A road domain full-factor credible perception method and system based on uncertainty quantification

Abstract

The present invention proposes a method and system for credible perception of the entire road domain based on uncertainty quantification. First, a roadside lidar is used to collect 3D point cloud data in real traffic scenes to create a data set. Subsequently, in the "single frame" detection process, a deep evidence occupancy grid model is constructed to divide the roadside space into grids of uniform size, predict the probability of any grid space being occupied, and quantify the model's confidence in the output result, that is, whether the grid is occupied or idle, that is, the uncertainty. Subsequently, a detection frame generation algorithm can be used to generate a detection frame for the target, thereby improving the accuracy and robustness of target detection in easily occluded environments. Finally, a multi-target tracking algorithm is developed, and an uncertainty-based trajectory management module is integrated to adaptively adjust the timing of trajectory initialization and trajectory deletion according to the target detection results, effectively overcoming the problem of trajectory breakage or loss caused by factors such as random occlusion, while reducing the generation of ghost trajectories.

Description

Technical Field

[0001] The present invention relates to the field of roadside perception technology, and in particular to a method and system for credible perception of all road elements based on uncertainty quantification. Background Art

[0002] Compared to vehicle-side perception, roadside perception can achieve full target perception and output continuous trajectory data across a long-range, wide-angle road area. This supports applications such as vehicle-road fusion perception and active traffic control, improving the operational efficiency and safety of transportation systems. However, in complex roadside traffic environments, random occlusions between vehicles pose significant challenges to multi-target perception and tracking. This leads to various credibility issues in trajectory data, such as "broken tracks" and "ghost tracks," which seriously impact the effectiveness of system control.

[0003] During object detection, physical occlusion often results in incomplete point cloud data for the target object, rendering its key features invisible and undermining model accuracy. In particular, in LiDAR-based roadside perception systems, distant objects are often easily occluded by nearby obstacles due to the system's long longitudinal sensing range. Furthermore, the dense near-field and sparse far-field nature of 3D point clouds complicates feature extraction of distant objects under occlusion. Furthermore, existing 3D object detection algorithms are mostly based on end-to-end neural network models, whose performance is highly dependent on the training dataset and are inherently limited to identifying the types of objects included during training. Therefore, when these models are deployed in real-world applications and encounter point cloud data generated by random occlusions or unlabeled objects in the training dataset, they are prone to significant uncertainty, typically manifesting as false positives (FP) or false negatives (FN). In object tracking, the uncertainty of single-frame detection significantly impacts inter-frame tracking performance. Specifically, consecutive false negatives can cause track fragmentation or disappearance. Furthermore, consecutive false positives can produce ghost tracks, incorrectly representing non-existent objects. Furthermore, distant targets can be completely obscured by nearby obstacles, making them undetectable. The duration of such obscurations is affected by a variety of factors, such as sensor mounting angle and the relative positions of vehicles, and exhibits significant variability. Existing object tracking algorithms generally lack specific optimization for these roadside scenarios.

[0004] To sum up, how to reliably perceive and track all targets in complex roadside traffic environments is a difficult problem that urgently needs to be overcome with the rapid development of intelligent networking and vehicle-road collaboration. Summary of the Invention

[0005] In view of the defects of the existing technology in practical application, the present invention provides a method and system for road domain full-factor credible perception based on uncertainty quantification. Figure 1 As shown, it includes data set construction based on roadside lidar point cloud, full-factor road area detection based on uncertainty quantification, and continuous multi-target tracking considering perception uncertainty. The specific technical solution is as follows:

[0006] (1) Dataset construction based on roadside lidar point cloud

[0007] Based on the 3D point cloud data collected by a calibrated roadside lidar installed in the road space, the present invention performs data preprocessing, 3D point cloud annotation, and true value data generation to create a roadside point cloud dataset for model training.

[0008] 1) Data preprocessing:

[0009] First, the raw point cloud data packet is received and parsed to obtain a PCD format file. Second, the point cloud is framed, overlaying multiple point cloud data packets onto the same frame. Third, the extrinsic parameters are changed. Because the roadside LiDAR has a pitch angle with the ground, the roadside LiDAR coordinate system is uniformly converted to a virtual LiDAR coordinate system using an extrinsic matrix. This coordinate system is based on the geometric center of the LiDAR sensor as the origin, with the x-axis parallel to the ground and pointing forward, the y-axis parallel to the ground and pointing to the left, and the z-axis perpendicular to the ground and pointing upward, conforming to the right-hand coordinate system rule. Finally, filtering is performed to remove noise and smooth the data.

