] Third, normalize the mean of the cross-modal matching quality index obtained above, that is, the temporal correlation coefficient, to the interval from 0 to 1 as the prior weight of the data quality of the current recognition result. The higher the cross-modal matching quality, the better the consistency of feature extraction and the higher the benchmark of the recognition result credibility. For example, the normalization method can adopt linear scaling, linearly map the value range from [-1, 1] to [0, 1]. If the original value of the cross-modal matching quality index is 0.6, the normalized value obtained after scaling is , which is directly used as the prior weight for calculation.

[0087] After that, calculate the arithmetic mean of the vehicle model consistency score and the license plate consistency score, multiply this mean by the normalized prior weight of the cross-modal matching quality to obtain the final spatio-temporal consistency score, and directly use it as the credibility identifier of the current recognition result.

[0088] For example, if the vehicle model consistency score is 0.8, the license plate consistency score is 0.98, and the normalized cross-modal matching quality prior weight is 0.8, then the final spatiotemporal consistency score is... This value is the credibility indicator of the current recognition result.

[0089] Finally, the credibility identifier is further corrected using the overall temporal correlation coefficient. Specifically, when the average temporal correlation coefficient of all retained channels is lower than a preset credibility threshold (e.g., the preset threshold can be set to 0.5), the output value of the current credibility identifier is reduced according to the proportion of correlation below the threshold. When the average temporal correlation coefficient is higher than or equal to the preset credibility threshold, the original spatiotemporal consistency score is directly retained as the credibility identifier output. The corrected credibility identifier will be output together with the vehicle type recognition result and license plate recognition result, directly provided to the ETC free-flow system to complete the toll deduction decision. If the credibility is lower than the threshold required for toll deduction, a manual review process can be triggered, which ensures the passage efficiency in normal scenarios and reduces the recognition error rate in low-quality scenarios.

[0090] Furthermore, the spatiotemporal consistency score also includes:

[0091] Based on the scale changes of the vehicle bounding box in multiple consecutive frames of vehicle passing images, the instantaneous velocity sequence of the vehicle relative to the gantry is calculated.

[0092] Verify whether the variance of the instantaneous velocity sequence exceeds a preset physical consistency threshold;

[0093] If the variance of fluctuation exceeds the physical consistency threshold, a motion consistency penalty between 0 and 1 is generated, and the value of the penalty is negatively correlated with the variance of fluctuation.

[0094] Multiplying the spatiotemporal consistency score by the motion consistency penalty term yields the credibility identifier after motion physics constraint correction.

[0095] In this embodiment, firstly, based on the fixed spacing of the capture points in the continuous vehicle passing frames and the timestamps of adjacent frames, the instantaneous speed of the vehicle passing each capture position is calculated. Since the vehicle usually maintains an approximately uniform speed during free-flowing traffic, the instantaneous speed sequence should not exhibit excessively large abnormal fluctuations. Therefore, the fluctuation variance of the instantaneous speed sequence is calculated. If the fluctuation variance exceeds a preset physical consistency threshold, it indicates that the identified vehicle position is abnormal, which is most likely caused by feature interference or cross-vehicle recognition errors. At this time, a corresponding motion consistency penalty is generated. The larger the fluctuation variance, the smaller the penalty value. The penalty value always remains between 0 and 1. Then, the original spatiotemporal consistency score is multiplied by the penalty, which can perform physical constraint correction on the credibility identifier, further improving the credibility identifier's ability to detect recognition errors.

[0096] Specifically, when a vehicle is traveling at a normal, constant speed, the positional changes between adjacent frames are stable, the fluctuation variance is smaller, and the penalty will be closer to 1, having almost no impact on the original score. However, if cross-vehicle recognition occurs, the local features of the preceding vehicle and the features of the following vehicle are spliced ​​together, causing a sudden change in the calculated instantaneous speed, and the fluctuation variance greatly exceeds the threshold. In this case, the penalty is set to a smaller value, such as 0.3, to lower the final credibility rating and trigger subsequent manual review, effectively avoiding billing disputes caused by cross-vehicle recognition errors.

[0097] Furthermore, the rapid vehicle identification method for ETC free-flow also includes:

[0098] When the credibility indicator is lower than the preset difficult case trigger threshold, the cross-modal matching quality index of the current vehicle is obtained;

[0099] If the cross-modal matching quality index is greater than or equal to the preset high quality threshold, the multi-frame vehicle passing images of the current vehicle, vehicle identity information, and fusion recognition results are packaged into high-value edge hard case samples.

