A vehicle seat active protection method and system based on occupant posture injury prediction

Abstract

The present application relates to the technical field of vehicle occupant safety protection, and discloses a vehicle seat active protection method and system based on occupant posture injury prediction. Multi-modal occupant data is subjected to time-space synchronization and credibility fusion to form an occupant state containing three-dimensional posture, human body parameters, contact relationship and uncertainty; a safe feasible region is generated in combination with a candidate collision scene and seat execution capability, a damage prediction model is used to evaluate the multi-part injury risk under different feasible seats and restraint states, the target state and adjustment trajectory are solved under the remaining collision time, human body clearance, safety belt geometry and mechanical constraint, and re-planning or degradation is carried out according to the posture, execution deviation and collision risk change in the execution process. Thus, the active protection action can take into account individual injury benefits, time accessibility and execution safety.

Description

Technical Field

[0001] This invention relates to the fields of vehicle occupant restraint systems, intelligent seat control, vehicle active safety, computer vision, millimeter-wave sensing, digital human body modeling, and collision biomechanics. In particular, it relates to a vehicle seat active protection method and system that predicts occupant collision damage based on occupant real-time posture, human body parameters, vehicle collision conditions, and seat execution capabilities before a potential vehicle collision occurs, and actively adjusts the state of the seat and related restraint devices within a limited pre-collision time window. Background Technology

[0002] With the development of vehicle electrification, intelligence, and autonomous driving technologies, vehicle interiors are increasingly catering to the needs of driving, working, entertainment, and rest. Vehicle seats are capable of fore-and-aft sliding, overall height adjustment, seat cushion tilt adjustment, backrest angle adjustment, headrest adjustment, leg rest extension, side wing support, and, in some models, seat rotation. This can result in occupants being in non-standard postures such as forward leaning, side leaning, twisting, semi-reclined, pelvic forward sliding, and asymmetrical leg support. Different postures alter the relative positions of the head, neck, chest, pelvis, and lower limbs with the vehicle's internal structure, airbags, and seatbelts, and affect the transmission path of restraint loads during a collision.

[0003] Existing occupant monitoring systems can typically identify whether an occupant is in place and some seating postures, and can also obtain information such as seat belt wearing status. However, discrete posture labels cannot fully describe the three-dimensional spatial relationship between the head, ribcage, pelvis, limbs, and the seat, steering wheel, dashboard, door, and seat belt. On the other hand, protective actions such as seat belt pretensioning, seat straightening, and seat retraction can be used in the pre-collision phase, but fixed action strategies usually cannot simultaneously consider occupant body shape, initial posture, collision direction, expected collision intensity, remaining action time, and the speed and mechanical limits of different seat actuators. If large-scale adjustments are performed based solely on the theoretical target position, situations may arise where the seat is not yet in place when the collision occurs, there is insufficient clearance between the occupant and the interior, the geometric relationship of the seat belt deteriorates, or conflicts occur between actuators.

[0004] Patent document CN110588452A discloses a vehicle seat adjustment method, system, and in-vehicle terminal. It acquires the body state information of an occupant in the target seat and the vehicle's collision perception information. When the collision perception information reaches a preset collision condition, it determines the vehicle motion information corresponding to the collision event and sets seat adjustment parameters such as backrest angle, seat movement distance, or seat cushion depression height based on the vehicle motion information and the occupant's body state information. This technology can implement collision-related seat adjustments for different occupants and different sitting postures. Its main focus is on using the occupant's body state and collision motion information to determine seat adjustment parameters. However, it does not form a complete closed loop in the process of using the occupant's multi-part injury prediction results to back-evaluate multiple candidate seat states and solving the adjustment trajectory within the safe feasible domain jointly defined by the remaining collision time, human body clearance, seat belt geometry, and actuator capability.

[0005] Patent document CN116484627A discloses a method for generating an occupant injury prediction model, an occupant injury prediction method, and a device. It trains the occupant injury prediction model by constructing an autonomous vehicle occupant collision simulation model, parameterized collision characteristic curves, and an occupant injury database. Vehicle feature data, current collision condition feature data, occupant feature data, and constraint configuration feature data are converted into model inputs to obtain the predicted occupant injury level. This technology can incorporate factors such as seat orientation, backrest tilt angle, and constraint configuration in autonomous vehicles into injury prediction. Its main focus is on the generation of the injury prediction model and the output of the injury level. However, it lacks a corresponding active protection control link to address the aforementioned needs, such as how to convert real-time injury prediction results into seat and constraint device adjustment trajectories subject to physical safety constraints and completed within a limited pre-collision time, and how to continuously replan based on changes in occupant posture and collision scenario during execution.

[0006] Furthermore, while high-precision finite element human body models can describe the dynamic responses of various parts of the human body during a collision, the online computation time is often difficult to adapt to short pre-collision time windows. Using only fixed rules or static lookup tables is also insufficient to cover the combined space of multiple body shapes, postures, collision directions, and seat degrees of freedom. Therefore, it is necessary to combine multimodal posture perception, parameterized digital human body, candidate collision scenarios, multi-part damage prediction, safe feasible domain constraints, multi-objective posture optimization, collaborative execution of seats and constraint devices, and closed-loop verification while ensuring real-time performance in vehicles. Summary of the Invention

[0007] The purpose of this invention is to provide a vehicle seat active protection method and system based on occupant posture injury prediction, in order to solve the problem that it is difficult to convert the individual occupant state and collision condition into a seat and restraint device protective action that can be safely performed within a limited pre-collision time under non-standard sitting posture.

[0008] To achieve the above objectives, this invention provides a vehicle seat active protection method based on occupant posture damage prediction, comprising: acquiring occupant multimodal perception data, seat state, and restraint device state to form an occupant state with uncertainty; generating one or more candidate collision scenarios based on potential collision objects and vehicle motion state; when the active adjustment triggering condition is met, generating a safe feasible region based on seat execution capability, estimated remaining collision time, human body gap, and seat belt geometry; using a damage prediction model to predict multi-part damage to candidate seat and restraint device states within the safe feasible region; solving for the target state and segmented adjustment trajectory within the safe feasible region, verifying their reachability, safety, and predicted benefits before execution; continuously updating the occupant state and collision scenario during execution, and maintaining, replanning, stopping, or downgrading accordingly.

[0009] Furthermore, occupant multimodal perception data can include in-vehicle RGB or infrared images, depth data, millimeter-wave human body echoes, seat pressure distribution, seat belt pull-out and buckle status, and seat position in each degree of freedom. Data from each perception branch is synchronized and transformed in a unified clock domain and vehicle coordinate system, and its reliability is determined based on the degree of occlusion, point cloud completeness, radar signal-to-noise ratio, pressure stability, and sensor self-diagnostic results to obtain continuous three-dimensional key points, attitude angles, contact relationships, and state uncertainties.

