Vehicle lane changing risk quantification method based on surrounding vehicle trajectory

Abstract

The invention discloses a surrounding vehicle trajectory-based vehicle lane changing risk quantification method, which comprises the following steps: S1, a multi-source sensing and trajectory extraction step: acquiring motion state data and historical trajectory sequences of a vehicle and surrounding vehicles in real time through a multi-source sensor system of the vehicle, wherein the sensing range covers vehicle track records within the range of 200 meters in front of the vehicle and vehicle motion state sensing within the range of 80 meters around the vehicle; and S2, a surrounding vehicle track deviation characteristic calculation step: calculating a track deviation angle sequential sequence for the vehicle in front of the current lane, the vehicle in front of the target lane, the vehicle behind the target lane and the vehicle in the adjacent lane, defining a track deviation angle theta as an included angle between the instantaneous movement direction of the vehicle and the advancing direction of the center line of the lane where the vehicle is located, and calculating through the following formula: theta = arctan (v < y > / v < x >). The method has the advantages of high risk identification perspectiveness, high intention judgment accuracy and good multi-vehicle risk fusion effect.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and vehicle safety control technology, specifically to a method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories. Background Technology

[0002] Lane changing is a core interaction scenario in the road traffic system and a key risk source that can trigger side collisions, rear-end collisions, and other accidents. With the development of intelligent transportation systems and autonomous driving technology, higher demands are being placed on the accurate quantification and early warning of dynamic risks during lane changing.

[0003] Currently, existing technologies mainly assess lane-changing risks through environmental perception and model prediction. For example, Chinese invention patent CN110085056A proposes a lane-changing risk identification method based on gridded space and kinetic energy loss. This method calculates the potential conflict between the vehicle and surrounding vehicles during lane changing by establishing a vehicle motion trajectory prediction model and classifies the risk level based on kinetic energy loss. However, this scheme is essentially a post-event risk assessment model. The gridded discrete space analysis method it relies on is difficult to capture the continuous dynamic characteristics of the vehicle trajectory in real time, especially its ability to identify the lateral deviation intention of surrounding vehicles in the early stage is insufficient. In addition, this method uses kinetic energy loss as a severity indicator. Although it can reflect the consequences of the collision, it cannot predict the behavior intention of surrounding vehicles before the conflict occurs, resulting in a delay in the warning time and failing to meet the needs of active safety systems for forward-looking risk perception.

[0004] Another Chinese invention patent, CN112863245A, proposes a real-time prediction method for vehicle lane-changing trajectories based on deep neural networks. It identifies driving states through a hidden Markov model and uses a long short-term memory network to predict lane-changing trajectories. This solution improves the accuracy of trajectory prediction, but its technical focus is on predicting the lane-changing trajectory of the vehicle itself. It does not involve the identification of the lane-changing intentions of surrounding vehicles and the dynamic risk assessment of the vehicle's own lane-changing intentions. In other words, this solution lacks a quantitative mechanism for the risk of surrounding vehicle behavior and cannot solve the problem of chain risk response caused by the proactive entry or avoidance of surrounding vehicles in multi-vehicle interaction scenarios.

[0005] In summary, while existing technologies each have their own focus in trajectory prediction or risk assessment, none of them have achieved full-chain risk quantification, from extracting dynamic features of vehicle trajectories to early identification of lane-changing intentions, and then to the fusion of risks in multi-vehicle interactions. This technological gap makes it difficult for existing systems to provide timely and reliable lane-changing decision support in real and complex traffic scenarios, and also limits the efficiency and safety of advanced autonomous driving systems in dense traffic environments.

[0006] Therefore, there is an urgent need for a new technical solution that can capture the trajectory deviation characteristics of surrounding vehicles at the perception level, identify their lane-changing intentions at the prediction level, and achieve dynamic fusion of multi-source risks at the decision-making level. Summary of the Invention

[0007] The purpose of this invention is to provide a vehicle lane-changing risk quantification method based on weekly vehicle trajectories, which has the advantages of strong forward-looking risk identification, high accuracy of intent judgment and good multi-vehicle risk fusion effect, and solves the problems in the prior art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories includes the following steps:

[0010] S1. Multi-source perception and trajectory extraction steps: The vehicle's multi-source sensor system acquires the motion status data and historical trajectory sequences of the vehicle and surrounding vehicles in real time. The perception range covers the trajectory records of vehicles within 200 meters in front of the vehicle and the motion status perception of vehicles within 80 meters around the vehicle.

