Traffic intersection self-adaptive control traffic signal machine and traffic control method

Abstract

The invention relates to the technical field of traffic control, and discloses a traffic intersection adaptive control traffic signal machine and a traffic control method, and the method comprises the steps: obtaining local point cloud data of a to-be-turned vehicle through a multi-frame image collection and three-dimensional reconstruction technology; according to the geometrical characteristics of the vehicle local point cloud, dynamically distributing weight coefficients of shape completion and semantic completion, and fusing the two completion results to generate a high-precision vehicle global point cloud; and then the vehicle length value is calculated through bounding box fitting, and when the number of vehicles meeting the length condition reaches a threshold value, the system automatically prolongs the green light duration of the U-turn signal lamp. According to the scheme, the problem of size calculation errors caused by local missing of vehicle images in a traditional method is effectively solved, accurate recognition and passing demand matching of large turning-around vehicles (such as long-wheelbase trucks) are achieved, the intersection space-time resource utilization rate in a complex traffic scene is remarkably improved, and the overall passing efficiency and safety are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic control, in particular to a traffic intersection adaptive control traffic signal machine and a traffic control method. BACKGROUND

[0002] With the acceleration of urbanization and the rapid growth of motor vehicle ownership, the traffic efficiency and safety of traffic intersections have become increasingly prominent. Especially in complex traffic scenarios, such as the conflict between turning vehicles and straight vehicles, and the mixed traffic of non-motor vehicles and pedestrians, the traditional timing control traffic signal machine is difficult to dynamically adjust the signal timing according to the real-time traffic conditions, leading to unbalanced allocation of intersection resources, increased vehicle queue length, and aggravated traffic delay. According to statistics, traffic congestion caused by unreasonable signal timing accounts for more than 30%, and the turning demand of turning vehicles is often ignored due to the large turning radius and complex operation, further exacerbating the waste of time and space resources at the intersection.

[0003] In related technologies, due to factors such as camera angle limitation, vehicle occlusion, or lighting conditions, the collected vehicle image data sometimes has local missing (such as missing in the tail or head area), leading to insufficient vehicle model accuracy, making it difficult to accurately calculate vehicle size parameters, and thus unable to dynamically adjust the traffic signal according to the vehicle size information of the passing vehicles. SUMMARY

[0004] The main purpose of the present application is to provide a traffic intersection adaptive control traffic signal machine and a traffic control method, which aims to solve the technical problem that in the prior art, due to local occlusion, it is difficult to accurately calculate the vehicle size parameters, and thus unable to dynamically adjust the traffic signal.

[0005] To achieve the above purpose, in a first aspect, the present application provides a traffic intersection adaptive control traffic signal machine for dynamically adjusting traffic control signals, comprising:

[0006] An image acquisition module configured to acquire multiple frames of image data of a target traffic intersection vehicle to be turned;

[0007] A point cloud extraction module configured to perform three-dimensional reconstruction on the image data, identify and extract vehicle local point cloud data;

[0008] A completion strategy generation module configured to determine the weight coefficients of shape completion and semantic completion according to the geometric features of the vehicle local point cloud, and generate a target completion strategy according to the weight coefficients of shape completion and semantic completion;

[0009] A point cloud completion module configured to perform shape completion and semantic completion operations according to the target completion strategy, and fuse the two completion results of shape completion and semantic completion to generate a vehicle global point cloud;

[0010] a size analysis module configured to perform a bounding box fitting on the vehicle global point cloud, calculate a size parameter of the fitted bounding box to obtain a vehicle length value;

[0011] a signal control module configured to generate a target traffic control signal for prolonging a green light passing time length of a U-turn signal lamp in a case where a target vehicle quantity determined according to the vehicle length value is greater than or equal to a quantity threshold.

[0012] In a possible implementation, the image data is three-dimensionally reconstructed, and vehicle local point cloud data is identified and extracted, including:

[0013] a number of key features of the vehicle in each frame of image data and a missing degree of the vehicle in each frame of image data are calculated;

[0014] a cost constraint is constructed according to the number of key features and the missing degree of the vehicle, wherein the cost constraint satisfies a cost function as follows: + (1- ) * , is an adaptive weight coefficient based on ambient light, is the number of key features of the vehicle in each frame of image data, is a maximum number of key features in the multi-frame image data, is the missing degree of the vehicle in each frame of image data, is a normalized adjustment factor;

[0015] a cost value corresponding to each frame of image data is calculated, and an image frame with a minimum cost value is determined as a target image frame;

[0016] the target image frame is three-dimensionally reconstructed, and vehicle local point cloud data is identified and extracted.

[0017] In a possible implementation, a weight coefficient of shape completion and semantic completion is determined according to a geometric feature of the vehicle local point cloud, including:

[0018] a plurality of geometric features of the vehicle local point cloud are extracted, wherein the geometric features include a surface curvature feature, an edge feature, and a symmetry feature;

[0019] the plurality of geometric features are valued and normalized to construct a geometric feature vector;

[0020] the geometric feature vector is input into a pre-trained weight decision network to output a shape completion weight coefficient α and a semantic completion weight coefficient β, wherein α+β=1.

[0021] In a possible implementation, according to the target completion strategy, shape completion and semantic completion operations are respectively performed, and shape completion and semantic completion results are fused to generate a vehicle global point cloud, including:

[0022] The vehicle local point cloud is voxelized to generate a three-dimensional voxel grid representation;

[0023] The three-dimensional voxel grid is input into a pre-trained three-dimensional encoder-decoder network to output a high-density point cloud prediction;

[0024] A Poisson surface reconstruction algorithm is used to generate a shape completion result to obtain a first vehicle global point cloud;

[0025] The vehicle local point cloud is input into a point cloud feature recognition and classification network to output a vehicle type prediction probability distribution P = {P1, P2, P3};

[0026] According to the vehicle type prediction probability distribution, a corresponding type of vehicle model set T = {T1, T2, T3} is retrieved from a semantic prior library;

[0027] The feature matching degree of each vehicle model and the local point cloud is calculated, and the model with the highest matching degree is selected to obtain a target model;

[0028] A deformation field estimation algorithm is used to deform the target model into a semantic completion result that is best aligned with the local point cloud to obtain a second vehicle global point cloud;

[0029] According to the weight coefficients α and β in the target completion strategy, the first vehicle global point cloud and the second vehicle global point cloud are fused to obtain a vehicle global point cloud.