[0100] If the cross-modal matching quality index is less than the preset high quality threshold, the current sample is marked as a low-quality noise sample and is either discarded directly or only retained for statistical information without participating in model updates.

[0101] In this embodiment, firstly, the preset difficult case trigger threshold is the confidence screening threshold during the model update stage. In practical applications, it can be adjusted according to the storage capacity and upload bandwidth of the edge node. For example, it can be set to 0.5. When the confidence indicator of the recognition result is lower than the preset difficult case trigger threshold, it indicates that the current recognition result has high uncertainty and belongs to the difficult case sample that needs to be manually reviewed. At this time, the difficult cases are distinguished according to the cross-modal matching quality index.

[0102] Specifically, if the cross-modal matching quality index is greater than or equal to the preset high quality threshold, it indicates that the currently collected features are consistent, but there is difficulty in identification. Such samples have high value for subsequent optimization of the model's recognition ability. Therefore, the multi-frame images of the current vehicle, the labeled vehicle identity information, and the fusion recognition results obtained this time are packaged and sent back to the server as high-value edge difficult example samples for subsequent model iteration updates, continuously improving the model's ability to identify complex and difficult examples.

[0103] If the cross-modal matching quality index is less than the preset high-quality threshold, it indicates that the sample itself is low-quality data caused by extreme scenarios such as sudden changes in illumination, vehicle occlusion, and severe speeding. Such samples have low reference value for model optimization. Therefore, they are marked as low-quality noise samples, and only relevant statistical data are retained for scene analysis. They are not included in the subsequent model training dataset to avoid low-quality noise interfering with the model training effect.

[0104] Furthermore, after obtaining high-value edge case samples, the process includes:

[0105] Upload the high-value, difficult-to-produce sample to the cloud.

[0106] After receiving the edge hard example samples, the cloud extracts the vehicle appearance features and license plate style features from them, and uses a conditional generative adversarial network to generate a synthetic training image with the same vehicle appearance and license plate style as the edge hard example samples.

[0107] An incremental training dataset is constructed based on the synthetic training images. The lightweight temporal fusion recognition model is incrementally fine-tuned to generate an updated lightweight temporal fusion recognition model, which is then distributed to the edge computing node.

[0108] In this embodiment, firstly, the number of high-value difficult example samples collected by edge nodes is limited, and directly using them for incremental model training is prone to overfitting. Therefore, a conditional generative adversarial network in the cloud is used to generate synthetic training images of the same style based on the vehicle appearance and license plate features of real difficult example samples, thereby expanding the training data scale of difficult example scenarios and improving the effect of incremental training.

[0109] After generation, the synthesized image is combined with the original real hard example samples to construct an incremental training dataset. The lightweight temporal fusion recognition model originally deployed on the edge nodes is incrementally fine-tuned. While retaining the original recognition ability of the model, the model's adaptability to such hard example scenarios is enhanced. After fine-tuning, the updated model is distributed to each edge computing node to complete the online iterative update of the model. It does not require the collection of large-scale full data for retraining, nor does it require manual re-labeling of a large amount of data. This can effectively reduce the cost of model iteration and improve the efficiency of model optimization.

[0110] In summary, compared with existing technologies, this application effectively solves the recognition error problems caused by illumination interference and multi-vehicle sequence overlap in the ETC free-flow scenario through multimodal feature fusion and spatiotemporal consistency verification. At the same time, through edge-side hard example screening and cloud-based incremental iteration, the model is continuously optimized online, balancing recognition accuracy and traffic efficiency, and can adapt to the actual deployment needs of the ETC free-flow toll system.

[0111] In practical applications, this method can run directly on the edge computing device on the gantry side without the need for additional hardware acquisition equipment, resulting in low deployment costs and the ability to quickly adapt to the upgrade and transformation needs of the existing ETC free-flow system.