[0011] S2. Calculation steps for vehicle trajectory deviation characteristics: For the vehicle in front in this lane, the vehicle in front in the target lane, the vehicle behind in the target lane, and vehicles in adjacent lanes, calculate their trajectory deviation angle time series. The trajectory deviation angle θ is defined as the angle between the instantaneous direction of the vehicle's movement and the direction of travel of the centerline of its lane, and is calculated using the following formula:

[0012] θ = arctan(v_y / v_x);

[0013] Where v_x is the vehicle's velocity component in the longitudinal direction of the lane, and v_y is the vehicle's velocity component in the lateral direction of the lane;

[0014] S3, the following steps are taken to identify the intention of the vehicle to change lanes: Based on the time sequence of the trajectory offset angle, the probability of the vehicle to change lanes is calculated by a pre-trained intention recognition model. The intention recognition model adopts a feedforward neural network structure. The input is the trajectory offset angle sequence of the current time and the previous N time. The output is the probability of the vehicle to change lanes to the left, maintain lane or change lanes to the right in the next T time. The value of T is 1-3 seconds.

[0015] S4. Multi-vehicle interaction risk dynamic fusion steps: Construct a dynamic risk graph model with the current vehicle as the root node, and include surrounding vehicles with non-zero lane-changing intention probability as risk nodes in the model. Calculate the spatiotemporal conflict risk value based on the current vehicle's preset lane-changing path and the predicted trajectory of each risk surrounding vehicle within a future time T (the same as the value of T in step S3). Then, weight and fuse the risk values ​​according to the lane-changing intention probability to obtain a comprehensive risk index.

[0016] S5. Risk Warning and Decision Recommendation Generation Steps: Compare the comprehensive risk indicators with dynamically adjusted thresholds to determine the risk level, and generate decision recommendations for the vehicle based on the risk level and the probability of the vehicle's intention to change routes.

[0017] Preferably, in step S2, the calculation of the trajectory offset angle further integrates visual perception data, obtains the lateral offset d_lat of the target vehicle relative to the lane line through the vehicle-mounted camera, and calibrates the trajectory offset angle using the following formula:

[0018] θ_calibrated=γ×θ+(1-γ)×arctan(d_lat / L)

[0019] Where γ is the fusion weighting coefficient, with a value ranging from 0.6 to 0.8, and L is the predetermined observation distance benchmark, with a value of 50 meters.

[0020] Preferably, in step S3, the intention recognition model is trained using a historical trajectory dataset, including a sequence of vehicle trajectory offset angles and corresponding lane-changing behavior labels. The network parameters are optimized through supervised learning, the intention probability output is calculated through forward propagation, and the cross-entropy loss function is minimized through backpropagation.

[0021] Preferably, in step S4, the calculation of the spatiotemporal conflict risk value specifically includes: predicting the trajectories of the current vehicle and the risky vehicle within the next time period T, calculating the minimum distance D_min between the two trajectories in the spatiotemporal domain and the time TTC to reach the minimum distance, and the spatiotemporal conflict risk value R_conflict is calculated using the following formula:

[0022] R_conflict=(P_intent_left+P_intent_right)×exp(-λ×D_min) / TTC

[0023] Where λ is the distance sensitivity coefficient, with a value ranging from 0.1 to 0.3, and P_intent_left and P_intent_right are the probabilities of the surrounding vehicle's intention to change lanes to the left and right, respectively.

[0024] Preferably, in step S4, the comprehensive risk index is calculated by weighted summation of the spatiotemporal conflict risk values ​​of all risk cycles. The comprehensive risk index R_combined is calculated using the following formula:

[0025] R_combined=Σ_j[w_j×R_conflict_j]

[0026] Where j is the index of the risky weekly vehicle, and w_j is the risk weight of the j-th weekly vehicle.

[0027] Preferably, the formula for calculating the risk weight w_j is:

[0028] w_j=α×(1 / dist_j)+β×|Δv_j|

[0029] Where α and β are weighting coefficients, with α ranging from 0.6 to 0.8 and β ranging from 0.2 to 0.4. dist_j is the longitudinal relative distance between the vehicle and the j-th surrounding vehicle in the lane (in meters), and Δv_j is the longitudinal relative speed between the vehicle and the j-th surrounding vehicle (in m / s).

[0030] Preferably, in step S5, the dynamically adjusted threshold is determined based on the real-time traffic scenario by querying a predefined scenario-threshold mapping table; the scenario-threshold mapping table is generated through training on historical risk event data, wherein: the threshold range for unobstructed flow (traffic flow density < 10 vehicles / km / lane, average speed > 60km / h) is 0.3-0.4, and the threshold range for stable flow (10 vehicles / km / lane ≤ traffic flow density < 25 vehicles / km / lane, 40km / h < average speed ≤ 60km / h) is... The threshold range is 0.4-0.5. The threshold range for saturated flow (25 vehicles / km / lane ≤ traffic flow density < 40 vehicles / km / lane, 20km / h < average speed ≤ 40km / h) is 0.5-0.6. The threshold range for congested flow (traffic flow density ≥ 40 vehicles / km / lane, average speed ≤ 20km / h) is 0.6-0.7. Real-time traffic scenarios are divided into four types based on the traffic flow density and average speed of the road where the vehicle is located: smooth flow, stable flow, saturated flow, and congested flow.