[0010] 2) 3D point cloud annotation:

[0011] Based on the collected roadside point cloud data, the 3D target frame in the road environment is manually annotated, and information such as target category, occlusion, truncation, and trajectory ID is recorded. The target frame annotated in this step is the true value data corresponding to the original point cloud (the result after manual calibration), which is used to train the neural network model in the second step of road domain full-element detection. The trained model can output a high-accuracy target detection frame based on the original point cloud data (the target detection frame is the output result of the model, which has not been manually calibrated and has uncertainty). An example of the point cloud annotation process is as follows: Figure 10 shown.

[0012] Target categories: There are 10 categories in total, including pedestrians, motor vehicles, etc., as shown in Table 1:

[0013] Table 1 Target categories and annotation information

[0014]

[0015]

[0016] 3) Ground truth data generation:

[0017] After the annotation is completed, the point cloud data is uniformly converted to the BEV (Bird's Eye View) perspective. Then, the target detection range is determined and divided into uniformly sized grids (0.16m×0.16m), as shown in the schematic diagram. Figure 12 As shown in , each grid should contain several point clouds; based on the ground truth data annotated by the 3D point cloud, the position of the target can be obtained, and based on its position, whether the grid is occupied by the target can be determined. The schematic diagram of the ground truth generation result is shown in Figure 11 As shown in the figure, white grids represent objects occupied by the target (because the target size is much larger than the grid size, the white grids appear connected), gray grids represent unoccupied grids, and black grids indicate that there is no point cloud data in the grid. The script was used to batch generate the training dataset for the deep evidence occupancy grid model.

[0018] (2) Road area full element detection based on uncertainty quantification

[0019] In the first part of the present invention, the roadside detection range is divided into 0.16m×0.16m grids. This part will model the grid state. The occupancy state of each grid at time k can be expressed by the binomial random variable o k ={O,F} representation, where O means the grid is occupied and F means the grid is idle.

[0020] After the Deep Evidence Occupancy Grid Model (DEOG) proposed in this paper is trained using the training dataset generated in the first part, it can input the original point cloud data in real time to predict the occupancy of any grid in real time, and obtain occupancy grid maps (OGMs). Subsequently, a detection box generation and uncertainty quantification algorithm is designed to further process the OGMs to output multi-target detection boxes with uncertainty parameters. The specific process is as follows: Figure 2 shown.

[0021] 1) Deep Evidence Occupancy Grid Model Construction

[0022] The DEOG model uses a neural network structure to analyze and extract key features from point cloud data. It also incorporates evidence theory to design a detection head and define a model loss function. The model is trained to gather evidence indicating whether a grid is occupied or free. This evidence refers to the support quantification extracted from point cloud features, which is used to infer the occupancy status of each grid. During the inference process, the more evidence the model collects, the lower the uncertainty level of the corresponding prediction. This section will detail the DEOG model building process and loss function design. The specific steps are as follows:

[0023] a. Construction of Deep Evidence Occupancy Grid Model DEOG

[0024] The DEOG model mainly consists of three parts, as shown in the following diagram: Figure 3 shown.

[0025] First, the feature encoder network creates a set of pillars with infinite spatial range in the z-axis direction perpendicular to the ground and converts the reflection points measured by the lidar into stacked pillar tensors and pillar index tensors. The encoder uses the stacked pillars to learn a set of dense features and generate a 2D pseudo image that can be input into a convolutional neural network.

[0026] Subsequently, a 2D convolutional backbone network consisting of two sub-networks is used. The first sub-network collects features of different spatial scales through a top-down structure, and the second sub-network upsamples and concatenates all features to process the 2D pseudo-image into a high-dimensional representation.

[0027] Finally, we designed the Evidential Detection Head (EDH), the core component of the DEOG model. This module combines Dempster-Shafer Theory (DST) with Subjective Logic (SL) to collect evidence pointing to the corresponding grid state during the inference process. The detection head outputs a prediction of the grid occupancy state and, based on the collected evidence, quantifies the model's confidence in the prediction. The detection head's functionality is achieved through the following two steps:

[0028] 1) Based on the DST theory, the belief mass is assigned to all possible power sets in the discriminant framework Θ = {O, F}. m(O) represents the belief quality of the grid being occupied, m(F) represents the belief quality of the grid being idle, and m(Θ) represents the belief quality distributed to the entire framework. At this point, any category will have equal probability, i.e., the uncertainty is maximized. It is represented as an empty set with a value of 0, and the sum of the belief masses of all subsets is 1.