[0112] In summary, the embodiments of this application have at least the following technical effects:

[0113] This application provides a fast vehicle recognition method for ETC free-flow tolling. First, by temporally aligning consecutive vehicle images, it extracts geometric contours, license plate illumination invariance, and multi-dimensional dynamic features of local motion from multiple frames of the same vehicle, addressing the issues of insufficient recognition information and unstable results from a single frame image. Second, a lightweight heterogeneous fusion architecture is used to build a temporal fusion recognition model, adapting to the limited computing resources of edge computing nodes and reducing recognition latency. Third, a credibility indicator is obtained by combining spatiotemporal consistency scoring with temporal correlation and motion physical constraints for dual correction, which can measure the reliability of the recognition results and improve recognition accuracy in complex interference scenarios. Finally, through a cloud-edge collaborative hard example screening and incremental update mechanism, the model is continuously optimized using high-value hard examples collected from the edge side, balancing deployment flexibility and dynamic improvement of model performance, adapting to the multiple requirements of ETC free-flow tolling scenarios for recognition accuracy, low latency, and deployment scalability.

[0114] Through the above technical solution, this application solves the technical problems of insufficient single-frame recognition information, poor result stability, difficulty in deploying large models at the edge, high recognition latency, and insufficient recognition accuracy in complex interference scenarios under existing ETC free-flow scenarios. It can achieve fast, stable, and accurate ETC vehicle recognition under the limited computing power of edge computing nodes, providing a reliable decision-making basis for highway free-flow toll collection and ensuring toll accuracy and traffic efficiency. At the same time, the incremental optimization mechanism of cloud-edge collaboration allows the model to continuously adapt to constantly changing actual traffic scenarios, possessing good scalability and application prospects.

[0115] Example 2, as Figure 2 As shown, based on the same inventive concept as the vehicle rapid identification method for ETC free flow provided in Embodiment 1, this application also provides a vehicle rapid identification system for ETC free flow, including:

[0116] Information acquisition module 11 is used to acquire images of vehicles to be identified and to perform time-series alignment on the images of vehicles to be identified;

[0117] The feature sequence extraction module 12 is used to extract multi-dimensional dynamic feature sequences of the same vehicle in continuous spatial positions. The multi-dimensional dynamic feature sequences include geometric contour change features of the vehicle body, illumination invariance features of the license plate area, and inter-frame offset features of local texture of the vehicle.

[0118] The recognition model training module 13 is used to input the multi-dimensional dynamic feature sequence and the vehicle identity information into a lightweight temporal fusion recognition model deployed on an edge computing node, and perform cross-modal correlation alignment to obtain a unified multi-modal temporal feature tensor. The lightweight temporal fusion recognition model outputs the fusion recognition result of the current vehicle, which includes vehicle model recognition result, license plate recognition result, and credibility identifier.

[0119] The identification result output module 14 is used to calculate the spatiotemporal consistency score based on the fusion identification result, and use the spatiotemporal consistency score as the credibility identifier of the fusion identification result. The credibility identifier is corrected using the temporal correlation coefficient, and the final vehicle identification result is output for ETC free flow tolling decision.

[0120] In one embodiment, temporal alignment of the image of the vehicle to be identified includes:

[0121] Perform correlation matching on the same vehicle target between adjacent frames;

[0122] When the cross-union ratio of bounding boxes between adjacent frames is greater than a preset association threshold, a consistent mapping of vehicle identities between frames is established.

[0123] When the cross-union ratio of the bounding boxes between adjacent frames is less than or equal to the association threshold, a secondary matching is performed to complete vehicle alignment in low overlap scenarios.

[0124] In one embodiment, extracting a multi-dimensional dynamic feature sequence of the same vehicle at continuous spatial locations includes:

[0125] Extract the minimum bounding rectangle of the vehicle from multiple consecutive frames of vehicle images, calculate the aspect ratio change rate and area change rate of the minimum bounding rectangle between adjacent frames, concatenate the aspect ratio change rate and the area change rate along the time axis to form the geometric deformation trajectory feature vector of the vehicle, and use the geometric deformation trajectory feature vector of the vehicle as the geometric contour change feature of the vehicle body.

[0126] Each frame of the passing vehicle image is decomposed into an illumination component and a reflection component. The reflection component is extracted as a license plate candidate region after illumination normalization. Orientation gradient histogram features and local binary pattern features are extracted from the license plate candidate region. The orientation gradient histogram features and the local binary pattern features are serially fused to obtain the illumination-invariant features of the license plate region.

[0127] The optical flow field between adjacent frames is calculated, pixel-level motion vectors are extracted within the vehicle bounding box, the optical flow field is divided into blocks for statistical analysis, the mean and variance of the optical flow vectors in each block are calculated to form local motion feature vectors, and the local motion feature vectors are used as inter-frame offset features of the vehicle's local texture.

