A risk assessment and trajectory optimization method based on millimeter wave radar

Abstract

The application is suitable for the field of automatic driving technology, and provides a risk assessment and trajectory optimization method based on a millimeter wave radar, which comprises the following steps: collecting surrounding target information and vehicle state information, performing effective target screening and confidence normalization on the target information, calculating a confidence weighted global time to collision (TTC) risk index based on the screened effective target, constructing an asymmetric safety corridor based on a Frenet coordinate system, and constructing a longitudinal adaptive safety acceleration constraint; integrating the above constraints into an MPC optimization problem to solve an optimal control sequence; and simultaneously monitoring the global TTC risk index and the safety corridor width in real time through an independent MCU module, and outputting an emergency braking instruction when it is determined that the dangerous state condition is met. The application explicitly integrates the millimeter wave radar sensing uncertainty into the control logic, realizes horizontal and vertical collaborative optimization and hardware level safety bottom, and effectively solves the safety control failure problem in the sensing degradation scene.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and in particular relates to a risk assessment and trajectory optimization method based on millimeter-wave radar. Background Technology

[0002] Autonomous driving technology is a core direction for intelligent transportation and automotive industry innovation, and is gradually expanding from limited scenarios to open urban roads. The safe operation of the system highly depends on the coordinated cooperation of environmental perception, decision-making and planning, and motion control. Environmental perception is the basic input link for autonomous driving, and commonly used devices include visual cameras, LiDAR, and millimeter-wave radar. Among them, visual cameras are prone to detection failure in adverse lighting and weather conditions such as night, backlight, rain, and fog. LiDAR's detection accuracy decreases in rain, fog, and haze due to echo scattering and energy attenuation. Millimeter-wave radar, with its stable operation in all weather conditions, strong resistance to environmental interference, and reliable ranging and speed measurement accuracy, has become the core perception device for ensuring the safe driving of autonomous vehicles, and is widely used in key functions such as following the vehicle ahead, collision warning, and dynamic obstacle avoidance.

[0003] However, existing technologies generally suffer from the following technical problems: In real-world complex traffic scenarios, millimeter-wave radar is susceptible to multipath reflection, ground clutter interference, and weak target echoes, frequently resulting in target misses, increased false alarm rates, and decreased detection accuracy, leading to perception degradation. This is especially true in complex conditions such as high-speed driving, multi-vehicle interaction, lane changes, tunnel entrances and exits, and intersections, where perception uncertainty is further amplified, causing vehicle state estimation lag, reduced stability of multi-target tracking, and drastic fluctuations in perception data over time. Current autonomous driving control solutions often use fixed safety distances and fixed threshold collision times for risk assessment and trajectory planning, relying entirely on deterministic perception results and failing to incorporate target detection confidence into risk calculations, thus failing to address the potential risks caused by perception distortion. Furthermore, current lateral safety constraints often employ fixed-width symmetrical designs, which are unsuitable for adapting to... In real-world road conditions such as uneven distribution of obstacles on the left and right sides of the road, asymmetrical lane structure, and narrow lane encounters, longitudinal control uses fixed acceleration or fixed safety distance constraints without dynamically adjusting the constraint strength based on real-time risk levels. Furthermore, most solutions separate lateral obstacle avoidance and longitudinal following control, lacking a unified dynamic safety constraint and collaborative optimization mechanism. When perception data fluctuates or obstacles suddenly intrude into the lane, problems such as control constraint failure, infeasibility of trajectory optimization, and severe control output jitter can easily occur. In addition, the system relies solely on the upper-level software controller to make safety decisions, lacking a hardware-level safety protection mechanism independent of the main controller. When extreme dangerous situations such as complete target loss, drastic reduction in collision time, or insufficient safety corridor width occur, hard real-time safety intervention cannot be triggered quickly, and the vehicle is prone to entering an uncontrollable state, making it impossible to achieve a dynamic balance between driving safety and road traffic efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a risk assessment and trajectory optimization method based on millimeter-wave radar, aiming to solve the problems existing in the background technology.

[0005] This invention is implemented as follows: a risk assessment and trajectory optimization method based on millimeter-wave radar, comprising the following steps:

[0006] Step 1: Collect surrounding target information and vehicle status information and perform standardized processing. At the same time, calculate the global collision time (TTC) risk index and output it synchronously.