[0029] 2) Use SL to formalize the belief quality distribution process of the above DST into a Dirichlet distribution to quantify the belief quality and uncertainty of each power set. The specific process is as follows:

[0030] SL defines belief quality as b A , the uncertainty of the grid is u. e A,A∈Θ and Θ={O,F}, represents the evidence corresponding to a single category (occupied or free), which is the feature that the neural network learns from the high-dimensional features output by the convolutional backbone network to divide the grid into free / occupied. A ≥0. b A and u can be calculated by the following formula:

[0031]

[0032] ∑ A∈Θ b A +u=1 (2)

[0033] α A =e A +1 (3)

[0034] p A =α A / S (4)

[0035] Among them, b A - belief quality, determined by the evidence e corresponding to A A Sure;

[0036] e A - Evidence supporting the partitioning of the states of the occupied grid into states A, corresponding to A∈Θ;

[0037] Strength of the S-Dirichlet distribution;

[0038] α A - parameters of the Dirichlet distribution;

[0039] p A - the expected probability that grid i is occupied or free, which is 0 or 1;

[0040] K - the number of categories, the median value in this scenario is 2;

[0041] u - perceived uncertainty of each grid;

[0042] The more evidence supporting grid occupancy or vacancy is collected during the model reasoning process, the lower the uncertainty of the prediction result. When the model does not obtain any evidence, the uncertainty reaches its maximum and u is 1. The belief quality distribution process is defined as parameter α A =e A +1 Dirichlet distribution. S = ∑ A∈Θ α A is the strength of the Dirichlet distribution. The Dirichlet distribution parameterized by evidence is used to describe the probability mass function p A The probability density distribution of possible values ​​(free / occupied) is as follows:

[0043]

[0044]

[0045]

[0046] Where α={α A |A∈Θ},S A is an A-1 dimensional simplex, and B(α) is an A-dimensional polynomial beta function.

[0047] The evidence detection head module mainly consists of a 2D convolutional layer with two output channels, using the ReLU activation function to ensure that the output Dirichlet distribution parameter α A The feature map output by the detection head has the same size as the previously divided occupancy grid, each pixel corresponds to a grid in OGMs, and has two channels, each channel contains evidence e extracted from the point cloud features to support whether the grid i is occupied or idle i,A ,A∈Θ。

[0048] The collected evidence is converted into the estimated parameters of the Dirichlet probability density function (i.e., Equation 5) through Equation 3, and then the expected probability of grid i being occupied or idle and the uncertainty u of the prediction result are obtained using Equations 4 and 1, respectively.

[0049] b. Loss function design, that is, using the Kullback-Leibler divergence expansion loss function as the loss function of the DEOG model, that is, Formula 10.

[0050] This step integrates the Dirichlet distribution based on the square sum loss structure and designs the loss function of the DEOG model. The parameters of the DEOG neural network model are w, given a grid i, its expected occupied / idle probability p iA ,A∈Θ and the truth value y iA ,A∈Θ,y iA One-hot encoding is performed according to the actual state of the grid. If the state of grid i is occupied, it is 1, and if it is idle or unknown, it is 0.

[0051] The specific loss function is as follows, according to the properties of variance Formula 11 can be further simplified to Formula 12.

[0052]

[0053]

[0054] During the training process, the DEOG model mines and generates evidence corresponding to occupancy or vacancy based on the high-dimensional features of the point cloud data. Considering that the network will inevitably generate evidence pointing to the wrong label during the training process. When the DEOG model correctly predicts the grid occupancy state, that is, the evidence collected pointing to the correct state is greater than the wrong evidence, in this case the misleading evidence will not affect the result. However, once the prediction is wrong, the evidence supporting the wrong result should be reduced to zero, that is, S = K. At this time, the Dirichlet distribution will become a uniform distribution, and the uncertainty is maximized, that is, u = 1. Therefore, the present invention uses the Kullback-Leibler divergence expansion loss function.

[0055] When the grid occupancy state is predicted incorrectly, the parameterized Dirichlet distribution of the model output approaches the uniform distribution, which reduces the total evidence to zero and maximizes the uncertainty of the prediction results.

[0056]

[0057] Among them, the annealing coefficient is λ t =min(1,t / 10), its value will increase from 0 to 1 as the iteration period t increases. represents the Dirichlet parameter after removing non-misleading evidence from grid i.

[0058] The KL divergence term in the loss function can be expressed as:

[0059]

[0060] 2) Target detection box generation and detection box uncertainty quantification

[0061] This step is based on the output of the DEOG model (evidence occupancy grid map), and takes steps such as grid group identification, detection box generation, and uncertainty quantification to output a multi-target detection box with perceptual uncertainty. Its main attributes are the center position (x c ,y c ), the width and length of the detection frame (w, l), the deflection angle θ of the detection frame along the y-axis, and the uncertainty value u of the detection frame c The specific process is as follows: Figure 4 shown.

[0062] a. Grid group identification

[0063] Based on the grid occupancy results, independent grid groups (grids predicted to be occupied in the evidence occupancy grid map) are determined. Each grid group represents an object, supporting the subsequent object detection and segmentation tasks. This is mainly divided into two traversals: the first traversal marks all occupied grids, discovers all connected areas and possible equivalent labels (union-find in Table 2); the second traversal corrects all labels to ensure that each grid in the same connected area has the same label. The grid group identification diagram is shown in the figure below. Figure 5 The specific implementation method is shown in Table 2.