[0030] In a possible implementation, according to the target completion strategy, shape completion and semantic completion operations are respectively performed, and shape completion and semantic completion results are fused to generate a vehicle global point cloud, including:

[0031] The first vehicle global point cloud and the second vehicle global point cloud are aligned to the same coordinate system and the same number of target points are collected;

[0032] The target points (X n1 ,Y n1 ,Z n1 ) in the first vehicle global point cloud are fused with the corresponding target points (X n2 ,Y n2 ,Z n2 ) in the second vehicle global point cloud to obtain fused point coordinates (X n1 *α+X n2 *β, Y n1 *α+Yn2 *β, Z n1 *α+Z n2 *β);

[0033] According to the fused point coordinates (X n1 *α+X n2 *β, Y n1 *α+Y n2 *β, Z n1 *α+Z n2 *β) form a candidate vehicle global point cloud;

[0034] The candidate vehicle global point cloud is subjected to contour smoothing to obtain a vehicle global point cloud.

[0035] In a possible implementation, the first vehicle global point cloud and the second vehicle global point cloud are aligned to the same coordinate system, and the same number of target points are collected, comprising:

[0036] The first vehicle global point cloud is subjected to geometric analysis to generate a first geometric centerline representing the length direction of the vehicle;

[0037] The second vehicle global point cloud is subjected to geometric analysis to generate a second geometric centerline representing the length direction of the vehicle;

[0038] A first reference point is selected at the edge position of the first vehicle global point cloud, and a second reference point is selected at the same edge position of the second vehicle global point cloud;

[0039] The Euclidean distance from the first reference point to the first geometric centerline is calculated to obtain a first distance, and the Euclidean distance from the second reference point to the second geometric centerline is calculated to obtain a second distance;

[0040] The ratio of the first distance to the second distance is calculated to obtain a reference ratio;

[0041] A density-based clustering algorithm is performed on the first vehicle global point cloud to generate M target point sets Q;

[0042] The second vehicle global point cloud is subjected to clustering collection of the same number of target points to generate M target point sets Q', wherein the ratio of the corresponding points in the set Q and the set Q' to the corresponding axis is the reference ratio.

[0043] In a possible implementation, the vehicle global point cloud is subjected to bounding box fitting, and the size parameters of the fitted bounding box are calculated to obtain a vehicle length value, comprising:

[0044] The vehicle global point cloud is subjected to pose alignment in the length direction to obtain a target vehicle global point cloud;

[0045] The target vehicle global point cloud is fitted with a minimum outer bounding box.

[0046] The size of the minimum outer bounding box in the length direction is taken as the vehicle length value.

[0047] In a possible implementation, in a case where the target vehicle quantity determined according to the vehicle length value is greater than or equal to the quantity threshold, the target traffic control signal is generated, including:

[0048] In a case where the target vehicle quantity determined according to the vehicle length value is greater than or equal to the quantity threshold, the ratio of the green light duration and the red light duration of the U-turn signal is generated according to the target vehicle quantity, wherein the target vehicle quantity is positively correlated with the green light duration of the U-turn signal.

[0049] In a possible implementation, after the target traffic control signal is generated, the method further includes:

[0050] Obtaining weather condition information of the current traffic intersection, and determining a weather influence coefficient according to the weather condition information;

[0051] The generated ratio of the duration is corrected according to the weather influence coefficient, so as to dynamically adjust the green light duration of the U-turn signal.

[0052] In a second aspect, the application further provides a traffic control method applied to a traffic signal machine, and the method includes:

[0053] Obtaining multiple frames of image data of a vehicle to be U-turned at a target traffic intersection;

[0054] Performing three-dimensional reconstruction on the image data, identifying and extracting local point cloud data of the vehicle;

[0055] Determining a weight coefficient of shape completion and semantic completion according to geometric features of the local point cloud of the vehicle, and generating a target completion strategy according to the weight coefficient of shape completion and semantic completion;

[0056] According to the target completion strategy, performing shape completion and semantic completion operations respectively, and fusing two completion results of shape completion and semantic completion to generate global point cloud of the vehicle;

[0057] Fitting a bounding box to the global point cloud of the vehicle, calculating three-dimensional size parameters of the fitted bounding box to obtain a vehicle length value;

[0058] In a case where the target vehicle quantity determined according to the vehicle length value is greater than or equal to the quantity threshold, a target traffic control signal is generated, and the target traffic control signal is used to prolong the green light duration of the U-turn signal.

[0059] Different from the prior art, the traffic intersection adaptive control traffic signal machine provided by the embodiment of the application first acquires local point cloud data of a vehicle to be turned around through multi-frame image acquisition and three-dimensional reconstruction technology, dynamically allocates weight coefficients of shape completion and semantic completion according to geometric features of the local point cloud of the vehicle, and fuses two completion results to generate high-precision global point cloud of the vehicle; then the length value of the vehicle is calculated through boundary box fitting, and when the number of vehicles meeting the length condition reaches a threshold value, the system automatically prolongs the green light duration of the turning signal light. The scheme effectively solves the size calculation error problem caused by local missing of the vehicle image in the traditional method, realizes accurate identification of large turning vehicles (such as long-wheelbase trucks) and matching of traffic demand, significantly improves the utilization rate of intersection space-time resources in a complex traffic scene, reduces the secondary queuing or intersection lock caused by insufficient turning radius, and finally improves the overall traffic efficiency and safety. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings shown.

[0061] Figure 1 The module structure schematic diagram of the traffic signal machine in some embodiments of the present application;

[0062] Figure 2 The flowchart of the traffic control method in some embodiments of the present application;

[0063] Figure 3 The schematic diagram of the point cloud fusion process in some embodiments of the present application.