[0007] Step 2: Filter the surrounding target information to obtain a set of valid targets, and normalize the original detection confidence values ​​of the surrounding target information to obtain normalized confidence.

[0008] Step 3: Based on the vehicle state information and the set of effective targets, calculate the relative motion state between the vehicle and each effective target, calculate the projected collision time for targets with an approaching trend, and introduce the normalized confidence score for weighted correction, and select the minimum weighted TTC as the global risk index.

[0009] Step 4: Establish reference lines based on the Frenet coordinate system, obtain lane boundary information and calculate the remaining space on the left and right sides; based on the effective target set and the normalized confidence level, construct risk factors on the left and right sides respectively, dynamically adjust the safety corridor boundary, and generate asymmetric safety corridor constraints.

[0010] Step 5: Based on the global risk index, construct a longitudinal adaptive safety acceleration constraint.

[0011] Step 6: Integrate the asymmetric safety corridor constraint and the longitudinal adaptive safety acceleration constraint, construct the MPC optimization problem based on the vehicle dynamics model, and use the rolling time-domain optimization strategy to solve the optimal control sequence and output it to the vehicle actuator.

[0012] Step 7: Monitor the global risk index and the width of the safety corridor corresponding to the asymmetric safety corridor constraint in real time. When it is determined that the preset dangerous state conditions are met, output an emergency braking command.

[0013] The present invention provides a risk assessment and trajectory optimization method based on millimeter-wave radar, which has the following beneficial effects:

[0014] This invention constructs a risk assessment mechanism that jointly models target state and detection confidence, achieves dynamic collision risk quantification based on improved TTC, generates an asymmetric safety corridor by combining target spatial distribution and establishes longitudinal adaptive acceleration constraints, explicitly integrates perception uncertainty into the control logic, and dynamically adjusts the collision time according to perception uncertainty. The safety boundary automatically tightens or loosens according to the real-time risk level, solving the problem of risk assessment distortion in traditional fixed threshold assessment methods under perception degradation scenarios. At the same time, it relies on model predictive control to achieve lateral obstacle avoidance and longitudinal following coordination optimization, avoiding constraint failure and output jitter caused by lateral and longitudinal separation control, and significantly improving control robustness in complex dynamic scenarios. In addition, an independent MCU hardware safety loop is added, which can directly trigger hard real-time emergency braking without relying on upper-layer software under extreme conditions, fundamentally solving the safety control failure problem caused by millimeter-wave radar perception degradation. While ensuring driving safety, it improves traffic efficiency and enables the autonomous driving system to have practical and highly reliable operation capabilities in complex dynamic environments. Attached Figure Description

[0015] Figure 1 This invention provides a system framework diagram for a risk assessment and trajectory optimization method based on millimeter-wave radar. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0018] like Figure 1 As shown, this invention constructs a complete closed-loop control framework of "millimeter-wave radar perception → real-time risk assessment → adaptive constraint generation → MPC optimization control → hardware safety fallback". The system first collects the status of surrounding targets and the vehicle's own status via millimeter-wave radar, and sends this data to the TTC risk assessment module to calculate the global collision time risk index. This index is used to generate an asymmetric safety corridor and longitudinal adaptive acceleration constraints, serving as the dynamic boundary for MPC optimization control. Simultaneously, it is transmitted to an independent MCU module, which determines in real time whether the vehicle is in an extremely dangerous state based on minimum safety requirements. If the determination is safe, the domain controller executes MPC trajectory optimization and outputs control commands to drive the vehicle; if the determination is unsafe, the MCU directly triggers emergency braking, bypassing the upper-level software to achieve hardware-level safety fallback.

[0019] The autonomous driving domain controller connects to an external millimeter-wave radar sensor, a vehicle status acquisition unit, and an independent MCU (Microcontroller Unit) as safety fallback hardware, establishing a complete control architecture encompassing perception input, data processing, optimization decision-making, execution output, and hardware protection. The millimeter-wave radar is responsible for collecting target information about surrounding traffic participants, the vehicle status acquisition unit is responsible for acquiring the vehicle's motion status, and the MCU, as an independent hardware safety module, works in parallel with the domain controller.