[0064] Table 2 Connected domain statistics algorithm based on two traversals

[0065]

[0066] b. Detection box generation

[0067] Based on the independent grid groups of multiple targets obtained in the above steps, the present invention uses an oriented bounding box (OBB) algorithm to generate target detection boxes. The specific implementation method is shown in Table 3.

[0068] Table 3 Oriented bounding box algorithm based on independent grid groups

[0069]

[0070] c. Uncertainty Quantification

[0071] Based on the grid group identification and detection frame generation, the contour and posture information of each vehicle is obtained. It is also necessary to convert the perceptual uncertainty u of all grids contained in the detection frame into a unified scalar value, that is, the uncertainty u of the entire target detection frame c Since the accuracy and robustness of the results have spatial correlation within the sensing range of the roadside sensor, such as physical occlusion between vehicles and aging of some sensor areas, the present invention adopts a spatial mean filtering algorithm to calculate the uncertainty value u of the detection frame. c The algorithm diagram is as follows: Figure 6 The specific formula is as follows:

[0072]

[0073] The output of the filter g(x,y) is the product of the filter coefficient w and the grid f surrounded by the filter. To achieve the function of mean filtering, the filter weight w(s,t) is generally set to 1, and the center coefficient w(0,0) is aligned with the grid at position (x,y). x and y are variable so that each filter element in w can access each grid in f. For the grid covered by the same detection frame, g(x,y) is calculated once for each point, and then the g(x,y) of the covered grid is averaged to obtain the uncertainty value u of the detection frame. c .

[0074] (3) Multi-target continuous tracking considering perception uncertainty:

[0075] This step uses inter-frame displacement estimation and trajectory association to match the current frame target detection frame with the trajectory predicted based on the previous frame, and updates the successfully matched trajectory. For unmatched trajectories and detection frames, an uncertainty-based trajectory management module is designed to adaptively adjust the trajectory initialization and trajectory termination strategies. The specific process is as follows Figure 7 shown.

[0076] 1) Inter-frame displacement estimation and state update

[0077] The present invention is based on the Kalman filter algorithm, which models the current state of the target object and predicts its state at the next moment. The constant linear velocity model is used to approximate the inter-frame displacement of each target object. The state modeling is as follows:

[0078] S T =[x c ,y c ,w,l,θ,U,v x ,v y ] (13)

[0079] Among them, (x c ,y c ,w,l,θ,u c ) is the target detection box attribute of the current frame, U is the cumulative uncertainty of the trajectory, and its value is given by u c with U t-1 Joint decision, v x ,v y are the vehicle's lateral and longitudinal velocities, respectively.

[0080] When a detection box successfully matches a historical trajectory, the state space of the target trajectory is updated based on the matching target detection result. If no detection box matches the trajectory, its state is simply estimated using the linear velocity model without correction.

[0081] 2) Inter-frame trajectory association

[0082] This step calculates the matching cost matrix between the target detection box of the current frame and the estimated value through the constant linear velocity model (based on the prediction result of the trajectory of the previous frame). The cost is the intersection over union (IOU) of the detection value and the estimated value, and the formula is as follows:

[0083]

[0084] Here, X and Y represent the detected and estimated bounding boxes, respectively, and Square() represents the area of ​​the shape. Subsequently, the Hungarian algorithm is used to optimize the matching of the currently detected bounding boxes of all objects in the scene with the estimated bounding boxes based on the previous frame. The optimization goal is to maximize the sum of the Intersection over Union (IoU) of all objects. A higher IoU value indicates greater overlap between the two bounding boxes and better matching quality.

[0085] 3) Track status update

[0086] For successfully associated trajectories, the state space of each trajectory is updated based on the Bayesian rule using the current measurement value;

[0087] 4) Trajectory management based on uncertainty

[0088] When a roadside target enters or leaves the sensing range, a unique ID needs to be created or deleted accordingly. This step dynamically adjusts the target trajectory initialization and termination strategies through an uncertainty update function.

[0089] a. Uncertainty update function construction:

[0090] Define the cumulative uncertainty update principle. For continuously detected and successfully matched trajectories, the uncertainty update function will gradually reduce its uncertainty value, reflecting the increasing certainty of the existence of the trajectory.

[0091] For trajectories that fail to be successfully matched in consecutive frames, the uncertainty update function will increase its uncertainty value σ uncertainty , which indicates the increase in uncertainty about the trajectory. The specific formula is as follows:

[0092]

[0093] b. Trajectory initialization judgment:

[0094] Use piecewise function to construct trajectory initialization discriminant function, where Select a range for the adaptive parameter, such as T trial =[2,4,6,8] This value means continuous T trial The target's trajectory will be initialized only when the frame matching is successful.