[0064] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the drawings. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0066] It should be noted that all the direction indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, motion condition, etc. between the components in a certain specific posture (as shown in the drawings), and if the specific posture changes, the direction indications will also change accordingly.

[0067] In addition, the descriptions involving "first", "second" and the like in the present application are only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" can explicitly or implicitly include at least one of the features. In addition, "and / or" throughout the text includes three schemes, taking A and / or B as an example, including A technical solution, B technical solution, and A and B simultaneously satisfying the technical solution. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist, nor within the protection scope required by the present application.

[0068] The current traffic monitoring system mainly uses a camera to obtain two-dimensional image data, and realizes dynamic control of traffic signals through image analysis technologies such as target detection and trajectory tracking. However, the traditional scheme needs to process a large amount of pixel data, which puts high requirements on the computing power of edge computing devices. In order to solve the high computing power consumption problem of traditional image analysis, the present application converts two-dimensional images into three-dimensional point cloud data, and uses the sparsity (data volume is only 1 / 50-1 / 100 of the original image) and geometric feature advantages of point cloud data to realize the computing power optimization of traffic state analysis.

[0069] The inventive concept of the present application is: first, the local point cloud data of the vehicle to be turned around is obtained through multi-frame image acquisition and three-dimensional reconstruction technology, the weight coefficients of shape completion and semantic completion are dynamically allocated according to the geometric features of the vehicle local point cloud, and high-precision vehicle global point cloud is generated by fusing the two completion results; then the vehicle length value is calculated through the boundary box fitting, and when the number of vehicles meeting the length condition reaches the threshold value, the system automatically prolongs the green light duration of the turning signal light. This scheme effectively solves the size calculation error problem caused by the local missing of vehicle images in the traditional method, realizes the accurate identification of large turning vehicles (such as long-wheelbase trucks) and the matching of traffic demand, and improves the overall traffic efficiency and safety.

[0070] In this way, the existing camera equipment is used to collect original data, and relevant algorithms are used to perform three-dimensional reconstruction and other processing operations on image data, which effectively solves the computing power optimization problem of traffic state analysis without the need to replace the collection equipment to reduce costs.

[0071] For example, as shown in FIG. 1, the system includes a camera 101, a server 102, and a traffic signal 103. Figure 1As shown, the traffic signal machine 100 of the present application includes an image acquisition module 110, a point cloud extraction module 120, a completion strategy generation module 130, a point cloud completion module 140, a size analysis module 150, and a signal control module 160.

[0072] The image acquisition module 110 is configured to acquire / receive multi-frame image data of a target vehicle at a turning intersection; the completeness of vehicle information in the multi-frame image data can be different, for example, the multi-frame image data can include vehicle A (complete vehicle body information), or can include vehicle B (incomplete vehicle body information, with partial missing), etc. The multi-frame image data can also include vehicle A (completeness of 60%), or can include vehicle A (completeness of 80%) and the like.

[0073] The point cloud extraction module 120 is configured to perform three-dimensional reconstruction on the image data, identify and extract local point cloud data of the vehicle; more specifically, the vehicle part in the image data is three-dimensionally reconstructed to form point cloud data representing the vehicle. For vehicles with complete vehicle information or no occlusion, image information can be used for vehicle length analysis; and for vehicles with occlusion or incomplete vehicle information, local point cloud data of the visible part is extracted by the point cloud extraction module 120, providing a structured input basis for subsequent point cloud completion.

[0074] The completion strategy generation module 130 is configured to determine the weight coefficients of shape completion and semantic completion according to the geometric features of the vehicle local point cloud, and generate a target completion strategy according to the weight coefficients of shape completion and semantic completion.

[0075] The point cloud completion module 140 is configured to perform shape completion and semantic completion operations according to the target completion strategy, and fuse the two completion results of shape completion and semantic completion to generate a global point cloud of the vehicle.

[0076] The size analysis module 150 is configured to fit a bounding box to the global point cloud of the vehicle, calculate the size parameters of the fitted bounding box, and obtain the length value of the vehicle.

[0077] The signal control module 160 is configured to generate a target traffic control signal for prolonging the green light passing time length of the U-turn signal lamp in the case where the target vehicle quantity determined according to the vehicle length value is greater than or equal to the quantity threshold. More specifically, in a statistical time length period, if the number of large vehicles (such as the vehicle body length greater than 5m) for U-turn is large and greater than the set quantity threshold, it indicates that the current U-turn vehicle flow has a passing efficiency bottleneck (the turning radius requirement increases due to the long vehicle body, and the U-turn time is prolonged), and the number of vehicles that can complete U-turn passing per unit time is smaller. At this time, the target traffic control signal is generated to prolong the green light passing time length of the U-turn signal lamp, so as to improve the dispersing capacity of the U-turn vehicle flow at the intersection by increasing the effective passing time.

[0078] As shown in Figures 1-3 The following will be described by taking an example of an intersection signal machine executing a traffic control method. It should be noted that although a logical sequence is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here. Please refer to the accompanying Figure 2 The method includes the following steps S100-S600:

[0079] Step S100, acquiring multiple frames of image data of vehicles to be U-turned at a target intersection;

[0080] Specifically, the multiple frames of image data containing vehicle information can be obtained by continuous shooting of a camera device. The multiple frames of image data can include vehicle information of the same vehicle at different moments and in different states, or vehicle information of different vehicles at the same moment or at different moments, etc.

[0081] For example, the embodiment of the present application sets a visual sensor on the U-turn lane to acquire multiple frames of image data of vehicles to be U-turned at a target intersection.

[0082] Step S200, performing three-dimensional reconstruction on the image data to identify and extract vehicle local point cloud data;

[0083] The three-dimensional reconstruction on the image data can be three-dimensional reconstruction on all images (multiple frames), or three-dimensional reconstruction on a specific image (single frame).