[0128] Furthermore, cross-modal correlation alignment is performed to obtain a unified multimodal temporal feature tensor, including:

[0129] Calculate the temporal correlation coefficient between the rate of change of geometric contour and texture offset of the same vehicle in consecutive frames, and screen out the feature channels that are consistent with the trend of geometric deformation and texture motion.

[0130] Spatial attention weighting of geometric contour features is applied by utilizing the illumination-invariant features of the license plate region.

[0131] After dynamically aligning the geometric contour variation features and texture offset features in the time dimension, they are concatenated with the illumination-invariant features to form a unified multimodal temporal feature tensor.

[0132] The mean of the temporal correlation coefficient is calculated over the effective frame sequence, and this mean is used as the cross-modal matching quality index.

[0133] Furthermore, in one embodiment of the application, the lightweight temporal fusion recognition model is constructed using a heterogeneous fusion architecture of temporal convolutional networks and lightweight convolutional neural networks;

[0134] The lightweight convolutional neural network uses MobileNetV3 as its backbone network to extract spatial features of a single frame image.

[0135] The temporal convolutional network employs a dilated convolutional structure to perform temporal modeling of spatial features across multiple consecutive frames along the time dimension.

[0136] The receptive field length of the temporal convolutional network is matched with the number of input frames, and the dilation rate of the dilated convolution increases exponentially with the increase of the number of network layers.

[0137] Further, in one embodiment, based on the fusion recognition result, a spatiotemporal consistency score is calculated, and the spatiotemporal consistency score is used as a credibility identifier of the fusion recognition result. The credibility identifier is then corrected using a temporal correlation coefficient, including:

[0138] Obtain N vehicle model recognition results and N license plate recognition results corresponding to N consecutive frames of vehicle passing images;

[0139] The proportion of the most frequent vehicle type in the N vehicle type recognition results is used as the vehicle type consistency score.

[0140] Calculate the average string similarity between each pair of the N license plate recognition results, and use it as the license plate consistency score;

[0141] The cross-modal matching quality index is normalized and used as a prior weight for data quality.

[0142] The average of the vehicle model consistency score and the license plate consistency score is multiplied by the data quality prior weight to obtain the spatiotemporal consistency score, and this spatiotemporal consistency score is used as the credibility indicator of the current recognition result.

[0143] When the time-series correlation coefficient is lower than the preset confidence threshold, the output value of the current confidence identifier is reduced proportionally. When the time-series correlation coefficient is higher than or equal to the preset confidence threshold, the original spatiotemporal consistency score is maintained as the confidence identifier output.

[0144] Furthermore, the spatiotemporal consistency score also includes:

[0145] Based on the scale changes of the vehicle bounding box in multiple consecutive frames of vehicle passing images, the instantaneous velocity sequence of the vehicle relative to the gantry is calculated.

[0146] Verify whether the variance of the instantaneous velocity sequence exceeds a preset physical consistency threshold;

[0147] If the variance of fluctuation exceeds the physical consistency threshold, a motion consistency penalty between 0 and 1 is generated, and the value of the penalty is negatively correlated with the variance of fluctuation.

[0148] Multiplying the spatiotemporal consistency score by the motion consistency penalty term yields the credibility identifier after motion physics constraint correction.

[0149] Furthermore, the rapid vehicle identification method for ETC free-flow also includes:

[0150] When the credibility indicator is lower than the preset difficult case trigger threshold, the cross-modal matching quality index of the current vehicle is obtained;

[0151] If the cross-modal matching quality index is greater than or equal to the preset high quality threshold, the multi-frame vehicle passing images of the current vehicle, vehicle identity information, and fusion recognition results are packaged into high-value edge hard case samples.

[0152] If the cross-modal matching quality index is less than the preset high quality threshold, the current sample is marked as a low-quality noise sample and is either discarded directly or only retained for statistical information without participating in model updates.

[0153] Furthermore, after obtaining high-value edge case samples, the process includes:

[0154] Upload the high-value, difficult-to-produce sample to the cloud.

[0155] After receiving the edge hard example samples, the cloud extracts the vehicle appearance features and license plate style features from them, and uses a conditional generative adversarial network to generate a synthetic training image with the same vehicle appearance and license plate style as the edge hard example samples.

[0156] An incremental training dataset is constructed based on the synthetic training images. The lightweight temporal fusion recognition model is incrementally fine-tuned to generate an updated lightweight temporal fusion recognition model, which is then distributed to the edge computing node.