[0020] Step 1: Data Collection and Global TTC Risk Indicator Calculation and Output

[0021] In step 1, information about surrounding targets is collected in real time using millimeter-wave radar. The original state of each target can be represented as follows: ,in Let x be the horizontal and vertical positions of the target in the Cartesian coordinate system. The target's lateral and longitudinal velocities are given. Simultaneously, the radar outputs the raw detection confidence values ​​for each target. Synchronously acquire the vehicle's longitudinal position with status indicators. Horizontal position Driving speed Heading angle yaw rate Five core motion state variables. Two types of data are synchronously transmitted to the system front end at a fixed frequency (typically 20Hz~50Hz) for standardization processing.

[0022] The standardized and valid target data are sequentially fed into the risk assessment, constraint generation, and trajectory optimization stages. The system first calculates the Global Time-of-Collision (TTC) risk index through subsequent steps. This index is output synchronously in two paths: the first path is fed into the autonomous driving domain controller, serving as the core input parameter for asymmetric safety corridor generation, longitudinal adaptive constraint construction, and MPC trajectory optimization, corresponding to the lateral constraint boundary. With longitudinal constraint boundary ,in This represents the lateral position of the vehicle in the Frenet coordinate system. and The safety corridor boundary is dynamically adjusted according to TTC and target confidence. For longitudinal acceleration control, The maximum physical acceleration of the vehicle. This is an upper bound for adaptive safety acceleration based on TTC risk.

[0023] The second input goes directly to the independent MCU hardware module, serving as the core basis for determining extreme dangerous conditions and triggering emergency braking. Finally, the domain controller outputs smooth control commands, and the MCU module outputs hard real-time fallback commands, jointly driving the vehicle's actuators to complete safe driving and dynamic obstacle avoidance.

[0024] As a preferred implementation, to ensure operational consistency and communication reliability between the autonomous driving domain controller and the MCU hardware safety module, a bidirectional Life heartbeat interaction mechanism is established. The domain controller sends a Life signal to the MCU at fixed intervals (e.g., 50ms), containing information on system operating status, trajectory optimization execution status, current risk level, and control link health status. The MCU simultaneously returns its own operating status, braking execution availability, and safety monitoring status. An independent timeout detection mechanism is internally configured in the MCU. If it fails to receive a Life signal from the domain controller for several consecutive cycles (typically 3-5 cycles), or detects control output freeze, risk data update interruption, or communication anomalies, it immediately determines that the main control link has failed and enters a safety takeover mode, bypassing the domain controller to directly output hard real-time emergency braking commands to the vehicle actuators. Simultaneously, the domain controller continuously monitors the Life status information returned by the MCU. When it detects abnormal MCU offline status or execution failure, it automatically increases the control conservatism level, actively tightens the safety corridor boundary, and reduces the upper limit of longitudinal acceleration. This bidirectional heartbeat mechanism achieves fault isolation, redundant safety monitoring, and reliable safety protection under extreme conditions.

[0025] Step 2: Millimeter-wave radar target selection and confidence modeling

[0026] In step 2, the surrounding target information collected by the millimeter-wave radar is filtered to obtain a set of effective targets. At the same time, the original values ​​of the detection confidence of the surrounding target information are normalized to obtain normalized confidence.

[0027] Specifically, after the system completes the collection and synchronization of raw data, it first performs a triple effective target screening on the raw target data.

[0028] The first screening eliminates invalid targets that pose no potential collision threat from behind or to the sides of the vehicle. The criterion for this judgment is the target's longitudinal position relative to the vehicle. Horizontal position Only retain those that meet the requirements. and The goal, among which This is the lane expansion factor (typically 1.5~2.0). This is the current lane width. This step can quickly eliminate vehicles behind or to the sides and rear of the vehicle that pose no threat, reducing the computational load for subsequent actions.

[0029] The second layer of filtering removes environmental interference targets that exceed the lane width, retaining only targets located within a certain lateral range in front of the vehicle. This step effectively eliminates non-threatening radar echoes from stationary objects on the roadside, medians, guardrails, etc., preventing false alarms from interfering with risk assessment.