[0084] In an embodiment, the three-dimensional reconstruction on the image data to identify and extract vehicle local point cloud data includes:

[0085] Calculating the number of key features of the vehicle in each frame of image data and the vehicle missing degree in each frame of image data;

[0086] Constructing a cost constraint according to the number of key features and the vehicle missing degree, wherein the cost constraint satisfies the following cost function: +(1- )* , These are adaptive weighting coefficients based on ambient lighting. The number of key features of the vehicle in each frame of the image. The maximum number of key features in a multi-frame image dataset. The vehicle missingness in each frame of image data. This is the normalization adjustment factor;

[0087] Calculate the cost value corresponding to each frame of image data, and determine the image data with the smallest cost value as the target image frame;

[0088] The target image frame is reconstructed in three dimensions to identify and extract local point cloud data of the vehicle.

[0089] Specifically, key features can be features such as headlights, contour corners, and key contour lines. The more key features there are, the richer the "description" of the vehicle in that frame. Vehicle missingness refers to the proportion of vehicles missing. This proportion can be estimated through image analysis. The smaller the vehicle missing proportion, the richer the "description" of the vehicle in that frame. The coefficient is affected by ambient lighting. For example, in good lighting, image features are clear, and the system relies more on the "number of key features" for judgment; in poor lighting, it relies more on the "missingness" for judgment. That is, in good lighting... When the value is greater than 0.5, and the illumination is poor, Less than 0.5, under normal lighting conditions. It equals 0.5. In this way, the influence of the two indicators can be flexibly balanced according to the lighting conditions. The adjustment factor is designed to normalize the calculation results within a reasonable range, and can be adjusted as needed for different scenarios.

[0090] According to the cost function above, The closer , The smaller the value, the higher the "feature richness". The smaller, The smaller the value, the better the "vehicle integrity". Based on dynamic adjustments to the environment such as lighting, the image data with the lowest cost is determined as the target image frame. In other words, the image frame with "complete features and a complete vehicle" is ultimately selected as the target image frame for 3D reconstruction. Compared to a scheme that performs 3D reconstruction on all images (multiple frames), this effectively reduces the amount of data processing.

[0091] Step S300, determining a weight coefficient of shape completion and semantic completion according to the geometric features of the vehicle local point cloud, and generating a target completion strategy according to the weight coefficient of shape completion and semantic completion;

[0092] The vehicle local point cloud refers to a local point cloud obtained by performing three-dimensional reconstruction on the best image data (the target image frame obtained in step S200) of a vehicle with missing vehicle information. The geometric features of the vehicle local point cloud can include vehicle edge contour features, surface smoothness features, etc.

[0093] In an embodiment, determining the weight coefficient of shape completion and semantic completion according to the geometric features of the vehicle local point cloud comprises:

[0094] extracting a plurality of geometric features of the vehicle local point cloud, wherein the geometric features include surface curvature features, edge features, and symmetry features;

[0095] performing value assignment and normalization processing on the plurality of geometric features to construct a geometric feature vector;

[0096] inputting the geometric feature vector into a pre-trained weight decision network to output a shape completion weight coefficient a and a semantic completion weight coefficient β, wherein a+β=1.

[0097] Shape completion is based on geometric features (such as curvature and edge) to complete the missing part through mathematical methods (such as Poisson reconstruction). Semantic completion is based on prior knowledge (such as “a car has four wheels”) to complete the missing part through deep learning models (such as PointNet).

[0098] The surface curvature feature describes the bending degree of the point cloud surface (such as low plane curvature and high wheel curvature), the edge feature is to identify the point cloud boundary (such as the car body contour line), and the symmetry feature is to reflect the symmetry characteristics of the point cloud (such as the left-right symmetry of the vehicle).

[0099] It can be understood that the richer the surface curvature is, that is, the more the number of different curvatures is, the more detailed it is, and the shape completion can be more effective, at which time the assignment to the surface curvature feature is higher. The more the edge feature is, that is, the richer the contour line of the vehicle body is (such as the longer the edge contour line is), the more obvious the edge feature of the vehicle is, and the shape completion can be more effective, at which time the assignment to the edge feature is higher. The more obvious the symmetry feature is, the more effective the semantic completion can be, at which time the assignment to the symmetry feature is higher. After obtaining the assignments of the respective geometric features, the respective assignments are normalized to scale the respective assignments to [0, 1], and a geometric feature vector is constructed, for example, the geometric feature vector is (N1, N2, N3), wherein N1 is the normalized data of the surface curvature feature, N2 is the normalized data of the edge feature, and N3 is the normalized data of the symmetry feature. Finally, the geometric feature vector is input into the pre-trained weight decision network to output the shape completion weight coefficient a and the semantic completion weight coefficient β.

[0100] After obtaining the shape completion weight coefficient a and the semantic completion weight coefficient β, the target completion strategy is formed. Specifically, the target completion strategy determines the target proportion of the shape completion and the semantic completion in the overall completion task through the matching relationship of a and β.

[0101] Therefore, the greater the value of N1 is, the greater the weight coefficient a of the shape completion is; the greater the value of N2 is, the greater the weight coefficient a of the shape completion is; and the greater the value of N3 is, the greater the weight coefficient β of the semantic completion is.

[0102] It should be noted that the construction of the weight decision network can be realized by first building a basic model framework and then training and optimizing parameters through multi-scene geometric feature data. Specifically, first, a preliminary network structure is built as a basic model, and then geometric feature data covering different scenes (such as various vehicle local point cloud geometric feature samples) are collected, and the preliminary model is trained using these data, and the network parameters are iteratively optimized to finally obtain a weight decision network with stable decision-making ability.

[0103] In step S400, shape completion and semantic completion operations are respectively performed according to the target completion strategy, and vehicle global point cloud is generated by fusing the two completion results of shape completion and semantic completion.

[0104] After obtaining the target completion strategy in step S300, shape completion and semantic completion operations can be respectively performed according to the target completion strategy, and vehicle global point cloud is generated by fusing the two completion results of shape completion and semantic completion.