[0095] In order to prevent the target detection algorithm from generating false positives and causing mismatches, the target is usually detected and matched continuously in multiple frames before starting and outputting the target trajectory. i (i=1,2,3) is the uncertainty classification threshold of the potential trajectory, which is calculated by formula 15, for example, U i =[0.2, 0.6, 0.8], the lower the uncertainty level of the potential trajectory, the greater the confidence of the model in the existence of the trajectory, and the fewer the number of consecutive matching frames required to start the trajectory.

[0096]

[0097] c. Trajectory termination judgment:

[0098] Construct a trajectory termination discriminant function, where Select a range for the adaptation parameters. The lower the uncertainty level, the longer the number of frames to keep mismatched trajectories before terminating.

[0099] The present invention provides a device and system for a road domain full-quantity and full-element credible perception method based on uncertainty quantification. Figure 8 This is a schematic diagram of the system of the present invention, including modules such as roadside lidar, edge computing unit, and roadside communication equipment:

[0100] 1) Roadside LiDAR: A sensor device installed in the road space that is used to obtain point cloud data of the roadside traffic environment and provide high-resolution three-dimensional spatial information;

[0101] 2) Edge computing unit: Preprocesses the raw point cloud data acquired by the roadside LiDAR, inputs the processed data into the trusted perception system deployed in the computing unit, and outputs a continuous trajectory;

[0102] 3) Roadside communication equipment: Ensure real-time data transmission between the trusted perception system and the control center or other intelligent devices, and transmit the trajectory data of the entire road area and all elements obtained by the edge computing unit to the control center.

[0103] The road domain full-quantity and full-element trusted perception system is composed of the above-mentioned devices, all of which are installed in the road space. By executing the road domain full-quantity and full-element trusted perception method based on uncertainty quantification proposed by the present invention, it processes the roadside point cloud data and outputs a continuous trajectory.

[0104] Compared with the existing technology, the trust perception method and system provided by the present invention has the following key technical points and advantages:

[0105] (1) The present invention takes into account that in actual roadside applications, dynamic, complex and uncertain traffic environments are prone to cause a large amount of random occlusions, making the key features of the target object incomplete or generating point cloud data that is offset from the distribution of the training set. This poses a great challenge to the existing 3D target detection algorithm based on the end-to-end neural network model. Therefore, the DEOG model proposed in the present invention focuses on the problem of reliable perception of irregular point cloud data caused by random occlusion in complex roadside scenes. It divides the roadside space into grids of uniform size and predicts the probability of any grid space being occupied, weakening the semantic type of the target and focusing on the reasoning of spatial occupancy relationships. The accuracy and robustness of target detection are greatly improved.

[0106] (2) This paper integrates DS evidence theory and subjective logic into the DEOG model, enabling the model to collect evidence pointing to grid occupancy or vacancy from the high-dimensional features of the 3D point cloud, and to quantify the uncertainty of the prediction results. Its advantages are mainly reflected in two aspects:

[0107] Enhanced uncertainty management. Traditional neural network models, when faced with target types not seen during training, may produce incorrect predictions, but often with unreasonably high confidence levels. This phenomenon can lead to misjudgments in practical applications, increasing systemic risk. However, by incorporating DS evidence theory, the DEOG model naturally maintains a high level of uncertainty in its predictions when insufficient evidence is available.

[0108] Improve the interpretability of model predictions. When the uncertainty of the DEOG model output increases, it directly reflects that the model has low confidence in its prediction results, providing the system with a self-assessment capability for downstream decision support and risk management.

[0109] (3) The present invention integrates an uncertainty-based trajectory management module in the multi-target tracking process. According to the uncertainty update function, the uncertainty level of the trajectory is updated in real time. The management strategy of trajectory initialization and termination is adaptively adjusted based on uncertainty, which can avoid the phenomenon of trajectory breakage or ghost trajectory to a certain extent. That is, the uncertainty-based trajectory management module is integrated to adaptively adjust the timing of trajectory initialization (entering the detection area and continuously detecting and tracking for several frames) and trajectory deletion (leaving the detection area for several frames to detect and track) according to the target detection results, effectively overcoming the problem of trajectory breakage or loss caused by factors such as random occlusion, while reducing the generation of ghost trajectories and effectively improving the continuity and robustness of the output trajectory. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1 This is a flow chart of the road domain full-factor trustworthy perception method based on uncertainty quantification;

[0111] Figure 2 This is a flowchart for detecting all road elements based on uncertainty quantification;