[0030] The third screening step removes those below the minimum confidence threshold. The low reliability of millimeter-wave radar leads to false alarms and weak target detection. In cluttered environments or with weak target echoes, millimeter-wave radar outputs lower detection confidence levels. The system has a preset minimum confidence threshold. (Typical values ​​are 0.3~0.5), only retain [the relevant values]. The target. After three rounds of screening, the final set of effective targets containing only high-threat targets was obtained: .

[0031] After target screening is completed, detection confidence modeling is performed. The raw detection confidence output from the millimeter-wave radar is then used. Perform normalization processing and map to a unified standard. Standard range, i.e. ,in This represents the maximum confidence level of the radar output (typically 1). A value closer to 1 indicates higher reliability of the target detection result and lower perception uncertainty; a value closer to 0 indicates a more significant impact from clutter interference or missed detection, and higher perception uncertainty. This normalized confidence level serves as a core parameter throughout the subsequent risk assessment and constraint generation process, and is linked to the collision time indicator to achieve deep integration of perception quality and control logic.

[0032] Step 3: Real-time Collision Risk Assessment (TTC Quantification)

[0033] In step 3, based on the vehicle state information from step 1 and the set of valid targets from step 2, the relative motion state between the vehicle and each valid target is calculated. For targets with an approaching trend, the projected collision time is calculated, and the normalized confidence from step 2 is introduced for weighted correction. The minimum weighted TTC is selected as the global risk indicator.

[0034] After completing the effective target selection and confidence level normalization, the real-time collision risk assessment phase begins. First, the relative motion state between the vehicle and each effective target is calculated. Let the vehicle's state be denoted as... , The target state is , Then the relative position vector is The relative velocity vector is .

[0035] Then, it is determined whether there is an approaching trend. For targets that show an approaching trend with the vehicle, the projected collision time is calculated. The criterion for determining the approaching trend is... A negative dot product of the relative position vector and the relative velocity vector indicates that the target is approaching the vehicle. For targets that do not meet this condition (i.e., are moving away or remaining stationary), their projection TTC is directly assigned a value of 0. This avoids unnecessary calculations that consume system resources.

[0036] For targets exhibiting an approaching tendency, the projected collision time is defined as the time required for the vehicle and the target to collide according to their current relative motion state. The calculation formula is: The denominator Let be the square modulus of the relative velocity. This formula essentially projects the relative distance along the relative velocity direction to obtain the collision time along that direction.

[0037] Introducing detection confidence level based on projected TTC Perform a weighted adjustment. The weighted adjusted TTC is: .because ,therefore When the target detection confidence is low (i.e., perceptual uncertainty is high), The weighted TTC approaches 1, making it close to the original TTC. When the target detection confidence is high, the weighted TTC will be appropriately reduced, thereby increasing the risk level of the target. This design can accurately characterize the real collision threat under the uncertainty of millimeter-wave radar perception.

[0038] The minimum weighted TTC among all valid targets is selected as the global risk indicator, expressed as: This indicator automatically identifies the most threatening target in the current scene. A lower TTC value indicates a higher collision risk; a higher TTC value or an infinite value indicates no immediate collision risk.

[0039] The calculated global TTC index is synchronously transmitted to the autonomous driving domain controller and MCU hardware module, providing a unified risk assessment basis for both control units. The domain controller uses this index to generate dynamic safety constraints and incorporates them into MPC optimization; the MCU uses the same TTC index to independently monitor extreme dangerous states, realizing the sharing and independent use of risk indicators at the software decision-making layer and the hardware protection layer.

[0040] Step 4: Asymmetric security corridor generation

[0041] In step 4, a reference line is established based on the Frenet coordinate system to obtain lane boundary information and calculate the remaining space on the left and right sides; based on the effective target set in step 2 and the normalized confidence level, risk factors on the left and right sides are constructed respectively, the safety corridor boundary is dynamically adjusted, and asymmetric safety corridor constraints are generated.

[0042] After obtaining the global TTC risk indicators, the system proceeds to the lateral dynamic asymmetric safety corridor generation stage. The system first establishes a reference trajectory based on the Frenet coordinate system. The Frenet coordinate system uses the reference path (usually the lane centerline or planned trajectory) as a reference, decomposing the vehicle's position into longitudinal distances along the reference line. and lateral offset Let the lateral offset on the current reference trajectory be... The actual lateral offset of the vehicle is The lateral control objective is to make Stay in the safe corridor.