[0105] In an embodiment, according to the target completion strategy, shape completion and semantic completion operations are respectively performed, and shape completion and semantic completion results are fused to generate a vehicle global point cloud, including:

[0106] The vehicle local point cloud is voxel grid processed to generate a three-dimensional voxel grid representation;

[0107] The three-dimensional voxel grid is input into a pre-trained three-dimensional encoder-decoder network to output a high-density point cloud prediction;

[0108] A shape completion result is generated by using a Poisson surface reconstruction algorithm to obtain a first vehicle global point cloud;

[0109] The vehicle local point cloud is input into a point cloud feature recognition and classification network to output a vehicle type prediction probability distribution P = {P1, P2, P3};

[0110] According to the vehicle type prediction probability distribution, a corresponding type of vehicle model set T = {T1, T2, T3} is retrieved from a semantic prior library;

[0111] The feature matching degree of each vehicle model and the local point cloud is calculated, and the model with the highest matching degree is selected to obtain a target model;

[0112] A deformation field estimation algorithm is used to deform the target model into a semantic completion result that is best aligned with the local point cloud to obtain a second vehicle global point cloud;

[0113] According to the weight coefficients a and b in the target completion strategy, the first vehicle global point cloud and the second vehicle global point cloud are fused to obtain a vehicle global point cloud.

[0114] Specifically, in the shape completion process, first, the input vehicle local point cloud is converted into a three-dimensional voxel grid (Voxel Grid), and the data sparsity is reduced through spatial discretization while the key geometric structure (such as the vehicle body contour and the wheel position) is preserved. Then the voxel grid is input into a pre-trained three-dimensional encoder-decoder network (such as PointNet++), and the network learns the geometric distribution law of a large number of vehicle point clouds to output a high-density point cloud prediction result, filling the local missing area (such as the obscured tail or head). Finally, a Poisson surface reconstruction algorithm (Poisson Surface Reconstruction) is used to convert the high-density point cloud into a continuous surface to generate a first vehicle global point cloud after shape completion.

[0115] In the semantic completion process, first, the vehicle local point cloud is input into a point cloud feature recognition classification network (such as PointCNN), and a vehicle type prediction probability distribution P={P1, P2, P3} (such as a car, an SUV, and a truck) is output. Then, according to the prediction probability distribution P, a vehicle model set T={T1, T2, T3} (such as a car model library, an SUV model library, and a truck model library) of the corresponding type is retrieved from a semantic prior library. Then, the feature matching degree (similarity) of each candidate model and the vehicle local point cloud is calculated, and the model with the highest matching degree is selected as the target model. Then, a deformation field estimation algorithm is used to deform the target model into a semantic completion result that is best aligned with the vehicle local point cloud, generating a second vehicle global point cloud. After obtaining the first vehicle global point cloud (shape completion result) and the second vehicle global point cloud (semantic completion result), according to the weight coefficients α and β in the target completion strategy, the first vehicle global point cloud and the second vehicle global point cloud are fused to obtain a vehicle global point cloud.

[0116] In an embodiment, according to the weight coefficients α and β in the target completion strategy, the first vehicle global point cloud and the second vehicle global point cloud are fused to obtain a vehicle global point cloud, comprising:

[0117] Aligning the first vehicle global point cloud and the second vehicle global point cloud to the same coordinate system and collecting the same number of target points;

[0118] Weighted fusion of the target points (X n1 ,Y n1 ,Z n1 ) in the first vehicle global point cloud and the corresponding target points (X n2 ,Y n2 ,Z n2 ) in the second vehicle global point cloud to obtain fused point coordinates (X n1 *α+X n2 *β, Y n1 *α+Y n2 *β, Z n1 *α+Z n2 *β);

[0119] Forming a candidate vehicle global point cloud according to the fused point coordinates (X n1 *α+X n2 *β, Y n1 *α+Y n2 *β, Z n1 *α+Z n2 *β);

[0120] Performing contour smoothing on the candidate vehicle global point cloud to obtain a vehicle global point cloud.

[0121] Specifically, the first vehicle global point cloud and the second vehicle global point cloud can be aligned into the same coordinate system, and a plurality of key target points (such as key points of the vehicle body contour) in the first vehicle global point cloud and a plurality of key target points in the second vehicle global point cloud are weighted and summed to obtain a fusion coordinate of the key points. After obtaining the fusion coordinates of all target key points, a candidate vehicle global point cloud is formed according to the fused target key point coordinates (X n1 *α+X n2 *β,Y n1 *α+Y n2 *β,Z n1 *α+Z n2 *β).

[0122] Thus, the application realizes the complementary advantages of multi-source point clouds through weighted fusion and smoothing processing, and generates a high-quality vehicle global point cloud. The application applies the weight coefficient of the completion strategy to the point cloud fusion process, so that the fusion result can take into account the completion advantages of geometric details and prior semantics, and provides a more reliable data basis for subsequent vehicle size analysis and traffic signal control.

[0123] It can be understood that when performing the point cloud weighted fusion operation, it is necessary to ensure that each key point in the first vehicle global point cloud and the corresponding key point in the second vehicle global point cloud are accurately mapped based on semantic consistency. For example, the key point representing the vehicle lamp in the first vehicle global point cloud needs to correspond to the key point representing the vehicle lamp in the second vehicle global point cloud. This semantic-based key point accurate correspondence mechanism effectively avoids the common part misplacement problem (such as the erroneous fusion of the vehicle head in the first vehicle global point cloud and the vehicle tail in the second vehicle global point cloud) in traditional point cloud fusion.