[0010] Furthermore, based on seat pressure load, visual or depth profile and distance between key points of the human body, the occupant's body shape parameters are estimated. A baseline human body model is selected from the digital human body primitive library, and a parameterized digital human body model is formed through bone segment scaling, body segment mass distribution and soft tissue correction, so that the occupant's current posture and body shape can be used as input for collision damage calculation.

[0011] Furthermore, collision risk prediction can output collision probability, expected collision time, collision direction, relative speed, overlap rate, object category, and expected deceleration pulse based on the trajectory of potential collision objects, vehicle speed, steering, and braking states. When there is uncertainty in the collision direction, intensity, or time, multiple candidate collision scenarios with probability weights are generated.

[0012] Furthermore, the injury prediction model adopts a combination of offline high-precision sample generation and online lightweight prediction to output injury indicators or injury probabilities for the head, neck, chest, abdomen, pelvis and lower limbs, and output the probability of constraint anomalies such as seat belt slippage, lap belt upward movement, occupant descent or non-ideal airbag contact.

[0013] Furthermore, the safe feasible region is jointly defined by seat mechanical limits, speed and acceleration limits, actuator states, multi-degree-of-freedom interference relationships, effective action time, predicted gaps between the human body and interior components, and the positional relationships of the seat belt shoulder strap and lap belt relative to the human body. The safe feasible region is generated before damage optimization, thereby eliminating candidate states that cannot be reached within the remaining time or that pose an execution risk.

[0014] Furthermore, multi-objective attitude optimization simultaneously considers the comprehensive damage risk of multiple parts, the worst-case risk, seat movement amplitude, movement time, occupant disturbance, and model uncertainty. It can be solved using a hierarchical approach involving safe attitude library screening, candidate nearby local optimization, and trajectory reachability verification. As the remaining time before the collision is expected to decrease, the number of seat degrees of freedom involved in the adjustment is dynamically reduced.

[0015] Furthermore, the collaborative control sends commands to the backrest, slide rail, lift, seat cushion, headrest, leg rest or side wing actuators and seat belt reversible pretensioners according to the target state and segmented adjustment trajectory, and provides the updated occupant posture and seat status to the airbag controller; during execution, it determines whether replanning, stopping or downgrading is needed based on the actual seat position, motor current, occupant posture and collision risk.

[0016] Furthermore, the active protection process can consist of normal, prepared, active adjustment, collision hold, risk clearance, and fault degradation states. When multimodal perception or seat position feedback is insufficient to support the current action, the corresponding seat degrees of freedom are reduced or prohibited, while low-risk protective actions verified by the vehicle model are retained; after the collision risk is cleared, the seat state before the event is restored according to the low-speed trajectory when the recovery conditions are met.

[0017] This invention also provides a vehicle seat active protection system based on occupant posture damage prediction, including a multimodal occupant perception module, a spatiotemporal synchronization and credibility assessment module, an occupant posture reconstruction module, a human body parameter and digital human body module, a collision risk prediction module, a damage prediction module, a safe feasible domain generation module, a multi-target posture optimization module, a collaborative control module, and a closed-loop feedback and degradation module. Each module interacts with data in the order of occupant state formation, collision risk characterization, damage inference, safety constraints, optimization decision-making, collaborative execution, and feedback verification.

[0018] The present invention has achieved the following beneficial effects: By combining real-time occupant 3D posture, human body parameters, and candidate collision scenarios, multi-part damage prediction is performed under different candidate seat states, enabling the protection target to change with occupant body size, initial posture, and collision direction. Before damage optimization, a safe feasible region is generated using remaining time, actuator capability, human body clearance, seat belt geometry, and mechanical interference, ensuring that the output target has both reachability and execution safety. Through multi-scenario probability weighting, worst-case risk constraints, and prediction uncertainty handling, a relatively robust protection decision is maintained when the collision direction or intensity is not yet fully determined. During execution, the safe feasible region is continuously re-verified using occupant posture, seat feedback, and collision risk, allowing for replanning or downgrading of occupant active movement, execution deviations, and risk mitigation, while maintaining consistency with the seat belt and airbag control links. Attached Figure Description

[0019] Figure 1 This is a block diagram of the overall composition of the vehicle seat active protection system provided in an embodiment of the present invention.

[0020] Figure 2 A flowchart of a vehicle seat active protection method based on occupant posture injury prediction provided in an embodiment of the present invention.

[0021] Figure 3 A flowchart of multimodal occupant posture recognition and parameterized digital human body construction provided in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the candidate collision scene generation and multi-site damage prediction structure provided in an embodiment of the present invention.

[0023] Figure 5 This is a flowchart of safe and feasible domain generation and multi-target attitude optimization provided for embodiments of the present invention.

[0024] Figure 6 A schematic diagram of the coordinated control timing of the seat, seat belt, airbag and vehicle controller provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. The same or corresponding technical features in each embodiment can be referred to mutually. Without changing the technical relationships described in each embodiment, the sensor type, surrogate model type, and optimization algorithm can be selected based on the vehicle hardware configuration and verified calibration results. Thresholds related to collision triggering, safety limits, actuator speed, and minimum predicted benefit are all determined by vehicle collision simulation, sled testing, seat mechanism testing, hardware-in-the-loop testing, or whole-vehicle testing, and are version-managed in the mass production calibration file.

[0026] Example 1: As Figure 1 and Figure 2As shown, this embodiment presents the overall implementation flow of the vehicle seat active protection method. After the vehicle starts, the cockpit domain controller loads the seat mechanism calibration parameters corresponding to the vehicle model, in-vehicle sensor extrinsic parameters, interior geometric envelope, digital human body primitive library, damage prediction model, and safety posture library, and reads the current state of the backrest, slide rail, height adjustment, seat cushion, headrest, leg rest, side wings, and seat belt reversible pretensioner. During the normal phase, the system continuously acquires the occupant status at a lower update frequency, switching to a higher update frequency when the collision risk increases, to reduce the online computational burden during the pre-collision phase.

[0027] The occupant status is derived from in-vehicle images, depth data, millimeter-wave human body echo, seat pressure array, seat belt sensor, and seat position sensor as inputs. Data from different sources are timestamped or corrected in a unified clock domain and transformed to a unified spatial reference according to the extrinsic parameters between each sensor and the vehicle coordinate system. Subsequently, image occlusion, depth holes, millimeter-wave signal-to-noise ratio, pressure zero drift, and sensor self-diagnostic status are evaluated to obtain corresponding confidence levels. The fusion result provides at least the three-dimensional positions of the head, neck, left and right shoulders, sternum, lumbar spine, pelvis, hip, knee, and ankle; trunk forward or lateral tilt; pelvic pitch; head-neck relative angle; limb joint angles; and the contact relationships between the human body and the seat cushion, backrest, headrest, and seat belt.