[0043] Calculate the remaining space from the reference trajectory to the left and right lane boundaries respectively. Let the lateral offset of the left lane boundary be... The lateral offset of the right lane boundary is Then the distance from the reference trajectory to the left boundary is The distance to the right boundary is In asymmetrical road conditions (such as obstacles on one side of the lane, narrow shoulders, etc.), and The two are not equal, and this method naturally supports this asymmetry.

[0044] The system constructs left-side risk factors respectively. With right-side risk factors The formula for calculating the risk factor on the left is: ,in This indicates that it is located on the left side of the vehicle (i.e., lateral offset). The effective target set; The lateral offset of the target relative to the reference trajectory; The longitudinal distance between the target and the vehicle; Assuming a longitudinal distance attenuation constant (typically 50-100m), an exponential attenuation is introduced. This can reduce the contribution of distant targets to risk; The left-hand risk sensitivity coefficient (an adjustable parameter, typically ranging from 1.0 to 1.5); This is the normalized detection confidence score for the target. This factor comprehensively considers the degree of lateral encroachment by the target. (A larger value indicates deeper penetration), longitudinal distance attenuation Perceiving uncertainty and the choice of maximum risk .

[0045] The formula for calculating the risk factor on the right is: ,in Located on the right side of the vehicle (lateral offset) The effective target set, The right-hand side represents the risk sensitivity coefficient. and The higher the value, the higher the risk on the corresponding side.

[0046] The safety corridor boundaries are dynamically adjusted based on risk factors on both the left and right sides. The left boundary (upper limit) is: The right boundary (lower limit) is: When a certain risk factor When it approaches 0, the side boundary hardly shrinks, preserving the maximum usable space; when As the distance approaches 1, the side boundary contracts inward to near the reference trajectory, forming a very narrow safety corridor that forces vehicles away from the high-risk side. This asymmetric contraction mechanism is well-suited for complex scenarios such as narrow lane encounters, unilateral obstacle intrusions, and roadside parking.

[0047] Generated asymmetric safe corridor constraints The domain controller receives and integrates the data into the subsequent MPC trajectory optimization framework to achieve adaptive control for lateral obstacle avoidance. The corridor translates in real time with the reference trajectory, and each frame recalculates the left and right risk factors and updates the boundaries based on the latest perception data to ensure that the constraints always match the current risk.

[0048] Step 5: Construction of longitudinal adaptive safety acceleration constraints

[0049] In step 5, based on the global risk index in step 3, a longitudinal adaptive safety acceleration constraint is constructed.

[0050] While generating the horizontal asymmetric safety corridor, the system simultaneously constructs vertical adaptive safety acceleration constraints based on the global TTC index of the input domain controller. First, two key risk thresholds are preset: a risk warning threshold and a risk warning threshold. With emergency braking threshold And satisfy Typical parameter values ​​are: , It can be calibrated according to vehicle model and scenario.

[0051] The global TTC risk indicator is smoothly mapped to an upper bound of safety acceleration using a limiting function, expressed as follows: ,in The vehicle's maximum physical acceleration (typical value) ),function Input Limited to Within the range.

[0052] The regional interpretation mapping logic is represented as follows:

[0053] when hour, After clipping, it takes 1, therefore Vehicles are allowed to travel at maximum acceleration to ensure traffic efficiency.

[0054] when When the mapping value changes linearly between (0, 1), As TTC decreases, the speed decreases linearly, and the vehicle automatically slows down to reduce risk.

[0055] when At that time, the mapping value After clipping, take 0, therefore The vehicle is only allowed to decelerate or maintain a constant speed, and is prohibited from actively accelerating, entering a forced deceleration preparation state.

[0056] The lower limit of longitudinal acceleration is fixed at the maximum braking deceleration. (typical value) ),Right now The final longitudinal acceleration constraint is: .

[0057] The longitudinal adaptive safety constraint, together with the asymmetric safety corridor constraint, is fed into the MPC optimization module of the domain controller, forming a dynamic safety constraint system that coordinates the longitudinal and lateral directions. Through the longitudinal constraint driven by TTC, the vehicle can maintain its dynamism under low-risk conditions and automatically reduce its speed under high-risk conditions, achieving a dynamic balance between driving safety and traffic efficiency.