[0031] Preferably, in step S1, the multi-source sensor system includes a camera, millimeter-wave radar, and lidar. The multi-sensor data is fused using a Kalman filter algorithm, including noise reduction, coordinate alignment, and time synchronization.

[0032] Preferably, in step S2, the sampling frequency of the trajectory offset angle time sequence is 10Hz, the length of the historical trajectory sequence covers at least 5 seconds, and the value of N is 10.

[0033] Preferably, in step S4, the update frequency of the dynamic risk graph model is 10Hz, and the node risk and edge weight are recalculated based on the latest sensor data at each time step.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention achieves early identification of lane-changing intentions by surrounding vehicles through real-time calculation and analysis of trajectory offset angles. Specifically, by calculating the time sequence of trajectory offset angles of vehicles in front in the current lane, vehicles in front and behind in the target lane, and vehicles in adjacent lanes, it can effectively capture subtle changes in the lateral movement of vehicles. When a surrounding vehicle intends to change lanes, its trajectory offset angle will exhibit a specific change pattern. This method identifies these patterns through a feedforward neural network, accurately predicting the intention 1-3 seconds before the lane-changing behavior occurs, thus solving the problem of delayed warnings in traditional methods.

[0036] 2. This invention employs multi-source perception fusion and a dynamic risk graph model, significantly improving the accuracy of risk assessment in multi-vehicle interaction scenarios. Through data fusion from cameras, millimeter-wave radar, and lidar, combined with trajectory offset angle features obtained from visual calibration, a dynamic risk graph with the vehicle itself as the root node is constructed. This model treats surrounding vehicles with lane-changing intentions as risk nodes. Based on the spatiotemporal conflict risk formula R_conflict=(P_intent_left+P_intent_right)×exp(-λ×D_min) / TTC, the conflict degree of each node is quantified, and a comprehensive risk index is obtained through weighted summation, effectively solving the assessment problem of multi-vehicle risk coupling.

[0037] 3. The adaptive threshold mechanism based on traffic flow status ensures the applicability of the system in different scenarios. Through a predefined scenario-threshold mapping table, the risk threshold is dynamically adjusted according to real-time traffic density and vehicle speed, so that the system can maintain optimal performance under different traffic conditions such as smooth flow, stable flow, saturated flow and congested flow, avoiding false alarms or missed alarms caused by fixed thresholds. Attached Figure Description

[0038] Figure 1 The diagram shown is an overall flowchart of the method of the present invention;

[0039] Figure 2 The flowchart shown is for the trajectory offset angle calculation and intent recognition of the present invention;

[0040] Figure 3 The diagram shown is a schematic representation of the dynamic risk diagram model of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] To address the issues of delayed risk warnings, lack of vehicle intent recognition, and insufficient consideration of multi-vehicle interaction risks in existing technologies, the following technical solution is proposed. Please refer to [link / reference needed]. Figures 1-3 ;

[0043] A method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories includes the following steps:

[0044] S1. Multi-source perception and trajectory extraction steps: The vehicle's multi-source sensor system acquires the motion status data and historical trajectory sequences of the vehicle and surrounding vehicles in real time. The perception range covers the trajectory records of vehicles within 200 meters in front of the vehicle and the motion status perception of vehicles within 80 meters around the vehicle.

[0045] S2. Calculation steps for vehicle trajectory deviation characteristics: For the vehicle in front in this lane, the vehicle in front in the target lane, the vehicle behind in the target lane, and vehicles in adjacent lanes, calculate their trajectory deviation angle time series. The trajectory deviation angle θ is defined as the angle between the instantaneous direction of the vehicle's movement and the direction of travel of the centerline of its lane, and is calculated using the following formula:

[0046] θ = arctan(v_y / v_x);

[0047] Where v_x is the vehicle's velocity component in the longitudinal direction of the lane, and v_y is the vehicle's velocity component in the lateral direction of the lane;

[0048] S3, the following steps are taken to identify the intention of the vehicle to change lanes: Based on the time sequence of the trajectory offset angle, the probability of the vehicle to change lanes is calculated by a pre-trained intention recognition model. The intention recognition model adopts a feedforward neural network structure. The input is the trajectory offset angle sequence of the current time and the previous N time. The output is the probability of the vehicle to change lanes to the left, maintain lane or change lanes to the right in the next T time. The value of T is 1-3 seconds.

[0049] S4. Multi-vehicle interaction risk dynamic fusion steps: Construct a dynamic risk graph model with the current vehicle as the root node, and include surrounding vehicles with non-zero lane-changing intention probability as risk nodes in the model. Calculate the spatiotemporal conflict risk value based on the current vehicle's preset lane-changing path and the predicted trajectory of each risk surrounding vehicle within a future time T (the same as the value of T in step S3). Then, weight and fuse the risk values ​​according to the lane-changing intention probability to obtain a comprehensive risk index.