[0028] The environmental perception system identifies potential collision objects and predicts the collision probability, estimated remaining time, collision direction, relative speed, overlap rate, object mass or stiffness category, and deceleration pulse based on available information from the forward-facing camera, millimeter-wave radar, and lidar, as well as the vehicle's lateral and longitudinal motion. For situations where there is uncertainty regarding direction, speed, or estimated collision time, multiple candidate collision scenarios are generated near the predicted values ​​and assigned probability weights. When the collision probability reaches a pre-condition, the system increases the perception and calculation frequency and invokes cached safe posture candidates; only when the active adjustment condition is met and the remaining time allows for execution does the actual seat adjustment phase begin.

[0029] Upon entering the active adjustment phase, the system does not directly assign a theoretically optimal seat position based on the damage prediction results. Instead, it first generates a safe and feasible region based on the seat's mechanical limits, actuator speed and acceleration, current motor status, remaining collision time, predicted clearances between the human body and interior components such as the steering wheel or dashboard, and the positional relationship of the seat belt relative to the clavicle and iliac crest regions. The safe and feasible region starts from the current measured seat status and excludes candidate states that exceed mechanical limits, are unreachable within the remaining time, pose a risk of human body clamping or collision along the trajectory, or increase the probability of shoulder strap slippage or lap belt upward movement.

[0030] The damage prediction module inputs the current occupant state, candidate collision scenarios, and candidate seat and restraint device states within the feasible safety domain into the online proxy model, obtaining damage indices or probabilities for the head, neck, chest, abdomen, pelvis, and lower limbs, and outputting prediction uncertainties. The optimization module then integrates the action range, action time, occupant disturbance, and uncertainty to solve for the target backrest angle, slide rail position, headrest position, side wing support amount, or seatbelt pretension amount and its segmented trajectory within the feasible domain. The optimization results are then re-verified for mechanical limits, human body clearance, seatbelt geometry, completion time, and minimum prediction benefit; if any verification fails, the action range is reduced or a conservative strategy is switched.

[0031] In a potential frontal collision scenario, the vehicle is traveling at 70 km / h as indicated in the briefing. The front passenger is slightly leaning forward and looking down at the in-vehicle screen. The system detects a torso lean angle of approximately 22°, and millimeter-wave distance measurement shows a decrease in the distance between the chest and the dashboard. The pressure array indicates that the pelvis remains in the center of the seat cushion. With approximately 1.2 seconds remaining before the collision, the system determines the reachable zone based on the current backrest and seat rail status. It compares the risk of injury to multiple body parts among candidates that meet knee clearance and seatbelt geometry requirements. Ultimately, a combination of seatbelt pretensioning, backrest adjustment towards a more upright position, and seat rail rearward adjustment, along with headrest adjustment, can be used. The approximately 5° backrest adjustment, approximately 34 mm rearward movement of the seat rail, and approximately 15 mm headrest adjustment in the briefing example can be considered as one possible implementation after vehicle model calibration.

[0032] The collaborative control module drives the reversible seatbelt pretensioner and seat actuator in the optimized output sequence, and continuously reads encoder position, motor current, occupant posture, and collision risk. If the occupant actively leans forward, their arms enter the lateral movement area, or the trajectory of the collision object changes significantly during adjustment, the safe and feasible domain is regenerated based on the current measured state, and the trajectory that has not yet been executed is recalculated. If the collision risk is eliminated, further adjustment is stopped and the recovery logic is entered. If the position feedback is abnormal or the actuator stalls, the corresponding degree of freedom is immediately prohibited from continuing to move.

[0033] This embodiment addresses the challenge of fixed seat return strategies simultaneously adapting to individual occupant differences, collision condition variations, and limited action time by organizing occupant status, candidate collision scenarios, multi-site damage prediction, safe feasible domain, and execution feedback into a continuous closed loop. Its technical advantage lies in ensuring that the target state with lower damage is constrained by time accessibility and physical safety before output, and enabling replanning or downgrading when the execution state changes, thereby improving the consistency between active protection actions and actual pre-collision conditions.

[0034] Example 2: As Figure 3As shown, this embodiment illustrates multimodal occupant posture recognition and reliability fusion. An in-vehicle RGB camera, infrared camera, or depth camera is used to obtain the occupant's external contour and key point candidates. Millimeter-wave radar is used to obtain chest-abdomen distance, micro-movements of the human body, and partial limb spatial information. A seat pressure array is used to obtain left and right ischial peak values, pelvic center, back contact boundary, and leg support area. Seatbelt pull-out sensors and buckle sensors are used to characterize the current usage status of the shoulder straps and lap belt. Seat position sensors are used to provide the position of each degree of freedom. All sensor data retains its original timestamp and is mapped to a unified vehicle clock domain before entering the fusion module.

[0035] The camera and millimeter-wave radar are converted to the vehicle coordinate system according to the factory-set extrinsic parameters, while the seat pressure array updates its coordinate transformation relationship based on the seat slide rails, height adjustment, backrest angle, and rotation status. When extrinsic parameter drift caused by long-term vehicle use is detected, online extrinsic parameter correction within a limited range can be performed using known seat geometry, fixed interior features, or stable human contact positions. However, the online correction results do not change mechanical limits and safety parameters. The converted data is interpolated or held at the same fusion time to avoid spatial misalignment of key points caused by different sampling frequencies.

[0036] Each perception branch generates a corresponding confidence level. The vision branch integrates key point confidence, occlusion ratio, and image brightness; the depth branch integrates effective point ratio and hole rate; the millimeter-wave branch integrates signal-to-noise ratio, target clustering stability, and distance continuity; and the pressure branch integrates total load rationality, contact area continuity, and zero-point drift state. Sensor self-diagnostic results serve as a hard constraint on confidence levels; when a branch is deemed faulty, its output is not used in decisions regarding high-risk seat actions.

[0037] Based on the above multimodal fusion results, the occupant posture, human body parameters, contact relationships, and seat and seat belt states at the same fusion moment are uniformly organized, and the occupant state vector can be represented as: ; In the formula, This represents the crew state vector at the current fusion moment; This represents the set of coordinates of three-dimensional key points of the human body organized in a preset key point order, where The keypoint index for the human body is defined by values ​​that increase from 1 to the total number of keypoints involved in the fusion. , Determined based on the definition of key points of the human skeleton used; This represents the set of joint angles and trunk posture angles, where This is the angular feature index, whose value increases from 1 to the total number of angular features. ; A set of contact features between the human body and the seat cushion, backrest, headrest and seat belt, including at least one of the contact area, contact center and the position of the seat belt relative to the human body; This represents the set of human body parameters used for digital human body scaling, including at least one of height, weight, shoulder width, hip width, sitting height, leg length, body type, and soft tissue correction parameters; Indicates at least one of the following: seatbelt buckle, pull-out amount, wearing position, and reversible pretension state; This represents the set of measured positions for each controllable degree of freedom of the current seat; This represents the set of uncertainties corresponding to the aforementioned crew state variables, which can be characterized using covariance, confidence intervals, or normalized uncertainty scores. (Subscript) Used to identify quantities related to occupant status. , and These are used as distinguishing markers for contact features, human body parameters, and seat or constraint state-related quantities, respectively, and are not used as summation indexes.