[0058] Step 6: Constrained Adaptive MPC Trajectory Optimization

[0059] In step 6, the asymmetric safety corridor constraint of step 4 and the longitudinal adaptive safety acceleration constraint of step 5 are integrated. Based on the vehicle dynamics model, the MPC optimization problem is constructed, and the optimal control sequence is solved by the rolling time-domain optimization strategy and output to the vehicle actuator.

[0060] After integrating the lateral asymmetric safety corridor constraint and the longitudinal adaptive acceleration constraint, the process proceeds to the constraint-adaptive MPC trajectory optimization stage. The controller uses a linear discrete vehicle dynamics model as the prediction basis, and the model expression is as follows: , where the state vector This includes vehicle longitudinal position, lateral position, speed, heading angle, and yaw rate; control inputs. This represents the longitudinal acceleration and the front wheel steering angle. (Matrix) and The model is obtained by discretizing the vehicle's kinematics or dynamics model. This model has a simple and efficient structure, making it suitable for the real-time computing needs of in-vehicle embedded platforms.

[0061] The MPC objective function balances trajectory tracking accuracy and control smoothness. The optimization objective is to minimize the tracking error and the rate of change of the control input, specifically expressed as: ,in For predicting the time domain (typically 10-20 steps), The reference trajectory state (usually derived from the upper-level path planning) The state error weight matrix is ​​a positive definite diagonal matrix. To control the input weight matrix (positive definite diagonal matrix), The first term penalizes trajectory tracking deviation, the second term penalizes control action amplitude, and the third term is a risk constraint. Together, they ensure trajectory tracking accuracy and the smoothness of control output.

[0062] The optimization problem must satisfy the following three core constraints:

[0063] 1. Vehicle dynamics constraints: Furthermore, the control input must meet the physical limit. .

[0064] 2. Lateral Asymmetric Safety Corridor Constraint: Within each prediction step, the vehicle's lateral offset must lie within the dynamic corridor, i.e. This constraint is linearized after transformation to the Frenet coordinate system, forming a convex inequality constraint.

[0065] 3. Longitudinal adaptive acceleration constraint: ,in It is calculated in real time from step 5.

[0066] All of the above constraints together constitute a standard convex optimization problem (quadratic programming), which can be solved in milliseconds using an efficient solver (such as OSQP).

[0067] The system employs a rolling time-domain optimization strategy, synchronously updating millimeter-wave radar sensing data, confidence information, and TTC risk indicators in each frame to resolve the aforementioned convex optimization problem, thereby obtaining the optimal control sequence for a future time domain period. Only the first control variable of the sequence (i.e., the optimal longitudinal acceleration at the current moment) is taken. and front wheel cornering The output is sent to the vehicle's actuators. This process is repeated in the next instant, forming a closed-loop control. This mechanism effectively addresses fluctuations in perception data and sudden obstacle intrusions, improving control robustness in complex traffic environments.

[0068] To further ensure the smoothness of the control output, a control increment penalty term can be added to the objective function. ,in , This is the incremental weight matrix. Meanwhile, when drastic changes in the perceived data render the constraints temporarily infeasible, the system can employ the slack variable method to introduce soft constraints on the safety corridor constraint, prioritizing the existence of a solution to the optimization problem and preventing control failure.

[0069] Step 7: MCU hardware-level security fallback execution

[0070] In step 7, the global risk index of step 3 and the safety corridor width corresponding to the asymmetric safety corridor constraint of step 4 are monitored in real time. When it is determined that the preset dangerous state conditions are met, an emergency braking command is output.

[0071] Running synchronously with the domain controller's trajectory optimization process, the independent MCU hardware module performs hardware-level safety fallback control based on globally transmitted TTC metrics. The MCU module operates entirely independently, without relying on upper-layer sensing data or software algorithms, and possesses the highest control priority. Internally, it integrates an independent clock, independent memory, and dedicated braking drive circuitry, monitoring global TTC metrics and safety corridor width in real time. .