[0050] S5. Risk Warning and Decision Recommendation Generation Steps: Compare the comprehensive risk indicators with dynamically adjusted thresholds to determine the risk level, and generate decision recommendations for the vehicle based on the risk level and the probability of the vehicle's intention to change routes.

[0051] Preferably, in step S2, the calculation of the trajectory offset angle further integrates visual perception data, obtains the lateral offset d_lat of the target vehicle relative to the lane line through the vehicle-mounted camera, and calibrates the trajectory offset angle using the following formula:

[0052] θ_calibrated=γ×θ+(1-γ)×arctan(d_lat / L)

[0053] Where γ is the fusion weighting coefficient, with a value ranging from 0.6 to 0.8, and L is the predetermined observation distance benchmark, with a value of 50 meters.

[0054] Preferably, in step S3, the intention recognition model is trained using a historical trajectory dataset, including a sequence of vehicle trajectory offset angles and corresponding lane-changing behavior labels. The network parameters are optimized through supervised learning, the intention probability output is calculated through forward propagation, and the cross-entropy loss function is minimized through backpropagation.

[0055] Preferably, in step S4, the calculation of the spatiotemporal conflict risk value specifically includes: predicting the trajectories of the current vehicle and the risky vehicle within the next time period T, calculating the minimum distance D_min between the two trajectories in the spatiotemporal domain and the time TTC to reach the minimum distance, and the spatiotemporal conflict risk value R_conflict is calculated using the following formula:

[0056] R_conflict=(P_intent_left+P_intent_right)×exp(-λ×D_min) / TTC

[0057] Where λ is the distance sensitivity coefficient, with a value ranging from 0.1 to 0.3, and P_intent_left and P_intent_right are the probabilities of the surrounding vehicle's intention to change lanes to the left and right, respectively.

[0058] Preferably, in step S4, the comprehensive risk index is calculated by weighted summation of the spatiotemporal conflict risk values ​​of all risk cycles. The comprehensive risk index R_combined is calculated using the following formula:

[0059] R_combined=Σ_j[w_j×R_conflict_j]

[0060] Where j is the index of the risky weekly vehicle, and w_j is the risk weight of the j-th weekly vehicle.

[0061] Preferably, the formula for calculating the risk weight w_j is:

[0062] w_j=α×(1 / dist_j)+β×|Δv_j|

[0063] Where α and β are weighting coefficients, with α ranging from 0.6 to 0.8 and β ranging from 0.2 to 0.4. dist_j is the longitudinal relative distance between the vehicle and the j-th surrounding vehicle in the lane (in meters), and Δv_j is the longitudinal relative speed between the vehicle and the j-th surrounding vehicle (in m / s).

[0064] Preferably, in step S5, the dynamically adjusted threshold is determined based on the real-time traffic scenario by querying a predefined scenario-threshold mapping table. The scenario-threshold mapping table is generated through training on historical risk event data, wherein: the threshold range for unobstructed flow (traffic flow density < 10 vehicles / km / lane, average speed > 60km / h) is 0.3-0.4, and the threshold for stable flow (10 vehicles / km / lane ≤ traffic flow density < 25 vehicles / km / lane, 40km / h < average speed ≤ 60km / h) is... The threshold range is 0.4-0.5. The threshold range for saturated flow (25 vehicles / km / lane ≤ traffic flow density < 40 vehicles / km / lane, 20km / h < average speed ≤ 40km / h) is 0.5-0.6. The threshold range for congested flow (traffic flow density ≥ 40 vehicles / km / lane, average speed ≤ 20km / h) is 0.6-0.7. Real-time traffic scenarios are divided into four types based on the traffic flow density and average speed of the road where the vehicle is located: smooth flow, stable flow, saturated flow, and congested flow.

[0065] Preferably, in step S1, the multi-source sensor system includes a camera, millimeter-wave radar, and lidar. The multi-sensor data is fused using a Kalman filter algorithm, including noise reduction, coordinate alignment, and time synchronization.

[0066] Preferably, in step S2, the sampling frequency of the trajectory offset angle time sequence is 10Hz, the length of the historical trajectory sequence covers at least 5 seconds, and the value of N is 10.

[0067] Preferably, in step S4, the update frequency of the dynamic risk graph model is 10Hz, and the node risk and edge weight are recalculated based on the latest sensor data at each time step.

[0068] Example 1: Lateral Cut-in Risk Warning in Highway Scenarios

[0069] In this embodiment, the vehicle is traveling at 100 km / h in the middle lane of a highway and intends to change lanes to the right into the adjacent lane. The specific implementation steps are as follows:

[0070] S1. Multi-source perception and trajectory extraction: The vehicle-mounted millimeter-wave radar detects a vehicle (denoted as vehicle A) 85 meters to the right front. Its initial state is: the lateral distance from this vehicle is 3.5 meters, and the relative speed is -5 km / h (decelerating). The camera also detects that the lateral offset d_lat of vehicle A relative to its lane line is 0.2 meters.