[0038] In the aforementioned occupant state vector, all items originate from the same fusion moment. The keypoint set is used to calculate the gap between the human body and the interior trim, the posture angle set is used to describe the relative posture of the torso, pelvis, head and neck, and limbs, the contact features are used to describe the contact between the human body and the seat cushion, backrest, headrest, and seat belt, the human body parameters are used for subsequent digital human body scaling, the seat belt state and seat state are used to constrain the injury prediction input, and the state uncertainty is used for risk conservative correction. This avoids directly using discrete labels such as "standard," "forward," and "backward" as the sole control basis.

[0039] When multiple observations exist for the same human keypoint, the fusion module determines the fusion coordinates based on the confidence level of each branch. If the visual keypoint is obscured by a seatbelt, clothing, or arm, it can be completed using millimeter-wave distance, pressure contact boundary, and historical bone segment length. If the completion result causes significant changes in the length of adjacent bone segments within a short period or joint angles exceeding the reachable range of the human body, the keypoint is projected back into the kinematically feasible region of the human body through constraint correction with bone length and joint limits. Subsequently, the keypoint trajectory is smoothed using extended Kalman filtering, unscented Kalman filtering, factor graphs, or validated time-series models, while outputting the position covariance or equivalent uncertainty.

[0040] During normal operation, the attitude fusion update frequency can be calibrated within the range of 5 Hz to 30 Hz, depending on the vehicle's computing power and sensor frequency. During the pre-collision phase, this can be increased to 20 Hz to 100 Hz. The low confidence threshold for sensors can be calibrated within the range of 0.2 to 0.6, incorporating occlusion tests, nighttime tests, and sensor fault injection tests. If the threshold is lowered, the weight of the corresponding branch is reduced or the branch is removed. The above ranges are intended to illustrate an feasible configuration and do not require all vehicle models to use the same frequency or threshold.

[0041] This embodiment converts multi-source sensor outputs into continuous three-dimensional attitude and uncertainty by using time synchronization, coordinate unification, branch reliability assessment, and human kinematic constraints. This solves the problem that a single camera cannot stably provide occupant status data suitable for injury calculation under conditions of occupant obstruction, lighting changes, thick clothing, or vehicle vibration. Its technical advantage lies in providing input with unified spatial semantics and reliability information for subsequent calculation of human body gaps, injury prediction, and generation of the safe and feasible domain.

[0042] Example 3: This example illustrates the estimation of human body parameters and the establishment of a parameterized digital human body model. After posture fusion, the system first extracts features that can be used for body shape estimation from seat pressure load, visual or depth contours, and the relative distances between key points on the human body. Weight can be estimated based on the load at the four corners of the seat or the equivalent total load; shoulder width, hip width, and seat height can be estimated based on the depth contour and known dimensions of the seat; leg length can be estimated based on the bone segment lengths between key points of the hip, knee, and ankle. User-configured age, gender, or vehicle account information can be used as auxiliary information; when not configured, a neutral model is used with increased safety margins, and identity recognition is not a necessary prerequisite for active protection actions.

[0043] The system pre-establishes a digital human body primitive library. Primitives can be categorized by height percentile, weight percentile, and body type. Each primitive includes at least the head, neck, chest, abdomen, pelvis, left and right upper arms, forearms, thighs, and calves, as well as joints between adjacent segments. During online use, the scaling factor for each bone segment is determined based on the ratio of the distance between current keypoints to the length of a reference bone segment. The mass distribution of each body segment is determined based on the total load and anthropometric proportions. Soft tissue envelope correction is determined by combining visual contour and pressure distribution. For visual contour enlargement caused by heavy clothing, historical body type records and pressure loads can be used to constrain the contour, avoiding the direct interpretation of clothing thickness as soft tissue size.

[0044] Parametric digital human models do not require online high-precision finite element analysis (FEM). Their function is to transform information such as occupant body shape, joint posture, external envelope, and segmental mass into unified parameters usable for damage prediction models and safety clearance calculations. Online simulations can employ multi-rigid-body models, reduced-order dynamic models, or validated surrogate models; high-precision FEM human models or dummy models are primarily used for offline sample generation and model calibration. This approach preserves the influence of body shape and posture on collision response while avoiding direct solution of large-scale finite element models within a short pre-collision timeframe.

[0045] To prevent misuse of individual parameters due to occupant switching, the system establishes occupant session identifiers based on seat occupancy status, door status, seat pressure distribution, and boarding events. When a seat changes from vacant to occupied, an occupant re-enters the vehicle after leaving, or the difference between the body shape estimate and the current session parameters reaches the vehicle model calibration update condition, the original session ends and its individualized scaling factor is cleared, retaining only the vehicle model-wide primitive library and safety parameters. The new session re-estimates human body parameters after obtaining stable pressure and posture observations.

[0046] For short-term changes in occupant posture, the system only updates joint angles, key point positions, and contact relationships, avoiding frequent modifications to slow variables such as height and weight. For cases where body shape parameters show stable changes over multiple consecutive normal cycles, the relevant scaling coefficients are then updated smoothly. By separating fast and slow variables, the system avoids misinterpreting actions such as bending over or raising legs as changes in body shape.

[0047] When applied to child safety seats or special rear-seat restraint scenarios, a dedicated child human body primitive can be invoked, and the child seat type, installation orientation, and fixation status can be added to the digital human body and restraint configuration. Only the seat freedom degrees allowed by the vehicle model calibration participate in subsequent optimization; if the child seat fixation relationship does not allow the vehicle seat to be actively adjusted, the corresponding action is directly prohibited in the safe feasible domain.

[0048] This embodiment addresses the technical challenges of using only a standard human body model to cover occupants of different body types, and the difficulty of meeting real-time requirements in online finite element reconstruction, through non-invasive human body parameter estimation, primitive selection, and scaling by bone and body segment parameters. Its technical advantage lies in creating an individualized digital human body representation that can participate in collision damage prediction and human body clearance calculation, and reducing the risk of individual parameter mismatch through occupant session management.

[0049] Example 4: Figure 4 As shown, this embodiment illustrates collision risk prediction and candidate collision scenario generation. The collision risk prediction module receives the relative position and speed of potential collision objects, the vehicle's speed, steering wheel angle, yaw rate, braking pressure, available road surface adhesion information, and planned trajectory. It performs short-term prediction of the object trajectory and calculates the expected collision point, expected collision time, and unavoidable collision probability. The collision direction can be represented by categories such as frontal, left front offset, right front offset, left side, right side, rear, and rollover risk, or it can be represented using continuous azimuth angles in the vehicle coordinate system.