[0124] To improve the accuracy of point cloud fusion, in an embodiment, the first vehicle global point cloud and the second vehicle global point cloud are aligned into the same coordinate system, and the same number of target points are collected, including:

[0125] performing geometric analysis on the first vehicle global point cloud to generate a first geometric center line representing the length direction of the vehicle;

[0126] performing geometric analysis on the second vehicle global point cloud to generate a second geometric center line representing the length direction of the vehicle;

[0127] selecting a first reference point at the edge position of the first vehicle global point cloud, and selecting a second reference point at the same edge position of the second vehicle global point cloud;

[0128] calculating the Euclidean distance from the first reference point to the first geometric center line to obtain a first distance, and calculating the Euclidean distance from the second reference point to the second geometric center line to obtain a second distance;

[0129] a ratio of the first distance and the second distance is calculated as a reference ratio;

[0130] performing a density-based clustering algorithm on the first vehicle global point cloud to generate M target point sets Q;

[0131] performing the same number of target point clustering on the second vehicle global point cloud to generate M target point sets Q', wherein the ratio of the corresponding points in the set Q and the set Q' to the corresponding axis is the reference ratio.

[0132] Specifically, as shown in Figure 3 , first, a geometric analysis is performed on the first vehicle global point cloud to extract a first geometric center line L1 representing the length direction of the vehicle; the same analysis is performed on the second vehicle global point cloud to extract a second geometric center line L2. It can be understood that L1 is the center axis in the first global point cloud along the length direction of the vehicle, and L2 is the center axis in the second global point cloud along the length direction of the vehicle. A first reference point S1 (such as the upper left corner point of the vehicle) is selected at the edge position of the first global point cloud, and a second reference point S2 (consistent with the position of S1) is selected at the corresponding edge position of the second global point cloud. The Euclidean distance from S1 to L1 is calculated to obtain a first distance D1; the Euclidean distance from S2 to L2 is calculated to obtain a second distance D2. The reference ratio D1 / D2 is calculated based on D1 and D2. Subsequently, a density-based clustering algorithm (such as DBSCAN) is performed on the first global point cloud to generate M target point sets Q = {Q1, Q2, …, Q M}, ensuring that the target point set Q covers the entire key area of the vehicle appearance surface; the same clustering operation is performed on the second global point cloud to generate M target point sets Q' = {Q1', Q2', …, Q M '}. In the clustering process, the distance ratio of the corresponding points in Q and Q' to the geometric center line is forced to be equal to the reference ratio D1 / D2, i.e.

[0133] .

[0134] In this way, by constraining the distance ratio of the geometric center line and the reference point, the first vehicle global point cloud and the second vehicle global point cloud are accurately aligned in the length direction, eliminating the spatial offset caused by the difference in the completion strategy. It ensures that the geometric features (such as the vehicle body contour and component position) in the two global point clouds remain consistent in local and global scales, improving the accuracy of subsequent size analysis.

[0135] Step S500, performing a bounding box fitting on the vehicle global point cloud, calculating the three-dimensional size parameters of the fitted bounding box to obtain a vehicle length value;

[0136] After the point cloud fusion is performed to obtain the final vehicle global point cloud, a bounding box fitting can be performed on the vehicle global point cloud, and the three-dimensional size parameters of the fitted bounding box are calculated to obtain the vehicle length value.

[0137] The vehicle length value obtained above refers to the vehicle length value of all partially missing vehicles that turn around at a certain traffic intersection. The vehicle length of a vehicle that is not partially missing (information complete) can be obtained through image analysis, without the need for subsequent three-dimensional reconstruction, completion, and the like.

[0138] In an embodiment, the bounding box fitting is performed on the vehicle global point cloud, and the size parameters of the fitted bounding box are calculated to obtain the vehicle length value, including:

[0139] The vehicle global point cloud is subjected to pose alignment in the length direction to obtain a target vehicle global point cloud;

[0140] The target vehicle global point cloud is fitted with a minimum bounding box;

[0141] The size of the minimum bounding box in the length direction is taken as the vehicle length value.

[0142] Specifically, first, spatial rotation alignment is performed on the vehicle global point cloud, the vehicle length direction is adjusted to be perpendicular to the current observation angle through coordinate transformation, and a target vehicle global point cloud is generated; then the target vehicle global point cloud is fitted with a minimum bounding box, and finally the size of the minimum bounding box in the length direction is taken as the vehicle length value. In this way, the length value of the vehicle can be accurately obtained.

[0143] In step S600, in a case where the target vehicle quantity determined according to the vehicle length value is greater than or equal to the quantity threshold, a target traffic control signal is generated, and the target traffic control signal is used to prolong the green light passing time length of the U-turn signal light.

[0144] After the global point cloud processing of the target vehicles at the traffic intersection is completed, the cumulative number of long vehicle bodies (such as trucks, buses, and the like with a body length exceeding a threshold of 5 meters) can be counted based on the extracted vehicle length values. When the number of long vehicles in the counting period reaches or exceeds a preset quantity threshold, the system determines that there is a passing efficiency bottleneck in the current U-turn vehicle flow. Because the long vehicle body requires a larger turning radius and a longer turning time, the number of vehicles completing the U-turn per unit time is reduced, thereby causing the risk of intersection congestion. At this time, the target control signal of prolonging the green light time length of the U-turn is generated by dynamically adjusting the signal light timing, for example, the green light passing proportion of the U-turn lane is prolonged, so as to increase the effective passing time of the long vehicle U-turn, and finally improve the overall dispersion efficiency of the intersection U-turn vehicle flow.

[0145] In an embodiment, after the target traffic control signal is generated, the method further includes:

[0146] acquire weather condition information of the current traffic intersection, determine a weather influence coefficient according to the weather condition information;

[0147] correct the generated duration ratio according to the weather influence coefficient, to dynamically adjust the turning green light duration.

[0148] Specifically, after generating the target control signal of extending the turning signal green light duration, the system further integrates a real-time weather data interface to acquire the weather condition information (such as rainfall intensity, visibility, snow thickness, etc.) of the current traffic intersection. Based on a predefined weather-traffic efficiency mapping model, the influence coefficient of weather factors on vehicle turning speed is calculated (for example, the influence coefficient is 1.3 in heavy rain, indicating that the turning time is increased by 30%). Then, the initial generated green light duration extension ratio is dynamically corrected according to the influence coefficient, to generate the final adjusted turning green light duration, ensuring that the signal timing is accurately matched with the traffic capacity under actual weather conditions.