[0072] The MCU's default dangerous state determination formula is: unsafe ,in The emergency braking threshold preset in step 5, The minimum safe corridor width is defined (typically 0.5m to 1.0m). If either of these two conditions is met, the MCU immediately determines that the vehicle has entered an extremely dangerous state. This dual-determination mechanism considers both the temporal collision risk and the spatial lateral safety margin, covering a wider range of hazardous scenarios.

[0073] In hazardous conditions, the MCU bypasses the domain controller and directly outputs a hard real-time emergency braking command. The fallback control strategy is as follows: ,in Maximum braking deceleration (typical value) And it forces the vehicle not to perform steering operations, that is This avoids the risk of secondary collisions during emergency braking. Under normal operating conditions, the MCU does not interfere with the main control process, and the vehicle stably executes the optimal MPC instructions output by the domain controller.

[0074] A real-time Life heartbeat communication mechanism is established between the MCU and the autonomous driving domain controller via CAN bus, Ethernet, or an automotive-grade high-speed communication link. The MCU continuously monitors the domain controller's operating status and the health of the control link. The MCU will initiate a Life heartbeat communication when any of the following abnormal conditions are detected: the domain controller fails to send a Life signal beyond the timeout threshold, the TTC indicator stops updating for more than 50ms, or abnormal changes occur in the control command (such as a sudden change in the rate of acceleration exceeding a certain threshold). In the event of communication CRC check failure, the MCU does not need to wait for confirmation from the upper-layer software and directly enters the hardware-level safety takeover state, outputting an emergency braking command within 20ms~50ms to achieve millisecond-level response.

[0075] After the MCU triggers emergency braking, the system will continuously monitor whether the fault has been cleared. When the fault is detected for multiple consecutive cycles (e.g., 10 cycles), the system will continue to monitor whether the fault has been cleared. And the width of the safety corridor At this time, the MCU can gradually exit the takeover state and return control to the domain controller. Simultaneously, the MCU records the security intervention event in its internal non-volatile memory for subsequent diagnosis and analysis.

[0076] This invention employs a dual-redundancy safety architecture, combining MCU hardware as a safety net with domain controller software optimization. Under normal operating conditions, the MCU acts as a passive monitor; in scenarios such as sensor degradation, extreme risks, or main controller failure, the MCU becomes the final executor. Both share the same TTC risk indicators, but their decision-making logic and execution paths are completely independent, thoroughly overcoming the shortcomings of traditional solutions that lack ultimate hardware-level safety protection.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A risk assessment and trajectory optimization method based on millimeter-wave radar, characterized in that, The risk assessment and trajectory optimization method based on millimeter-wave radar includes: Step 1: Collect surrounding target information and vehicle status information and perform standardized processing; at the same time, calculate the global collision time (TTC) risk index and output it synchronously. Step 2: Filter the surrounding target information to obtain a set of valid targets, and normalize the original detection confidence values ​​of the surrounding target information to obtain normalized confidence. Step 3: Based on the vehicle state information and the set of effective targets, calculate the relative motion state between the vehicle and each effective target, calculate the projected collision time for targets with an approaching trend, and introduce the normalized confidence score for weighted correction, and select the minimum weighted TTC as the global risk index. Step 4: Establish reference lines based on the Frenet coordinate system, obtain lane boundary information and calculate the remaining space on the left and right sides; based on the effective target set and the normalized confidence level, construct risk factors on the left and right sides respectively, dynamically adjust the safety corridor boundary, and generate asymmetric safety corridor constraints. Step 5: Based on the global risk index, construct a longitudinal adaptive safety acceleration constraint; Step 6: Integrate the asymmetric safety corridor constraint and the longitudinal adaptive safety acceleration constraint, construct the MPC optimization problem based on the vehicle dynamics model, and use the rolling time-domain optimization strategy to solve the optimal control sequence and output it to the vehicle actuator; Step 7: Monitor the global risk index and the width of the safety corridor corresponding to the asymmetric safety corridor constraint in real time. When it is determined that the preset dangerous state conditions are met, output an emergency braking command.

2. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 1, characterized in that, The surrounding target information mentioned in step 1 includes the original state of each target. and the raw detection confidence value for each target. ,in: The horizontal and vertical positions of the target in the Cartesian coordinate system The target's lateral and longitudinal velocities; The vehicle status information includes the vehicle's longitudinal position. Horizontal position Driving speed Heading angle yaw rate .

3. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 2, characterized in that, The effective target screening described in step 2 is a triple screening: First step: Based on the target's longitudinal position relative to the vehicle. Horizontal position Targets behind and to the sides of the vehicle are removed, and only those that meet the requirements are retained. and The goal, among which The lane expansion coefficient, This represents the current lane width. The second layer filters out targets whose lateral position exceeds the lane width, retaining only targets located within a limited lateral range in front of the vehicle. ; The third step: Eliminating confidence levels Below the minimum confidence threshold The goal is to retain only The goal is to obtain an effective set of objectives. .

4. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 3, characterized in that, The normalization process described in step 2 specifically includes: The original value of the detection confidence level Mapped to interval, i.e. ,in This represents the maximum confidence level of the radar output.

5. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 1, characterized in that, In the relative motion state described in step 3, Record the vehicle status as , The target state is , Then the relative position vector is The relative velocity vector is ; The condition for the approximation trend is: If the dot product of the relative position vector and the relative velocity vector is negative, it indicates that the target is approaching the vehicle. For targets that do not meet the approach trend condition, the direct projection TTC is assigned a value of [value missing]. Avoid invalid calculations; The projection collision time ,in The square modulus of the relative velocity; The minimum weighted collision time (TTC) among all valid targets is selected as the global risk indicator. .

6. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 1, characterized in that, In step 4, the vehicle position is decomposed into longitudinal distances along the reference line based on the Frenet coordinate system. and lateral offset Let the lateral offset of the current reference trajectory be... ; Let the lateral offset of the left lane boundary be... The lateral offset of the right lane boundary is Then the distance from the reference trajectory to the left boundary Distance from the reference trajectory to the right boundary ; Constructing left-side risk factors With right-side risk factors Left-side risk factors ; Right-side risk factors ; in, This represents the set of valid targets located to the left of the vehicle. The set of valid targets located on the right side of the vehicle. The lateral offset of the target relative to the reference trajectory. The longitudinal distance between the target and the vehicle; The longitudinal distance attenuation constant; The risk sensitivity coefficient on the left side; The risk sensitivity coefficient on the right side; Normalized detection confidence for the target; The safety corridor boundary is dynamically adjusted based on the left-hand and right-hand risk factors, wherein: The upper limit of the left boundary of the safety corridor ; The lower right boundary of the safety corridor boundary ; The asymmetric safe corridor constraint is then expressed as: .

7. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 1, characterized in that, Step 5 includes: Preset risk warning threshold With emergency braking threshold ,and ; The global TTC risk indicator is smoothly mapped to an upper bound of safety acceleration using a limiting function. ,in This represents the vehicle's maximum physical acceleration. The longitudinal adaptive safety acceleration constraint is expressed as follows: The lower bound of longitudinal acceleration is fixed at the maximum braking deceleration. .

8. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 1, characterized in that, The vehicle dynamics model described in step 6 is a linear discrete model. , where the state vector Control input The longitudinal acceleration and the front wheel steering angle are given. The objective function of the MPC optimization problem is: ,in To predict the time domain, For reference trajectory state, Here is the state error weight matrix. To control the input weight matrix, ; The MPC optimization problem must simultaneously satisfy: Vehicle dynamics constraints: Furthermore, the control input must meet the physical limit. ; Asymmetric safety corridor constraint: Within each prediction step, the vehicle's lateral offset must lie within the dynamic corridor, i.e. ; Longitudinal adaptive safety acceleration constraints: .

9. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 8, characterized in that, In step 6, a control increment penalty term is added to the objective function of the MPC optimization problem. ,in , This is an incremental weight matrix; When the constraints are not feasible, the slack variable method is used to introduce soft constraints on the asymmetric safety corridor constraints.

10. The risk assessment and trajectory optimization method based on millimeter-wave radar according to claim 1, characterized in that, The preset hazardous condition mentioned in step 7 is unsafe. ,in This is the emergency braking threshold. Minimum safe corridor width; The emergency braking bottom-out control strategy is expressed as follows: That is, using maximum braking deceleration And forced front wheel steering angle .

Images