[0050] Based on the collision risk prediction results, the parameters of a single candidate collision scenario used for subsequent damage simulation are organized uniformly, and the candidate collision scenario vector can be represented as: ; In the formula, Represents a single candidate collision scene vector; This represents the collision probability of the candidate collision scenario, and is a dimensionless value between 0 and 1. It represents the remaining time from the current fusion moment to the expected collision moment, expressed in seconds or milliseconds, and uses a unified time unit in the same vehicle model calculation link; The relative speed between the vehicle and the potential collision object is expressed in meters per second or an equivalent speed unit after a standardized conversion. The direction of collision can be indicated by the vehicle coordinate system azimuth or the collision direction category code corresponding to the azimuth. This represents the expected collision overlap rate, a dimensionless proportion between 0 and 1. The mass category or equivalent stiffness category of the potential collision object is represented by the discrete category code in the vehicle calibration file; This represents the expected set of vehicle deceleration pulse parameters, which includes at least one or more of the following parameters: peak deceleration, pulse duration, and pulse shape. This represents the uncertainty of the candidate collision scenario, which can be obtained by combining the uncertainties of the object trajectory, collision direction, relative velocity, and predicted collision time using a unified scale. (Subscript) Used to identify quantities related to collision scenarios, subscript Used to indicate relative motion, subscript Used to identify deceleration pulse parameters.

[0051] In candidate collision scenarios, the collision probability is used to control the entry conditions for the preparatory and active adjustment states, the estimated remaining collision time is used to calculate the effective action time, relative velocity, collision direction, overlap rate, and potential object mass or stiffness category are used together to determine the collision intensity, and the estimated deceleration impulse parameter is used for damage prediction. Scenario uncertainty is used to describe the prediction errors of object trajectory, direction, velocity, and collision timing, and is transformed into a conservative margin in damage prediction and optimization.

[0052] The system employs a tiered triggering and hysteresis mechanism. A first threshold is used to enter the preparatory state, calibrated within the range of 0.3 to 0.7 through playback of hazardous vehicle scenarios. A second threshold is used to enter the active adjustment state, calibrated within the range of 0.6 to 0.95, and the second threshold is kept higher than the first. After the collision probability decreases, an exit threshold below the active adjustment threshold is applied, requiring a predetermined duration before exiting, to reduce frequent seat start-stop cycles caused by probability fluctuations around the threshold. The specific threshold values ​​are determined jointly based on vehicle collision prediction performance, false alarm rate, and actuator risk.

[0053] When environmental perception can only provide a single nominal collision direction, a candidate scene can be directly formed. When the uncertainty of direction, velocity, or predicted collision time is large, multiple scenes are formed around the predicted distribution. For example, unscented transformations, finite-number Monte Carlo sampling, scene tree expansion, or discrete binning can be performed on the collision azimuth, relative velocity, and predicted collision time, and approximate scenes can be merged to keep the number of scenes within the limits of onboard computing resources. The probability weights of each scene are obtained based on the output probability of the perception model, sampling weights, or historical calibration statistics. The weights of all retained scenes are normalized before entering the optimization process.

[0054] The deceleration impulses can be generated from vehicle collision databases by selecting similar templates based on potential object type, relative velocity, and overlap rate, and then scaling them. Alternatively, they can be generated from lightweight collision dynamics models that have been validated through vehicle testing or simulation. For cases where the stiffness of potential objects is difficult to determine, discrete impulse scenarios such as softer, nominal, and stiffer impulses can be generated to ensure that subsequent optimization does not rely on a single assumption about object stiffness.

[0055] This embodiment addresses the issues of uncertainty in predicting collision direction, intensity, and timing during the pre-collision phase, and the tendency for a single nominal condition to cause target deviation, through tiered triggering, hysteresis exit, and probabilistic weighted multi-scenario construction. Its technical advantage lies in providing damage prediction and robust optimization with a set of scenarios that simultaneously reflect both expected and unfavorable collision conditions, and directly constraining the accessibility of subsequent seat actions using the estimated remaining collision time.

[0056] Example 5: This example illustrates the construction and online simulation of the damage prediction model. In the offline phase, the sample space is first defined. Sample factors include at least occupant body shape, initial posture, seat degrees of freedom, collision direction, relative velocity, overlap ratio, seatbelt position, airbag configuration, and vehicle interior geometry. High-precision simulation samples can be selected using orthogonal experiments, Latin hypercube methods, or active learning to reduce duplicate samples. Each offline sample obtains the response of various parts of the human body through finite element human body models, dummy models, trolley tests, or multibody dynamics simulations, and checks energy balance, time step stability, and contact anomalies before being included in the training set.

[0057] The online proxy model can select one of the following based on the vehicle development conditions: gradient boosting tree, random forest, Gaussian process, radial basis function network, multilayer perceptron, temporal convolutional network, graph neural network, or physically constrained neural network. For ease of calibration, a sub-model can also be used to predict indicators for the head, neck, chest, abdomen, pelvis, and lower limbs separately. The model input consists of occupant status, digital human body shape parameters, candidate collision scenarios, seat and restraint device status, and input uncertainty; the output includes at least the available items from HIC, BrIC, Nij, chest 3 ms acceleration, chest compression, equivalent abdominal load, pelvic acceleration, femoral axial load, tibial indicators, and corresponding injury probabilities.

[0058] In addition to biomechanical injury outputs, the model can also output constraint anomaly probabilities, including shoulder girdle slippage relative to the clavicle region, lap belt upward displacement towards the abdomen, occupant descent, excessive forward head movement, and deviations in airbag contact timing or position. Different outputs are first normalized to a uniform risk scale based on regulatory limits, injury risk curves, or validated corporate safety objectives before being used for comprehensive risk calculations.

[0059] After normalizing the various biomechanical damage and constraint abnormality risks using a unified risk scale, the comprehensive damage risk can be calculated using the following formula: ; In the formula, The overall damage risk is represented by a dimensionless evaluation value obtained after normalizing each damage risk according to a unified scale. The total number of damage risk items included in the current working condition comprehensive evaluation is determined by the head, neck, chest, abdomen, pelvis, and lower limb injury indicators and restraint abnormality risk items actually used in the current vehicle model. This represents the damage risk item index, with values ​​ranging from 1 to... ; Indicates the first The normalized risk value is the result of mapping the safety target verified by regulations, damage risk curves, or vehicle models to a unified risk scale. Indicates the first The non-negative weights of each damage risk The sum after normalization is 1; This represents the non-negative penalty coefficient indicating the risk of the worst-case scenario. Indicates in all The maximum value is taken from the normalized damage risk.

[0060] The weights of each part in the formula are determined based on the collision direction and occupant category, with the total weights remaining constant at 1. The worst-case penalty coefficient is used to prevent optimization from reducing the risk of one part while significantly increasing the risk of another. For example, in a side-impact scenario, the weights of the proximal chest and pelvis can be increased, while in a rear-end collision scenario, the weights of neck-related risks can be increased. The weight changes only affect the optimization evaluation and do not modify the validated calculation methods for individual damage indicators.