[0149] Based on this, the traffic intersection adaptive control traffic signal machine provided by the embodiments of the present application first acquires the local point cloud data of the vehicle to be turned by multi-frame image acquisition and three-dimensional reconstruction technology, dynamically allocates the weight coefficients of shape completion and semantic completion according to the geometric features of the vehicle local point cloud, and fuses the two completion results to generate high-precision vehicle global point cloud. Then, the vehicle length value is calculated by boundary box fitting, and when the number of vehicles meeting the length condition reaches the threshold value, the system automatically extends the green light duration of the turning signal. This scheme effectively solves the size calculation error problem caused by local missing of vehicle images in traditional methods, realizes accurate identification of large turning vehicles (such as long-wheelbase trucks) and matching of traffic demand, significantly improves the utilization rate of intersection space-time resources in complex traffic scenarios, reduces the secondary queuing or intersection lock caused by insufficient turning radius, and finally improves the overall traffic efficiency and safety.

[0150] The traffic signal machine provided by the embodiments of the present application further includes a memory and a processor, wherein the memory is used to store computer readable instructions, and the processor is used to call the computer readable instructions to execute the traffic control method as described above.

[0151] The processor is configured to provide computing and control capabilities to control the traffic signal to perform corresponding tasks, for example, to control the traffic signal to perform the traffic control method in any of the above method embodiments. The method includes: acquiring a plurality of image data of vehicles to be turned at a target intersection; performing three-dimensional reconstruction on the image data, identifying and extracting local point cloud data of the vehicles; determining weight coefficients of shape completion and semantic completion according to geometric features of the local point cloud data of the vehicles, and generating a target completion strategy according to the weight coefficients of shape completion and semantic completion; performing shape completion and semantic completion operations according to the target completion strategy, and fusing the two completion results of shape completion and semantic completion to generate global point cloud data of the vehicles; fitting a bounding box to the global point cloud data of the vehicles, calculating three-dimensional size parameters of the fitted bounding box to obtain a vehicle length value; and in a case where a target vehicle quantity determined according to the vehicle length value is greater than or equal to a quantity threshold, generating a target traffic control signal, the target traffic control signal being used to prolong a green light passing time length of a turning signal light.

[0152] The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), a hardware chip, or any combination thereof. It can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The above-mentioned PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.

[0153] The memory is a non-transitory computer readable storage medium, which can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the traffic control method in the embodiments of the present application. The processor can implement the traffic control method in any of the above method embodiments by running the non-transitory software programs, instructions and modules stored in the memory.

[0154] In particular, the memory can include volatile memory (VM), such as random access memory (RAM); the memory can also include non-volatile memory (NVM), such as read-only memory (ROM), flash memory, a hard disk drive (HDD), or a solid-state drive (SSD) or other non-transitory solid-state memory device; the memory can also include a combination of the above kinds of memory.

[0155] In summary, the traffic signal of the present application adopts the technical solutions of any one of the above traffic control method embodiments, and therefore has at least the beneficial effects brought by the technical solutions of the above embodiments, which will not be repeated here.

[0156] The embodiments of the present application also provide a computer readable storage medium, such as a memory including program codes, which can be executed by a processor to complete the traffic control method in the above embodiments. For example, the computer readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD ROM), a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0157] The embodiments of the present application also provide a computer program product, which includes one or more program codes stored in a computer readable storage medium. The processor of the early warning system reads the program codes from the computer readable storage medium, and the processor executes the program codes to complete the steps of the traffic control method provided in the above embodiments.

[0158] Those of ordinary skill in the art can understand that all or part of the steps of the above embodiments can be completed by hardware, or by program codes related to hardware, which can be stored in a computer readable storage medium, such as a read-only memory, a magnetic disk or an optical disk, etc.

[0159] It should be noted that the apparatus embodiments described above are merely illustrative, and the units described as separate units can or can not be physically separate, and the units displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0160] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be implemented by means of software plus a general hardware platform, and of course can also be implemented by hardware. Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiment methods. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0161] The above description is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation made according to the inventive concept of the present application, and direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A traffic intersection adaptive control traffic signal machine for dynamically adjusting traffic control signals, characterized by, The method comprises the steps of: An image acquisition module configured to obtain multiple frames of image data of a vehicle to be turned around at a target intersection; A point cloud extraction module configured to perform three-dimensional reconstruction on the image data, identify and extract local point cloud data of the vehicle; A completion strategy generation module configured to determine the weight coefficients of shape completion and semantic completion according to the geometric features of the local point cloud of the vehicle, and generate a target completion strategy according to the weight coefficients of shape completion and semantic completion; A point cloud completion module configured to perform shape completion and semantic completion operations respectively according to the target completion strategy, and fuse the two completion results of shape completion and semantic completion to generate a global point cloud of the vehicle; A size analysis module configured to fit a bounding box to the global point cloud of the vehicle, calculate the size parameters of the fitted bounding box to obtain a length value of the vehicle; A signal control module configured to generate a target traffic control signal if the number of target vehicles determined according to the length value of the vehicle is greater than or equal to a threshold value, wherein the target traffic control signal is used to prolong the green light duration of the turning signal.

2. The traffic adaptive control signal at a traffic intersection as claimed in claim 1, wherein, The three-dimensional reconstruction of the image data, the identification and extraction of the local point cloud data of the vehicle, comprise: Calculating the number of key features of the vehicle in each frame of image data and the missing degree of the vehicle in each frame of image data; According to the number of key features and the vehicle missing degree, a cost constraint is constructed, wherein the cost constraint satisfies a cost function as follows: + (1- ) * , is an adaptive weight coefficient based on ambient light, is the number of key features of the vehicle in each frame of image, is the maximum number of key features in the multi-frame image data, is the vehicle missing degree in each frame of image data, is a normalized adjustment factor; Calculating the generation cost corresponding to each frame of image data, and determining the image frame with the minimum generation cost as the target image frame; Performing three-dimensional reconstruction on the target image frame, identifying and extracting local point cloud data of the vehicle.