[0112] Figure 3 Schematic diagram of the deep evidence occupancy grid model structure;

[0113] Figure 4 Methods for target detection box generation and uncertainty quantification;

[0114] Figure 5 Schematic diagram of the connected domain statistics algorithm based on two traversals;

[0115] Figure 6 Schematic diagram of uncertainty spatial mean filtering;

[0116] Figure 7 Flowchart for continuous tracking of multiple targets considering perception uncertainty;

[0117] Figure 8 A schematic diagram of a road domain full-scale, full-element trusted perception device and system;

[0118] Figure 9 This is a schematic diagram of output results for a specific embodiment;

[0119] Figure 10 Schematic diagram for 3D point cloud annotation;

[0120] Figure 11 Grid division and ground truth target box annotation within the roadside detection range;

[0121] Figure 12 Schematic diagram of roadside sensing range and grid division (0.16m*0.16m). DETAILED DESCRIPTION

[0122] The present invention relates to a method and system for realizing credible perception of all road elements based on uncertainty quantification. The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0123] This embodiment is based on the roadside perception system of the Intelligent Connected Vehicle Demonstration Zone in Jiading District, Shanghai, and carries out 3D point cloud data collection, data set construction, and model training and deployment.

[0124] This embodiment involves multiple roadside lidars and corresponding roadside edge computing and communication units, which together constitute the road domain full-scale, full-element trusted perception system proposed in this invention. The specific implementation method includes the following steps:

[0125] S1: Construction of roadside lidar point cloud dataset

[0126] This step is based on point cloud data collected by calibrated roadside lidar installed in the intelligent connected vehicle demonstration area. After data preprocessing, 3D point cloud annotation, and ground truth data generation, a roadside point cloud dataset for model training is created.

[0127] The roadside point cloud dataset created in this embodiment will include scenes with different weather conditions (sunny, cloudy, rainy) and different traffic flows (low flow, medium flow, high flow), which meets the model training requirements of the present invention.

[0128] S2: Road area full element detection based on uncertainty quantification

[0129] This step designs and trains a Deep Evidence Occupancy Grid (DEOG) model based on the 3D point cloud dataset obtained in step S1. After training, the model inputs real-time raw 3D point cloud data, and outputs occupancy grid maps (OGMs) and uncertainty quantification results in real time. A bounding box generation algorithm is developed based on the OGMs, outputting bounding boxes for all road features with perceptual uncertainty.

[0130] The algorithm is deployed in the edge computing unit. Specifically:

[0131] S21: DEOG model structure design

[0132] The DEOG model uses a neural network structure, which is divided into three parts: the feature encoder network, the 2D convolution backbone, and the evidence detection head. The first is the feature encoder network, which encodes the reflection points measured by the roadside lidar into denser features. Secondly, the convolution backbone network converts low-dimensional features into high-dimensional representations. Finally, the evidence detection head is a 2D convolution layer with two output channels and ReLU activation. Each pixel represents a grid i in the predicted OGM, where the channel contains evidence e of two states in the recognition framework. A ,A∈Θ, that is, whether there is a vehicle in the grid.

[0133] The collected evidence can be converted into the estimated parameters of the Dirichlet probability density function through formula 3, and then the probability of grid occupation or idleness and the uncertainty of the result can be obtained through formula 4 and 1.

[0134] The loss function of the DEOG model is designed based on the sum of squares loss and the parameter α of the Dirichlet distribution. Assume that a grid i is given, the grid has parameters w, the expected occupied / idle probability p i and the truth value y i ,y i One-hot encoding is performed based on the grid's actual occupied / unoccupied state. Furthermore, when the model makes an incorrect prediction during training, the evidence supporting the incorrect result is reduced to zero. The model's loss function, L(w), is designed using Equations 10 and 11.

[0135] S22: DEOG model training

[0136] The training data used by the model is a roadside 3D point cloud dataset. The specific hyperparameter settings are shown in Table 4. The loss function can measure the difference between the model's predicted results and the actual results. In this embodiment, as the model training process continues to iterate, the loss function achieves good convergence.

[0137] Table 4 Model hyperparameter settings

[0138] parameter Parameter value Batch size 16 Epochs 100 Learning rate 1e-4 Grid size 0.16m Min point distance 3m Max points per pillar 100 Max pillars 10000 Number features 9 Number channels 64

[0139] S23: Target detection box generation and uncertainty quantification:

[0140] This step is based on the occupancy grid map that takes into account the uncertainty of perception and outputs the corresponding target detection box so that the subsequent tracking algorithm can generate a continuous trajectory. The properties of the target detection box are (x c ,y c ,w,l,θ,u c To obtain the target detection frames of all vehicles in the perception range, it is necessary to count the connected domains in the grid, identify the boundaries of the target detection frames, and quantify the uncertainty of the detection frames.