[0061] The surrogate model simultaneously outputs prediction uncertainty or an equivalent confidence interval. Training, validation, and test sets are divided by scenario to avoid similar samples from the same collision condition being included in both training and testing simultaneously; boundary scenarios such as extreme pitching, larger body sizes, children, and side impacts have separate validation samples. Before deployment, upper bounds on the error and physical monotonicity are checked; for example, under similar conditions, an increase in relative collision velocity should not cause a systematic decrease in predicted damage. When the input exceeds the training coverage, the model outputs an out-of-distribution flag, and the control side increases the risk margin or switches to a validated conservative posture.

[0062] This embodiment addresses the challenges of high-precision human collision models requiring repeated online calculations within short pre-collision timeframes, and the difficulty of simple table lookups covering body shape, posture, and collision combination space, by employing a division of labor between offline high-precision samples and an online lightweight proxy model. Its technical advantage lies in its ability to rapidly compare the damage risks of multiple body parts among candidate seat states, while simultaneously transferring model uncertainty and constraint anomaly risks to subsequent optimization and degradation decisions.

[0063] Example 6: As Figure 5 As shown in this embodiment, the safe and feasible region is generated. The safe and feasible region is not simply a truncation of the results after optimization, but rather formed before damage optimization begins by layering mechanical feasibility, time availability, human body clearance, and seatbelt geometric constraints, starting from the current measured state of the seat and the restraint device. Only candidate target states and adjustment trajectories within the intersection of various constraints are included in the damage evaluation and optimization calculation.

[0064] Mechanical feasibility constraints consist of the minimum and maximum positions, maximum permissible velocities, maximum permissible accelerations, jerk limits when necessary, motor current and temperature limits, and mechanical interference relationships between multiple degrees of freedom. For example, when the leg rest is extended, the forward movement of the seat can be limited based on the mechanism envelope; when the seat rotates away from the forward direction of the vehicle, the angle range for the backrest to quickly return to center can be limited. Mechanical constraints are derived from seat design parameters, mechanism testing, and controller fault diagnosis parameters, and are updated based on the current actuator temperature and fault status.

[0065] The time reachability constraint is calculated by subtracting communication delay, computation delay, actuator setup time, and safety margin from the estimated remaining collision time to obtain the effective action time. The safety margin can be determined within the range of 50 ms to 300 ms, based on the briefing recommendations, combined with network latency, controller computation cycle, and actuator response test calibration. For each adjustable degree of freedom, the reachable position range within the effective action time is calculated based on the current measured position, permissible velocity, and acceleration. When multiple degrees of freedom operate in parallel, the power supply capacity or the controller's limitation on the number of simultaneous drives must also be considered.

[0066] Human body clearance constraints are determined using the minimum distance between the parametric digital human body outer envelope and interior surfaces such as the steering wheel, dashboard, center console, doors, B-pillar, roof, and adjacent seats. Candidate trajectories are sampled at discrete time points or approximated with continuous envelopes. Trajectories are excluded if the predicted clearance of any body part on the trajectory is less than the minimum clearance threshold corrected for human body model errors and sensor uncertainties. For cases where the occupant's arms, hands, or legs are close to seat side wings, slide rails, leg rests, or other motion mechanisms, the corresponding actuator torque is further restricted or that degree of freedom of movement is prohibited.

[0067] The seatbelt geometric constraints are determined based on the position of the shoulder belt relative to the clavicle region, the position of the lap belt relative to the iliac region, belt torsion, remaining retractor travel, and the direction of buckle force. If a candidate seat movement increases the probability of the lap belt shifting upwards to the abdomen, causes the shoulder belt to deviate from the clavicle region, or reduces retractor travel, it will not enter the safe and feasible region, even if a predicted individual injury metric decreases. This constraint is updated synchronously with the seat back angle, slide rail position, seat height, and occupant pelvic position.

[0068] The range of a single seat adjustment can be further limited based on the vehicle model's actuator capabilities and remaining time. The feasible ranges provided in the specifications include a single backrest angle adjustment of 0° to 20°, a single slide rail adjustment of 0 to 100 mm, a headrest quick adjustment of 0 to 40 mm, and an upper limit of the equivalent clamping force of the side wings of 20 N to 120 N. Specific values ​​are determined through seat mechanism testing, human comfort and clamping tests, and crash simulations. During operation, vehicle model calibration values ​​are used instead of fixed usage range endpoints.

[0069] This embodiment addresses the problem that seat targets with theoretically low damage may be unable to execute safely due to insufficient remaining time, insufficient human clearance, or deterioration of constraint geometry by establishing the intersection of mechanical, temporal, human body, and seatbelt constraints before damage optimization. Its technical advantage lies in limiting the optimization search to a set of states that the actual actuator can reach and whose trajectory process satisfies human safety conditions, reducing target distortion and mid-motion termination caused by truncation after optimization.

[0070] Example 7: As Figure 5As shown, this embodiment further illustrates multi-objective attitude optimization within the safe and feasible domain. Optimization variables include both the desired target state of the seat and restraint devices before a collision, and the start time, speed limit, sequence of actions, and segmented trajectory from the current measured state to the target state. The degrees of freedom that can participate in optimization include the vehicle-vehicle-available portions of the backrest angle, seat cushion tilt angle, slide rail position, seat height, headrest position, leg rest position, seat rotation angle, seatbelt reversible pretension, and side wing support.

[0071] By unifying the seat degrees of freedom and reversible constraint device target quantities that can participate in active protection in the current vehicle model, the control vectors for the seat and constraint devices can be expressed as follows: ; In the formula, This represents the control vector for the candidate seat and the reversible constraint device. Indicates the backrest angle; Indicates the seat cushion tilt angle; Indicates the front and rear position of the seat along the slide rail; Indicates seat height; This indicates the headrest height or fore-aft position; the specific direction is fixed by the vehicle model's calibration documents. Indicates the extension of the leg rest; This indicates the angle of rotation of the seat relative to the longitudinal reference direction of the vehicle; The target amount of the reversible seat belt pretensioner can be expressed as pretension displacement or pretension force, depending on the actuator type. The target support quantity of the flank actuator can be expressed as flank displacement or equivalent support force, depending on the actuator type. Angular quantities are uniformly expressed in degrees or radians within the same calculation link, while position and displacement quantities are uniformly expressed in millimeters or meters, and forces are expressed in Newtons. Control quantities of different dimensions are normalized dimensionlessly according to the calibration allowable range of the corresponding degree of freedom before entering the optimization calculation.

[0072] When the vehicle model lacks a certain degree of freedom, the corresponding variable is removed from the control vector and the safe feasible region, rather than participating in the calculation as a virtual actuator. The system first filters a small number of baseline candidate states from the safe attitude library based on the collision direction, occupant body shape, and current attitude. Then, it performs continuous variable local optimization near the candidate states. Finally, it performs reachability and collision verification on the obtained segmented trajectory. This hierarchical solution method can utilize pre-calculation results from the normal phase to narrow the online search range for the hazardous phase.