[0071] S2. Calculation of trajectory offset features: The initial trajectory offset angle of the vehicle A is calculated to be θ = arctan(0.35 / 27.7) = 0.72°. After calibration with visual data, θ_calibrated = 0.7 × 0.72° + (1 - 0.7) × arctan(0.2 / 50) = 0.71°.

[0072] S3. Lane Change Intent Recognition: Input the trajectory offset angle sequence of vehicle A in the last second (corresponding to N=10 sampling points, the sequence is [0.65°, 0.68°, ..., 0.71°]) into a pre-trained feedforward neural network. The output intent probabilities are: P_intent_left=0.05, lane keeping probability=0.35, P_intent_right=0.60, indicating that vehicle A has a clear intention to change lanes to the right.

[0073] S4. Dynamic Risk Fusion: Construct a dynamic risk map, take vehicle A as the risk node, predict the trajectories of the two vehicles in the next 2 seconds, and calculate the minimum distance D_min = 2.1 meters and TTC = 3.2 seconds;

[0074] Substitute into the formula to calculate the risk value of spatiotemporal conflict:

[0075] R_conflict_A=(0.05+0.60)×exp(-0.2×2.1) / 3.2≈0.133;

[0076] Risk weight w_A = 0.7 × (1 / 85) + 0.3 × |1.39| = 0.425;

[0077] The overall risk index R_combined = 0.133 × 0.425 = 0.057;

[0078] S5. Warning and Decision: The current traffic scenario is a stable flow with a threshold range of 0.4-0.5. The comprehensive risk index is 0.057 < 0.4, which is judged as low risk. The system prompts "Pay attention to vehicles on the right front. It is recommended to maintain the current speed."

[0079] Example 2: Multiple Risk Identification in Urban Expressway Congestion Scenarios

[0080] This vehicle is traveling at 30 km / h on an urban expressway with a traffic flow density of 35 vehicles / km·lane, which is considered congested. This vehicle is preparing to change lanes to the left, and the process is as follows:

[0081] S1. The multi-source sensing system simultaneously detects vehicle B (relative speed +5km / h) 50 meters to the left rear and vehicle C (relative speed -10km / h) 25 meters to the left front.

[0082] S2. The calculated trajectory offset angle sequence of vehicle B shows a continuous rightward deviation, with an average θ value of -2.5°; the trajectory offset angle of vehicle C fluctuates between -0.5° and +0.5°.

[0083] S3. Intent recognition model output: P_intent_left = 0.85 for car B and P_intent_left = 0.10 for car C.

[0084] S4. Establish a dynamic risk diagram containing two risk nodes:

[0085] Zhou Che B:

[0086] Calculate the risk weight w_B:

[0087] Parameters: dist_B = 50 meters, Δv_B = +5km / h ≈ +1.39m / s, take α = 0.7, β = 0.3.

[0088] Substituting into the formula: w_B=0.7×(1 / 50)+0.3×|1.39|=0.014+0.417=0.431.

[0089] Predict the trajectory and calculate: D_min = 3.5 meters, TTC = 4.1 seconds.

[0090] Calculate the spatiotemporal conflict risk value R_conflict_B (taking λ=0.2):

[0091] R_conflict_B=(0.85+0)×exp(-0.2×3.5) / 4.1=0.85×0.497 / 4.1≈0.103.

[0092] Zhou Che C:

[0093] Calculate the risk weight w_C:

[0094] Parameters: dist_C = 25 meters, Δv_C = -10 km / h ≈ -2.78 m / s, take α = 0.7, β = 0.3.

[0095] Substituting into the formula: w_C=0.7×(1 / 25)+0.3×|-2.78|=0.028+0.834=0.862.

[0096] Predict the trajectory and calculate: D_min = 8.2 meters, TTC = 6.3 seconds.

[0097] Calculate the spatiotemporal conflict risk value R_conflict_C (taking λ=0.2):

[0098] R_conflict_C=(0.10+0)×exp(-0.2×8.2) / 6.3=0.10×0.194 / 6.3≈0.003.

[0099] Calculate the comprehensive risk index:

[0100] R_combined=(0.103×0.431)+(0.003×0.862)≈0.044+0.003=0.047.

[0101] S5. Early Warning and Decision-Making: The threshold for congested traffic scenarios is 0.6-0.7. The comprehensive risk index of 0.047 is far below the lower threshold of 0.6, indicating a low overall risk. However, given that vehicle B has a high probability of intending to change lanes independently (P_intent_left=0.85) and a relatively small spatial-temporal distance (D_min=3.5 meters, TTC=4.1 seconds), the system still generates a warning message: "The vehicle behind you on the left intends to cut in; it is recommended to postpone lane changing and pay attention."