Claims

1. A method for rapid vehicle identification in ETC free-flow systems, characterized in that, include: Acquire images of the vehicle to be identified, and perform time-series alignment on the images of the vehicle to be identified; Extract the multi-dimensional dynamic feature sequence of the same vehicle at continuous spatial locations. The multi-dimensional dynamic feature sequence includes the geometric contour change features of the vehicle body, the illumination invariance features of the license plate area, and the inter-frame offset features of the local texture of the vehicle. The multi-dimensional dynamic feature sequence and the vehicle image are input into a lightweight temporal fusion recognition model deployed on an edge computing node, and cross-modal correlation alignment is performed to obtain a unified multi-modal temporal feature tensor. The lightweight temporal fusion recognition model outputs the fusion recognition result of the current vehicle, which includes vehicle model recognition result, license plate recognition result, and confidence identifier. Based on the fusion recognition result, a spatiotemporal consistency score is calculated, and the spatiotemporal consistency score is used as the credibility identifier of the fusion recognition result. The credibility identifier is corrected using the temporal correlation coefficient, and the final vehicle recognition result for ETC free-flow tolling decision is output.

2. The method for rapid vehicle identification for ETC free-flow according to claim 1, characterized in that, Performing time-series alignment on the vehicle image to be identified includes: Perform correlation matching on the same vehicle target between adjacent frames; When the cross-union ratio of bounding boxes between adjacent frames is greater than a preset association threshold, a consistent mapping of vehicle identities between frames is established. When the cross-union ratio of the bounding boxes between adjacent frames is less than or equal to the association threshold, a secondary matching is performed to complete vehicle alignment in low overlap scenarios.

3. The method for rapid vehicle identification for ETC free-flow according to claim 1, characterized in that, Extracting multi-dimensional dynamic feature sequences of the same vehicle at continuous spatial locations, including: Extract the minimum bounding rectangle of the vehicle from multiple consecutive frames of vehicle images, calculate the aspect ratio change rate and area change rate of the minimum bounding rectangle between adjacent frames, concatenate the aspect ratio change rate and the area change rate along the time axis to form the geometric deformation trajectory feature vector of the vehicle, and use the geometric deformation trajectory feature vector of the vehicle as the geometric contour change feature of the vehicle body. Each frame of the passing vehicle image is decomposed into an illumination component and a reflection component. The reflection component is extracted as a license plate candidate region after illumination normalization. Orientation gradient histogram features and local binary pattern features are extracted from the license plate candidate region. The orientation gradient histogram features and the local binary pattern features are serially fused to obtain the illumination-invariant features of the license plate region. The optical flow field between adjacent frames is calculated, pixel-level motion vectors are extracted within the vehicle bounding box, the optical flow field is divided into blocks for statistical analysis, the mean and variance of the optical flow vectors in each block are calculated to form local motion feature vectors, and the local motion feature vectors are used as inter-frame offset features of the vehicle's local texture.

4. The method for rapid vehicle identification for ETC free-flow according to claim 3, characterized in that, Cross-modal correlation alignment is performed to obtain a unified multimodal temporal feature tensor, including: Calculate the temporal correlation coefficient between the rate of change of geometric contour and texture offset of the same vehicle in consecutive frames, and screen out the feature channels that are consistent with the trend of geometric deformation and texture motion. Spatial attention weighting of geometric contour features is applied by utilizing the illumination-invariant features of the license plate region. After dynamically aligning the geometric contour variation features and texture offset features in the time dimension, they are concatenated with the illumination-invariant features to form a unified multimodal temporal feature tensor. The mean of the temporal correlation coefficient is calculated over the effective frame sequence, and this mean is used as the cross-modal matching quality index.

5. The method for rapid vehicle identification for ETC free-flow according to claim 1, characterized in that, The lightweight temporal fusion recognition model is constructed using a heterogeneous fusion architecture of temporal convolutional networks and lightweight convolutional neural networks; The lightweight convolutional neural network uses MobileNetV3 as its backbone network to extract spatial features of a single frame image. The temporal convolutional network employs a dilated convolutional structure to perform temporal modeling of spatial features across multiple consecutive frames along the time dimension. The receptive field length of the temporal convolutional network is matched with the number of input frames, and the dilation rate of the dilated convolution increases exponentially with the increase of the number of network layers.