[0073] When comprehensively evaluating the candidate target state and its segmented adjustment trajectory within the safe and feasible domain, the comprehensive optimization target can be calculated using the following formula: ; In the formula, This represents the comprehensive optimization objective used to compare the candidate target state and the adjustment trajectory. The smaller the value, the better the comprehensive evaluation under the current constraints. This represents the damage risk term obtained from the aforementioned comprehensive damage risk formula; The action range and energy consumption terms are obtained by normalizing the normalized displacement, velocity or acceleration requirements of each participating degree of freedom from the current state to the candidate target state and the estimated execution energy consumption. This indicates the action time and timeout penalty, which is determined based on the relative time taken up by the candidate trajectory's estimated completion time compared to the effective action time. A timeout penalty is added when the estimated completion time exceeds the effective action time. The penalty for occupant disturbance and safety and comfort is determined based on at least one of the following: attitude change caused by candidate trajectory, minimum clearance between human body, clamping risk, and actuator resistance state. The uncertainty penalty term is determined based on at least one of the following: occupant state uncertainty, candidate collision scenario uncertainty, damage model prediction uncertainty, and out-of-distribution flag. , , and These represent the non-negative weights of the action amplitude and energy consumption, action time and overtime penalty, occupant disturbance and safety / comfort penalty, and uncertainty penalty, respectively. Before being weighted, each objective item is converted to a uniform dimensionless evaluation scale. Each weight is calibrated by offline vehicle simulation, seat bench, hardware-in-the-loop, and whole vehicle verification, and there is no need to limit the sum of the four weights to 1.

[0074] The comprehensive objective includes a motion amplitude and energy consumption term to avoid unnecessary large movements, a motion time term to penalize trajectories that approach or exceed the effective motion time, a occupant disturbance term to limit abrupt attitude changes, clamping, and significant comfort shocks, and an uncertainty term to reduce the priority of aggressive actions when sensor reliability decreases or damage models are extrapolated. All weights are jointly calibrated through vehicle simulation, seat bench testing, hardware-in-the-loop testing, and vehicle verification, and safety control parameters are not directly modified via online self-learning.

[0075] When there are multiple candidate scenarios for collision direction, collision intensity, or expected collision time, the robust optimization objective used to compare candidate control schemes can be calculated using the following formula: ; In the formula, This represents the robust optimization objective used to compare candidate control schemes in a multi-candidate collision scenario; This represents the total number of candidate collision scenarios that are currently retained and participating in the robustness evaluation; This represents the candidate collision scene index, with values ​​ranging from 1 to... ; Indicates the first The normalized probability weights of each candidate collision scene are dimensionless values ​​between 0 and 1, and the sum of the normalized probability weights of all candidate scenes is 1. Indicates the first The evaluation value calculated for the same candidate control scheme under the aforementioned comprehensive optimization objective in a candidate collision scenario; This represents the non-negative weight of the worst-case candidate collision scenario item; Indicates all of the current situation Candidate collision scenarios Take the maximum value. When only one candidate collision scenario is retained, the aforementioned comprehensive optimization objective can be directly used for evaluation without performing multi-scenario robust aggregation.

[0076] The multi-scenario robust objective considers both high-probability and low-probability but severe-consequence scenarios. The probabilities of each candidate collision scenario are normalized before being used in the expected risk calculation, with the worst-case scenario weight used to control sensitivity to adverse scenarios. If a candidate state yields good returns in a high-probability scenario but significantly increases the risk to critical parts in another retained scenario, the worst-case scenario term will have its priority reduced. This avoids over-optimization on a single prediction result before the collision direction or intensity has fully converged.

[0077] The estimated remaining time after a collision is also used to dynamically select the degrees of freedom (DOFs) for participating actions. When the remaining time is sufficient, combinations of multiple DDFs such as backrest, slide rail, headrest, and side wings can be considered; as the remaining time decreases, priority is given to retaining reversible seatbelt pretensioning, headrest, or minor backrest actions that have been verified by the vehicle model and have a fast response time; when the effective action time is insufficient to complete any seat adjustment, no new seat movement commands will be issued. The optimization results must also meet a minimum predicted benefit threshold. The briefing suggests that the benefit should be determined based on vehicle model verification within the range of a 2% to 15% reduction in overall damage risk; if the benefit is below the threshold, the corresponding seat action should be canceled.

[0078] This embodiment addresses the issues of large multi-seat degree-of-freedom combination space, unsustainable optimization results in single collision scenarios, and the possibility of insufficient time to complete actions under short-term conditions by employing methods such as safety posture library selection, local continuous optimization, joint evaluation of expected and worst-case risks across multiple scenarios, and time-driven degree-of-freedom selection. Its technical advantage lies in prioritizing targets and trajectories with overall damage benefits and robustness to collision prediction errors from the set of executable states.

[0079] Example 8: As Figure 6As shown, this embodiment illustrates the coordinated execution and online replanning between the seat, seat belt, airbag, and vehicle controller. The multi-target attitude optimization module outputs the target seat state, the target seat belt reversible pretension, the action sequence of each actuator, speed limits, estimated completion time, and predicted benefits. The coordinated control module converts the segmented adjustment trajectory into position or speed commands recognizable by actuators such as the backrest, slide rail, lift, seat cushion, headrest, leg rest, and side wings, and reads the fault status and current measured position of each actuator again before issuing the commands.

[0080] The timing of actions is determined based on the interaction between the restraint devices and seat movements. For example, in a frontal collision involving a lurching occupant, reversible seatbelt pretensioning can be applied first to stabilize the pelvis and torso, followed by minor backrest and slide rail adjustments. When the headrest position needs to be changed, the relative sequence of headrest and backrest movements is determined based on the head-neck clearance and actuator speed. In a side-impact scenario, the coordinated timing of side wing support and seatbelt pretensioning can be determined based on the risk to the proximal chest and pelvis. All actions are constrained by the human body clearance and actuator limits within the safe feasible domain.

[0081] During execution, the seat position sensor continuously reports the actual position, the motor controller reports current, temperature, and stall status, the occupant perception chain continuously updates key points and contact relationships, and the collision risk prediction module continuously updates the estimated remaining collision time and candidate scenarios. If the deviation between the actual position and the planned trajectory exceeds the vehicle's calibrated allowable value, or if the occupant's actions such as bending over, turning around, or raising their hand change the body clearance, the system pauses the relevant degrees of freedom and regenerates the safe and feasible domain based on the current measured state.

[0082] Replanning does not recalculate from the initial state of the event; instead, it uses the current executed position, current occupant attitude, current collision scenario, and remaining time as a new starting point. Actions that are no longer reversible or have already been completed serve as new boundary conditions, while trajectories that have not yet been executed are re-optimized based on the updated safe and feasible domain. This prevents the controller from continuing to execute based on outdated, invalid assumptions. If the collision risk increases rapidly, leading to a significant reduction in remaining time, the system reduces the number of available degrees of freedom and prioritizes actions with faster responses.