3. The adaptive traffic signal controller for traffic intersections of claim 1, wherein, Determining the weight coefficients of shape completion and semantic completion according to the geometric features of the local point cloud of the vehicle, comprises: Extracting multiple geometric features of the local point cloud of the vehicle, wherein the geometric features include surface curvature features, edge features and symmetry features; Value assignment and normalization processing are performed on the multiple geometric features to construct a geometric feature vector; Input the geometric feature vector into a pre-trained weight decision network to output a shape completion weight coefficient α and a semantic completion weight coefficient β, wherein α+β=1.

4. The adaptive traffic control signal for traffic intersections of claim 1, wherein, According to the target completion strategy, shape completion and semantic completion operations are performed respectively, and the two completion results of shape completion and semantic completion are fused to generate a global point cloud of the vehicle, comprising: Performing voxel grid processing on the local point cloud of the vehicle to generate a three-dimensional voxel grid representation; Input the three-dimensional voxel grid into a pre-trained three-dimensional encoder-decoder network to output a high-density point cloud prediction; Using a Poisson surface reconstruction algorithm, a shape completion result is generated to obtain a first global point cloud of the vehicle; Input the local point cloud of the vehicle into a point cloud feature recognition and classification network to output a vehicle type prediction probability distribution P={P1, P2, P3}; According to the vehicle type prediction probability distribution, a vehicle model set T={T1, T2, T3} of the corresponding type is retrieved from a semantic priori library; Calculate the feature matching degree of each vehicle model and the local point cloud, and select the model with the highest matching degree to obtain a target model; Using a deformation field estimation algorithm, the target model is deformed into a semantic completion result that is best aligned with the local point cloud to obtain a second global point cloud of the vehicle; According to the weight coefficients α and β in the target completion strategy, the first vehicle global point cloud and the second vehicle global point cloud are fused to obtain a vehicle global point cloud.

5. The adaptive traffic signal controller for traffic intersections of claim 4, wherein, According to the weight coefficients α and β in the target completion strategy, the first vehicle global point cloud and the second vehicle global point cloud are fused to obtain a vehicle global point cloud, comprising: Aligning the first vehicle global point cloud and the second vehicle global point cloud to the same coordinate system and collecting the same number of target points; target points (X n1 ,Y n1 ,Z n1 ) in the first vehicle global point cloud and corresponding target points (X n2 ,Y n2 ,Z n2 ) in the second vehicle global point cloud are fused by weighting to obtain fused point coordinates (X n1 *α+X n2 *β, Y n1 *α+Y n2 *β, Z n1 *α+Z n2 *β); According to the fused point coordinates (X n1 *α+X n2 *β, Y n1 *α+Y n2 *β, Z n1 *α+Z n2 *β) to form a candidate vehicle global point cloud; Performing contour smoothing processing on the candidate vehicle global point cloud to obtain a vehicle global point cloud.

6. The traffic adaptive control signal at a traffic intersection as claimed in claim 5, wherein, Aligning the first vehicle global point cloud and the second vehicle global point cloud to the same coordinate system and collecting the same number of target points, comprising: Performing geometric analysis on the first vehicle global point cloud to generate a first geometric centerline representing the length direction of the vehicle; Performing geometric analysis on the second vehicle global point cloud to generate a second geometric centerline representing the length direction of the vehicle; Selecting a first reference point at the edge position of the first vehicle global point cloud and a second reference point at the same edge position of the second vehicle global point cloud; Calculating the Euclidean distance from the first reference point to the first geometric centerline to obtain a first distance, and calculating the Euclidean distance from the second reference point to the second geometric centerline to obtain a second distance; Calculating the ratio of the first distance and the second distance to obtain a reference ratio; Performing a density-based clustering algorithm on the first vehicle global point cloud to generate M target point sets Q; Collecting the same number of target points from the second vehicle global point cloud to generate M target point sets Q', wherein the ratio of the corresponding points in sets Q and Q' to the corresponding axis is the reference ratio.

7. The adaptive traffic control signal for traffic intersections of claim 1, wherein, Fitting a bounding box to the vehicle global point cloud to calculate the size parameters of the fitted bounding box to obtain a vehicle length value, comprising: Aligning the vehicle global point cloud in the length direction to obtain a target vehicle global point cloud; Fitting a minimum outer bounding box to the target vehicle global point cloud; Taking the size of the minimum outer bounding box in the length direction as the vehicle length value.

8. The adaptive traffic control signal for traffic intersections of claim 1, wherein, In the case where the number of target vehicles determined according to the vehicle length value is greater than or equal to the number threshold, generating a target traffic control signal, comprising: In the case where the number of target vehicles determined according to the vehicle length value is greater than or equal to the number threshold, generating a time length ratio of the U-turn green light and the red light according to the number of target vehicles, wherein the number of target vehicles is positively correlated with the U-turn green light time length.

9. The adaptive traffic control traffic signal of claim 8, wherein, After generating the target traffic control signal, further comprising: Obtaining weather condition information of the current traffic intersection, and determining a weather influence coefficient according to the weather condition information; According to the weather influence coefficient, the generated time length ratio is corrected to dynamically adjust the U-turn green light passing time length.

10. A traffic control method applied to a traffic signal, characterized by, The method comprises: Obtaining multiple frames of image data of vehicles to be U-turned at a target traffic intersection; Performing three-dimensional reconstruction on the image data, identifying and extracting vehicle local point cloud data; Determining the weight coefficients of shape completion and semantic completion according to the geometric features of the vehicle local point cloud, and generating a target completion strategy according to the weight coefficients of shape completion and semantic completion; According to the target completion strategy, shape completion and semantic completion operations are respectively performed, and shape completion and semantic completion results are fused to generate a vehicle global point cloud; A bounding box fitting is performed on the vehicle global point cloud, three-dimensional size parameters of the fitted bounding box are calculated, and a vehicle length value is obtained; In a case where a target vehicle quantity determined according to the vehicle length value is greater than or equal to a quantity threshold, a target traffic control signal is generated, and the target traffic control signal is used to prolong a green light passing time length of a U-turn signal light.

Images