[0102] Example 3: Training and Implementation of the Intent Recognition Model

[0103] This embodiment details the training process of the intent recognition model:

[0104] S1. Data Preparation: Using the NGSIM highway dataset, extract the trajectory offset angle sequences of 5000 lane change events. Each sequence is 2 seconds long (20 time points), and label the corresponding lane change direction.

[0105] S2. Network Structure: A three-layer feedforward neural network is used, with 20 neurons in the input layer, 15 neurons in the hidden layer, and 3 neurons in the output layer (corresponding to three intent categories).

[0106] S3, Training parameters: learning rate 0.001, batch size 32, training epochs 100;

[0107] S4. Performance metrics: Achieved an accuracy of 92.3% on the test set, with a particularly high accuracy of 87.5% for early lane change intention recognition (1.5 seconds before the lane change occurs).

[0108] Example 4: Implementation of the Adaptive Threshold Mechanism

[0109] This example illustrates the specific implementation of dynamic threshold adjustment:

[0110] S1. Data Foundation: Collect 1,000 hours of real driving data and label risk events under different traffic conditions;

[0111] S2, Threshold Training: Use a logistic regression model to establish a mapping relationship between traffic flow parameters and the optimal risk threshold, maximizing the F1 score for risk identification;

[0112] S3. Online Update: The system updates the local threshold mapping table every 5 minutes to ensure adaptation to real-time changes in traffic conditions;

[0113] S4. Performance Verification: In actual testing, compared with the fixed threshold scheme, the false alarm rate was reduced by 42% and the false negative rate was reduced by 35%.

[0114] Experimental Example 1:

[0115] 1. Experimental Objective

[0116] To verify the overall performance of this invention in a highway scenario, the key technical indicators to be evaluated include the accuracy of identifying lane-changing intentions of surrounding vehicles, the advance time of risk warning, and the false alarm rate.

[0117] 2. Test Methods

[0118] 2.1 Experimental Dataset

[0119] Using the HighD highway public dataset, 500 typical vehicle lane-changing events were selected as test samples, covering different traffic flow densities (unobstructed flow, steady flow, saturated flow) and weather conditions (dry, light rain).

[0120] 2.2 Comparison Objects

[0121] This invention is compared in parallel with the following three types of methods:

[0122] Comparison Method 1: Rasterized Kinetic Energy Loss Model Based on Patent CN110085056A

[0123] Comparison Method 2: LSTM trajectory prediction model based on patent CN112863245A

[0124] Traditional baseline methods: Time-based collision warning (TTC) method based on fixed thresholds

[0125] 2.3 Testing Process

[0126] 1. Perform data preprocessing on the test samples to extract key information such as motion state data and trajectory sequences of the vehicle and surrounding vehicles;

[0127] 2. The present invention and various comparative methods are used to identify and warn of lane-changing risks in the samples;

[0128] 3. Statistically analyze the performance metrics of various methods, including:

[0129] Intent recognition accuracy: The proportion of samples that correctly identify the lane-changing intent of a vehicle in the surrounding area out of the total number of samples;

[0130] Average warning lead time: The average time from when the system issues a warning to when a potential conflict occurs;

[0131] Risk false alarm rate: The number of warnings issued per unit of time (hour) when there is no actual risk.

[0132] 3. Test Results

[0133] Performance indicators The invention Comparative method 1 Comparative method 2 Conventional method Lane change intention recognition accuracy rate (%) 93.2 Not involved 68.5 Not involved Average warning lead time (s) 2.4 0.9 1.8 1.1 Risk false alarm rate (times / hour) 5.1 23.6 15.4 19.3

[0134] 4. Experimental Conclusions

[0135] This invention significantly outperforms existing comparative methods in three core indicators: accuracy of lane-changing intention recognition, early warning time of risk warning, and false alarm rate control. It achieves this through features extraction of vehicle trajectory deviation, early recognition of lane-changing intention, and dynamic fusion mechanism of multi-vehicle interaction risk. It can meet the actual needs of intelligent vehicle lane-changing risk advance warning in highway scenarios.

[0136] Experimental Example 2:

[0137] 1. Experimental Objective

[0138] The purpose of this invention is to verify its early warning effect on the chain risks caused by the active behavior of surrounding vehicles in complex multi-vehicle interaction scenarios, and to evaluate the timeliness of the model's response to dynamic risks and the rationality of its decision-making.

[0139] 2. Test Methods

[0140] 2.1 Simulation Scenario Construction

[0141] A multi-vehicle interaction scenario was built based on the Prescan simulation platform, with the following specific parameters:

[0142] This vehicle is currently traveling at an initial speed of 80 km / h in the middle lane and plans to change lanes to the left to reach the target lane.

[0143] The vehicle behind in the target lane has an initial speed of 90 km / h and an initial longitudinal distance of 60 m from this vehicle. It shows a strong intention to change lanes to the left (probability of lane change intention P_intent_left=0.85).

[0144] Traffic scenario: steady flow (traffic flow density 20 vehicles / km / lane, average vehicle speed 55km / h).