6. The method for rapid vehicle identification for ETC free-flow according to claim 1, characterized in that, Based on the fusion recognition result, a spatiotemporal consistency score is calculated, and the spatiotemporal consistency score is used as the credibility identifier of the fusion recognition result. The credibility identifier is then corrected using a temporal correlation coefficient, including: Obtain N vehicle model recognition results and N license plate recognition results corresponding to N consecutive frames of vehicle passing images; The proportion of the most frequent vehicle type in the N vehicle type recognition results is used as the vehicle type consistency score. Calculate the average string similarity between each pair of the N license plate recognition results, and use it as the license plate consistency score; The cross-modal matching quality index is normalized and used as a prior weight for data quality. The average of the vehicle model consistency score and the license plate consistency score is multiplied by the data quality prior weight to obtain the spatiotemporal consistency score, and this spatiotemporal consistency score is used as the credibility indicator of the current recognition result. When the time-series correlation coefficient is lower than the preset confidence threshold, the output value of the current confidence identifier is reduced proportionally. When the time-series correlation coefficient is higher than or equal to the preset confidence threshold, the original spatiotemporal consistency score is maintained as the confidence identifier output.

7. The method for rapid vehicle identification for ETC free-flow according to claim 6, characterized in that, The spatiotemporal consistency score also includes: Based on the scale changes of the vehicle bounding box in multiple consecutive frames of vehicle passing images, the instantaneous velocity sequence of the vehicle relative to the gantry is calculated. Verify whether the variance of the instantaneous velocity sequence exceeds a preset physical consistency threshold; If the variance of fluctuation exceeds the physical consistency threshold, a motion consistency penalty between 0 and 1 is generated, and the value of the penalty is negatively correlated with the variance of fluctuation. Multiplying the spatiotemporal consistency score by the motion consistency penalty term yields the credibility identifier after motion physics constraint correction.

8. The method for rapid vehicle identification for ETC free-flow according to claim 7, characterized in that, Also includes: When the credibility indicator is lower than the preset difficult case trigger threshold, the cross-modal matching quality index of the current vehicle is obtained; If the cross-modal matching quality index is greater than or equal to the preset high quality threshold, the multi-frame vehicle passing images of the current vehicle, vehicle identity information, and fusion recognition results are packaged into high-value edge hard case samples. If the cross-modal matching quality index is less than the preset high quality threshold, the current sample is marked as a low-quality noise sample and is either discarded directly or only retained for statistical information without participating in model updates.

9. The method for rapid vehicle identification for ETC free-flow according to claim 8, characterized in that, After obtaining high-value edge case samples, the following steps are included: Upload the high-value, difficult-to-produce sample to the cloud. After receiving the edge hard example samples, the cloud extracts the vehicle appearance features and license plate style features from them, and uses a conditional generative adversarial network to generate a synthetic training image with the same vehicle appearance and license plate style as the edge hard example samples. An incremental training dataset is constructed based on the synthetic training images. The lightweight temporal fusion recognition model is incrementally fine-tuned to generate an updated lightweight temporal fusion recognition model, which is then distributed to the edge computing node.

10. A vehicle rapid identification system for ETC free-flow, characterized in that, The method for rapid vehicle identification for ETC free-flow as described in any one of claims 1-9 includes: The information acquisition module is used to acquire images of the vehicle to be identified and to perform time-series alignment on the images of the vehicle to be identified. The feature sequence extraction module is used to extract multi-dimensional dynamic feature sequences of the same vehicle in continuous spatial positions. The multi-dimensional dynamic feature sequences include geometric contour change features of the vehicle body, illumination invariance features of the license plate area, and inter-frame offset features of local texture of the vehicle. The recognition model training module is used to input the multi-dimensional dynamic feature sequence and the vehicle identity information into a lightweight temporal fusion recognition model deployed on an edge computing node, and perform cross-modal correlation alignment to obtain a unified multi-modal temporal feature tensor. The lightweight temporal fusion recognition model outputs the fusion recognition result of the current vehicle, which includes vehicle model recognition result, license plate recognition result, and credibility identifier. The identification result output module is used to calculate the spatiotemporal consistency score based on the fusion identification result, and use the spatiotemporal consistency score as the credibility identifier of the fusion identification result. The credibility identifier is corrected using the temporal correlation coefficient, and the final vehicle identification result is output for ETC free-flow tolling decision.

Images