[0141] S231: Based on the connected domain statistics algorithm based on two traversals proposed in Table 2, all independent grid groups of single-frame OGMs are obtained.

[0142] S232: Based on the independent grid group obtained in the previous step, the OBB method (see Table 3) is used to calculate the size and direction of the bounding rectangle according to the geometric shape of the object itself, and the center position of the target (x c ,y c ), the width and length of the detection frame (w, l), and the deflection angle θ of the detection frame along the y-axis.

[0143] S233: Use the spatial mean filtering algorithm to calculate the uncertainty value u of the vehicle at the current moment c , for any point (x, y) in the independent grid group, the output of the filter g(x, y) is the product of the filter coefficient w(s, t) and the grid f surrounded by the filter. The specific formula is shown in Formula 12. Calculate the mean of g(x, y) of all points in the independent grid group to obtain the uncertainty u of the target detection box c

[0144] S24: The trained DEOG model will be deployed on the edge computing unit together with the detection box generation algorithm, which can perform real-time detection on the input roadside point cloud and output all the target detection boxes in the scene and the corresponding perception uncertainty.

[0145] Specifically, in the tth frame, the target detection module outputs the result as (n t is the number of detections, which can vary with the frame), each detection box Represented as a tuple (x c ,y c ,w,l,θ,u c ).

[0146] S3: Continuous Multi-Target Tracking Considering Perception Uncertainty

[0147] Receive the full-scale, full-element detection results of the road area from step S2 to conduct continuous multi-target tracking, and perform data association on multiple single-frame detection results (target detection boxes and corresponding perception uncertainties) to output a continuous trajectory. The algorithm is deployed in the edge computing unit. Specifically:

[0148] S31: The inter-frame displacement estimation module uses the Kalman filter to calculate all the trajectories to be associated in the t-1 frame. (Total m t-1 ) propagates to the next frame, named T estimation , obtained based on the constant velocity model, that is, x estimation =x+v x ,y estimation =y+v y .

[0149] For each trajectory in frame t-1 After propagating to the tth frame, the state of each target in each trajectory is predicted based on the inter-frame displacement: The predicted state is then sent to the trajectory association module.

[0150] S32: In order to convert the target detection result D of the tth frame t And the trajectory prediction result T of the t-1 frame estimation The data association module uses the Hungarian algorithm to use each pair of detection boxes and The intersection over union (IoU) between the two is calculated in dimension n. t ×m t-1 Affinity matrix.

[0151] Then, the Hungarian algorithm is used to solve the bipartite graph matching problem in polynomial time. min , reject the match.

[0152] The output of the data association module is the trajectory of the detection result at time t that successfully matches the predicted trajectory at time t-1 and updates the state. and mismatched trajectories And the unmatched target detection results where w t is the number of matches.

[0153] S33: Use the status update module to update the status of the successfully matched trajectory.

[0154] To consider T match The uncertainty in the target detection result D match , update T match The entire state space of each trajectory in , and obtain the associated trajectory of the t-th frame

[0155] S34: The uncertainty-based trajectory association module manages trajectory initialization and termination measurements.

[0156] After the matching is completed, the cumulative uncertainty of all trajectories is updated based on the uncertainty update function, that is, Formula 15. For the unmatched target detection box, the "trial period" T of the newly created trajectory is adaptively adjusted by the uncertainty value. trial ; For the trajectory that does not match the detection frame, the "temporary storage period" T of the retained trajectory is adaptively adjusted based on the accumulated uncertainty value storage The length of is shown in Formula 16.

[0157] In this embodiment, the specific parameters are selected as follows: The value is [2,2,3,3]; The value is [10,8,6,4]; U i (i=1,2,3) is the cumulative uncertainty classification threshold, and its value is [0.3,0.5,0.7].

[0158] Based on the above steps, the road domain full-factor trustworthy perception method and system constructed in this embodiment will process the roadside point cloud online and output a continuous trajectory. The results are as follows: Figure 9 shown.