[0145] 2.2 Test Content

[0146] The present invention and comparative method 1 (gridized kinetic energy loss model of patent CN110085056A) were used respectively to identify risks in this scenario, and the early warning timing, risk level judgment results and decision suggestions were recorded.

[0147] 3. Test Results

[0148] 3.1 Test Results of the Invention

[0149] 1. Lane change intention recognition: Accurately identify the left lane change intention of the vehicle behind in the target lane, and output the intention probability P_intent_left=0.83 (error ≤2.4% from the preset value of 0.85).

[0150] 2. Comprehensive risk calculation: The comprehensive risk index calculated based on the dynamic risk diagram model is 0.72;

[0151] 3. Risk Level Determination: Based on the stable flow scenario threshold (0.45), it is determined to be high risk;

[0152] 4. Decision Recommendation: Output a warning message: "A vehicle behind is rapidly approaching the target lane and intends to change lanes. It is recommended to abandon the lane change."

[0153] 3.2 Comparison of Test Results of Method 1

[0154] Because the distance between the vehicle and the vehicle behind in the target lane was greater than the safety threshold (i.e. critical safety distance) set by comparison method 1 in the initial stage, the model did not identify the potential risk; the alarm was only triggered when the time collision warning (TTC) between the two vehicles was less than 1.0s, which is 1.8s later than the present invention.

[0155] 4. Experimental Conclusions

[0156] This invention has a superior risk prediction and identification capability in complex multi-vehicle interaction scenarios. It can effectively capture the chain risks caused by dynamic behaviors such as lane changing by surrounding vehicles, provide earlier warnings and more targeted decision-making suggestions, and significantly reduce the risk of lane changing accidents in multi-vehicle interaction scenarios.

[0157] Experimental Example 3:

[0158] 1. Experimental Objective

[0159] The training effect and early recognition reliability of the vehicle lane change intention recognition model in this invention are verified.

[0160] 2. Test Methods

[0161] 2.1 Training Dataset

[0162] Using the NGSIM highway dataset, trajectory data of 5000 valid lane change events were extracted. Each event contains a 2-second sequence of vehicle trajectory offset angles (sampling frequency 10Hz) and is labeled with the corresponding lane change behavior (left lane change, lane keeping, right lane change).

[0163] 2.2 Model Structure

[0164] A three-layer feedforward neural network is used as the intent recognition model, and its specific structure is as follows:

[0165] Input layer: 20 neurons (corresponding to a 2-second trajectory offset angle sequence);

[0166] Hidden layer: 15 neurons, using ReLU activation function;

[0167] Output layer: 3 neurons (corresponding to three types of lane-changing intentions), with Softmax as the activation function.

[0168] 2.3 Training Parameters

[0169] Learning rate: 0.001;

[0170] Batch size: 32;

[0171] Training rounds: 100;

[0172] Loss function: Cross-entropy loss function;

[0173] Optimizer: Adam optimizer.

[0174] 2.4 Test Indicators

[0175] Overall recognition accuracy: The proportion of samples in the test set that correctly recognize the intent;

[0176] Early recognition accuracy: The percentage of correct intent recognitions performed 1.5 seconds before a lane change occurs.

[0177] 3. Test Results

[0178] Overall model recognition accuracy: 92.3%;

[0179] Early recognition accuracy 1.5 seconds before lane change: 87.5%;

[0180] Model training convergence: After 60 rounds of training, the loss function tends to stabilize, with no overfitting.

[0181] 4. Experimental Conclusions

[0182] The lane change intention recognition model used in this invention has good training results. It not only has a high overall recognition accuracy, but also can reliably identify lane change behavior 1.5 seconds before it occurs, providing core technical support for forward-looking early warning of lane change risks.

[0183] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0184] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories, characterized in that, Includes the following steps: S1. Multi-source perception and trajectory extraction steps: The vehicle's multi-source sensor system acquires the motion status data and historical trajectory sequences of the vehicle and surrounding vehicles in real time. The perception range covers the trajectory records of vehicles within 200 meters in front of the vehicle and the motion status perception of vehicles within 80 meters around the vehicle. S2. Calculation steps for vehicle trajectory deviation characteristics: For the vehicle in front in this lane, the vehicle in front in the target lane, the vehicle behind in the target lane, and vehicles in adjacent lanes, calculate their trajectory deviation angle time series. The trajectory deviation angle θ is defined as the angle between the instantaneous direction of the vehicle's movement and the direction of travel of the centerline of its lane, and is calculated using the following formula: θ = arctan(v_y / v_x); Where v_x is the vehicle's velocity component in the longitudinal direction of the lane, and v_y is the vehicle's velocity component in the lateral direction of the lane; S3, the following steps are taken to identify the intention of the vehicle to change lanes: Based on the time sequence of the trajectory offset angle, the probability of the vehicle to change lanes is calculated by a pre-trained intention recognition model. The intention recognition model adopts a feedforward neural network structure. The input is the trajectory offset angle sequence of the current time and the previous N time. The output is the probability of the vehicle to change lanes to the left, maintain lane or change lanes to the right in the next T time. The value of T is 1-3 seconds. S4. Multi-vehicle interaction risk dynamic fusion steps: Construct a dynamic risk graph model with the current vehicle as the root node, and include surrounding vehicles with non-zero lane-changing intention probability as risk nodes in the model. Calculate the spatiotemporal conflict risk value based on the current vehicle's preset lane-changing path and the predicted trajectory of each risk surrounding vehicle within a future time T (consistent with the value of T in step S3). Then, weight and fuse the risk values ​​according to the lane-changing intention probability to obtain a comprehensive risk index. S5. Risk Warning and Decision Recommendation Generation Steps: Compare the comprehensive risk indicators with dynamically adjusted thresholds to determine the risk level, and generate decision recommendations for the vehicle based on the risk level and the probability of the vehicle's intention to change routes.

2. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S2, the calculation of the trajectory offset angle further integrates visual perception data. The lateral offset d_lat of the target vehicle relative to the lane line is obtained through the vehicle-mounted camera, and the trajectory offset angle is calibrated using the following formula: θ_calibrated=γ×θ+(1-γ)×arctan(d_lat / L) Where γ is the fusion weighting coefficient, with a value ranging from 0.6 to 0.8, and L is the predetermined observation distance benchmark, with a value of 50 meters.

3. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S3, the intention recognition model is trained using a historical trajectory dataset, including the sequence of vehicle trajectory offset angles and the corresponding lane-changing behavior labels. The network parameters are optimized through supervised learning, the intention probability output is calculated through forward propagation, and the cross-entropy loss function is minimized through backpropagation.

4. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S4, the calculation of the spatiotemporal conflict risk value specifically includes: predicting the trajectories of the current vehicle and the vehicle at risk within the next time interval T, calculating the minimum distance D_min between the two trajectories in the spatiotemporal domain and the time TTC to reach that minimum distance, and the spatiotemporal conflict risk value R_conflict is calculated using the following formula: R_conflict=(P_intent_left+P_intent_right)×exp(-λ×D_min) / TTC Where λ is the distance sensitivity coefficient, with a value ranging from 0.1 to 0.3, and P_intent_left and P_intent_right are the probabilities of the surrounding vehicle's intention to change lanes to the left and right, respectively.

5. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S4, the comprehensive risk index is calculated by weighted summation of the spatiotemporal conflict risk values ​​of all risk cycles. The comprehensive risk index R_combined is calculated using the following formula: R_combined=Σ_j[w_j×R_conflict_j] Where j is the index of the risky weekly vehicle, and w_j is the risk weight of the j-th weekly vehicle.

6. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 5, characterized in that, The formula for calculating the risk weight w_j is: w_j=α×(1 / dist_j)+β×|Δv_j| Where α and β are weighting coefficients, with α ranging from 0.6 to 0.8 and β ranging from 0.2 to 0.

4. dist_j is the longitudinal relative distance between the vehicle and the j-th surrounding vehicle in the lane (in meters), and Δv_j is the longitudinal relative speed between the vehicle and the j-th surrounding vehicle (in m / s).

7. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S5, the dynamically adjusted threshold is determined based on the real-time traffic scenario by querying a predefined scenario-threshold mapping table; The scenario-threshold mapping table is generated through training on historical risk event data. The threshold range for smooth flow (traffic density < 10 vehicles / km / lane, average speed > 60km / h) is 0.3-0.4; for stable flow (10 vehicles / km / lane ≤ traffic density < 25 vehicles / km / lane, 40km / h < average speed ≤ 60km / h) it is 0.4-0.5; for saturated flow (25 vehicles / km / lane ≤ traffic density < 40 vehicles / km / lane, 20km / h < average speed ≤ 40km / h) it is 0.5-0.6; and for congested flow (traffic density ≥ 40 vehicles / km / lane, average speed ≤ 20km / h) it is 0.6-0.

7. Real-time traffic scenarios are categorized into these four types—smooth flow, stable flow, saturated flow, and congested flow—based on the traffic density and average speed of the road where the vehicle is located.

8. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S1, the multi-source sensor system includes a camera, millimeter-wave radar, and lidar. The multi-sensor data is fused using a Kalman filter algorithm, including noise reduction, coordinate alignment, and time synchronization.

9. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S2, the sampling frequency of the trajectory offset angle time sequence is 10Hz, the length of the historical trajectory sequence covers at least 5 seconds, and the value of N is 10.

10. The method for quantifying vehicle lane-changing risk based on weekly vehicle trajectories according to claim 1, characterized in that, In step S4, the dynamic risk graph model is updated at a frequency of 10Hz, and the node risk and edge weight are recalculated at each time step based on the latest sensor data.

Images