Claims

1. A road domain full-factor credible perception method based on uncertainty quantification, characterized by: The following steps are involved: 1) Dataset construction based on roadside lidar point cloud: 3D point cloud data is collected by a calibrated roadside lidar installed on the road. Data preprocessing, 3D point cloud annotation, and ground truth data generation are performed in sequence to create a roadside point cloud dataset for model training. 2) Road area full element detection based on uncertainty quantification: 2.1) Deep Evidence Occupancy Grid Model Construction: A deep evidence occupancy grid model is constructed based on DS evidence theory and subjective logic. After preliminary training, the model can input roadside 3D point cloud data and output an occupancy grid map within the perception range and the uncertainty of the prediction; 2.2) Object Detection Box Generation and Uncertainty Quantification: Based on the output of the trained deep evidence occupancy grid model, grid group identification, detection box generation, and uncertainty quantification are performed to output multi-object detection boxes with perceptual uncertainty. 3) Multi-target continuous tracking considering perception uncertainty: 3.1) Inter-frame displacement estimation: A kinematic model and Kalman filter algorithm are used to model the historical trajectory of the target object and predict the current state based on the historical trajectory. 3.2) Inter-frame trajectory association: Calculate the current measurement value, that is, the matching cost matrix between the target detection box and the displacement estimate, and use the Hungarian algorithm to optimize the multi-target trajectory association; 3.3) Trajectory state update: For successfully associated trajectories, the state space of each trajectory is updated based on the Bayesian rule using the current measurement value. 3.4) Uncertainty-based trajectory management: The cumulative uncertainty of the trajectory is calculated based on the uncertainty update function, and the trajectory initialization and trajectory termination strategies are adaptively adjusted.

2. The method for credible perception of all road elements based on uncertainty quantification according to claim 1, characterized in that: The 3D point cloud data preprocessing includes roadside 3D point cloud data preprocessing: data reception and reading, point cloud framing, external parameter changes, and filtering processing procedures.

3. The method for credible perception of all road elements based on uncertainty quantification according to claim 1, characterized in that: The 3D point cloud annotation is based on the collected roadside point cloud data, annotating the 3D target box in the road environment, and annotating the target category, occlusion, truncation, and trajectory ID information.

4. The method for credible perception of all road elements based on uncertainty quantification according to claim 1, characterized in that: The true value data generation is to select the detection range and divide it into grids of uniform size, determine whether each grid is occupied by the target based on the point cloud annotation results, and use scripts to batch process and generate a point cloud dataset.

5. The method for credible perception of all road elements based on uncertainty quantification according to claim 1, characterized in that: The depth occupancy grid model construction includes the following steps: 2.1.1) Deep Evidence Occupancy Grid Model Construction: The model is constructed into three parts, including a feature encoder network, a 2D convolutional backbone, and an evidence detection head; 2.1.2) Loss function design: Based on the sum-of-squares loss structure, we integrate the Dirichlet distribution and design the loss function of the deep evidence occupancy grid model.

6. The method for credible perception of all road elements based on uncertainty quantification according to claim 1, characterized in that: The multi-target detection frame generation includes the following steps: 2.2.1) Grid group identification: Based on the occupancy grid map, a connected domain analysis algorithm is used to identify all independent grid groups within the sensing range and assign them unique IDs; 2.2.2) Detection box generation: Based on the independent grid group, a directional bounding box algorithm is used to determine the center point, corner point coordinates, and motion direction parameters of the detection box; 2.2.3) Uncertainty Quantification: A spatial mean filtering algorithm is used to convert the perceptual uncertainty of the grid contained in the detection box into a unified scalar value to obtain the perceptual uncertainty parameter of the detection box.

7. The method for credible perception of all road elements based on uncertainty quantification according to claim 1, characterized in that: The uncertainty-based trajectory management includes the following steps: 3.4.1) Uncertainty update function construction: The cumulative uncertainty update principle is defined according to the trajectory association state. That is, for a successfully matched trajectory, the minimum value of the trajectory uncertainty value at the previous moment and the detection box perception uncertainty parameter at the current moment is taken as the cumulative uncertainty value of the trajectory at the current moment; for an unsuccessful trajectory, the uncertainty factor σ is increased based on the trajectory uncertainty value at the previous moment. uncertainty , get the cumulative uncertainty value of the trajectory at the current moment; 3.4.2) Trajectory initialization judgment: Based on the accumulated uncertainty value, a trajectory initialization discriminant function is constructed. The higher the uncertainty level, the more consecutive matching frames are required to start the trajectory. 3.4.2) Trajectory termination judgment: Based on the accumulated uncertainty value, a trajectory termination discriminant function is constructed. The lower the uncertainty level, the longer the number of frames that the mismatched trajectory is retained before termination.

8. A road domain full-quantity and full-element credible perception system based on uncertainty quantification, used to implement the road domain full-quantity and full-element credible perception method based on uncertainty quantification according to any one of claims 1 to 7, characterized in that: include: 1) Roadside LiDAR: A sensor device installed in the road space that is used to obtain point cloud data of the roadside traffic environment and provide three-dimensional spatial information; 2) Edge computing unit: Preprocesses the raw point cloud data acquired by the roadside LiDAR, inputs the processed data into the trusted perception system deployed in the computing unit, and outputs a continuous trajectory; 3) Roadside communication equipment: ensures real-time data transmission between the system and the control center or other intelligent devices, and transmits the trajectory data of the entire road area and all elements obtained by the edge computing unit to the control center.

Images