[0083] The collaborative control module simultaneously sends updated occupant posture, seat position, occupant body type, and restraint status to the airbag controller, enabling the airbag controller to select trigger parameters using its own validated strategy. The active protection system does not replace the airbag controller's independent safety judgment, nor does it directly modify deployment logic that has not been verified in the vehicle model; its role is to maintain consistency in the perception of the current occupant status and seat position among different controllers. The vehicle braking or autonomous driving controller can provide the active protection system with information on braking intervention and changes in collision risk to update the active protection status.

[0084] In the side-impact example, if the occupant's torso tilts slightly to the right and the system predicts a higher risk to the proximal chest and pelvis, a combination of proximal wing tightening, slight seat return, and seatbelt pretensioning can be considered within the safe and feasible domain. If pressure or posture sensing detects that the occupant's arm is between the wing and the door, the wing action is canceled, and only the backrest or seatbelt strategy that satisfies other constraints is retained. This process allows the execution layer to perform real-time physical verification of the optimization objective.

[0085] This embodiment solves the problems of separate actions by the seat, reversible seat belt, and airbag-related controllers based on different occupant states, and the failure of the original trajectory due to occupant's active movement during execution, by unifying the action timing, actuator feedback, and online replanning starting from the current state. Its technical advantage lies in enabling active protection commands to be corrected in a timely manner according to actual execution deviations and changes in occupant spatial relationships, while maintaining state coordination between different restraint devices.

[0086] Example 9: This example illustrates the state management, failure degradation, and risk mitigation recovery of active protection events. Under normal conditions, the system continuously updates occupant body shape, habitual posture, seat reference position, and sensor health, but does not drive active protection actions. When the collision probability reaches a first threshold, it enters a preparatory state, increasing the frequency of perception and damage prediction updates and preparing candidate safe postures. When a second threshold is reached and the estimated remaining collision time is within the executable range, it enters an active adjustment state. Upon confirmation of a collision, new posture optimization ceases, and the system maintains or completes permitted actions according to a validated strategy.

[0087] Perception degradation is graded based on the degree of observability. When only the RGB image's reliability decreases due to nighttime backlighting, the weight of the RGB branch is reduced, and infrared depth, millimeter-wave distance, and pressure arrays are used to continue recovering the torso posture. When the uncertainty of key points increases, damage prediction is given a conservative margin and the range of seat movements is limited. If the depth camera further fails and the remaining perception cannot reliably support fine 3D posture, multi-degree-of-freedom optimization relying on fine posture is stopped, and only the reversible pretensioning of the seat belt or headrest movement verified by the vehicle model is retained.

[0088] Actuator degradation is prioritized to prevent further uncontrollable motion. If a seat position sensor fails, the corresponding degree of freedom is prohibited from continuing active movement. If the motor current continuously exceeds the stall threshold or the temperature reaches the protection condition, the actuator is immediately stopped, and the current position is used as the new boundary condition. Whether other degrees of freedom continue to execute is determined by the updated safe feasible domain. In the event of abnormal seat belt pretensioner or communication link status, the system does not assume that the action has been completed, but recalculates the available protection strategy based on the measured status.

[0089] Active resistance by the occupant is also considered a degradation condition. When the system detects continuous reverse movement, hands or feet near the moving mechanism, or pressure distribution indicating a clamping tendency, it first limits the torque or speed of the corresponding motor and issues audible, tactile, or visual cues; if the clamping risk still exists, it stops the corresponding degree of freedom. At this point, subsequent optimizations will no longer treat this degree of freedom as an available variable, thus avoiding repeatedly issuing the same unexecutable action.

[0090] When automatic emergency braking successfully avoids a collision or a potential collision object leaves the predicted path, and the collision probability drops below the exit threshold for a predetermined time, the system enters a risk-relief state. A 2-second duration can be used as a calibration example. After risk relief, the system first removes any unnecessary reversible pretensioning, alerts the occupant to the change in risk status, and maintains the current child safety seat position. Only upon occupant confirmation or when vehicle model policy allows for recovery does the system return to the pre-event recorded comfort position at a lower speed. Human clearance and actuator status are monitored throughout the recovery process.

[0091] After the event, anonymized perception features, candidate scenarios, prediction results, execution trajectories, fault states, and event outcomes are recorded for offline model playback, version verification, and subsequent calibration. New models or calibration parameters must be deployed after simulation regression, hardware-in-the-loop testing, seat bench testing, and vehicle testing. Online data acquisition does not directly change seat safety limits, collision trigger thresholds, or airbag-related safety parameters.

[0092] This embodiment addresses the issues of potential secondary risks arising from partial sensor or actuator failure, occupant voluntary movement, and continued execution of the original actions after a collision risk has been eliminated, through controlled transitions between normal, standby, active adjustment, collision hold, risk clearance, and fault degradation. Its technical advantage lies in enabling the active protection system to reduce its functional boundaries when information is insufficient or execution conditions change, and to exit and resume in a controlled manner after the risk has been cleared.

[0093] Example 10: This example illustrates the module composition and data connection relationships of a vehicle seat active protection system. The system includes a multimodal occupant perception module, a spatiotemporal synchronization and reliability assessment module, an occupant posture reconstruction module, an anthropomorphic parameter and digital human body module, a collision risk prediction module, a damage prediction module, a safe feasible domain generation module, a multi-target posture optimization module, a cooperative control module, and a closed-loop feedback and degradation module. These modules can be deployed in the cockpit domain controller, central computing platform, or independent safety controller, or distributed across multiple controllers according to computing power and functional safety levels.

[0094] The multimodal occupant perception module connects to in-vehicle cameras, depth sensors, millimeter-wave radar, seat pressure arrays, seat belt sensors, and seat position sensors, outputting timestamped perception data via in-vehicle Ethernet, CAN FD, LIN, or a dedicated serial bus. The spatiotemporal synchronization and reliability assessment module receives the perception data and performs time interpolation, extrinsic parameter transformation, outlier suppression, and reliability calculation. The occupant posture reconstruction module, based on the synchronized multi-source data, outputs 3D human body key points, posture angles, contact relationships between the human body and the seat and seat belt, and state uncertainties.

[0095] The human body parameter and digital human body module receives posture reconstruction results, seat pressure load, and optional account configurations. It builds a parameterized digital human body model from the primitive library and provides the human body's external envelope, bone segment parameters, and body type to the safety feasible domain generation module and the damage prediction module. The collision risk prediction module obtains the trajectory of potential collision objects, the vehicle's status, and planning information from the vehicle's environmental perception and motion control links. It then forms one or more candidate collision scenarios and outputs the collision probability and the estimated remaining time before the collision.

[0096] The safe feasible region generation module connects seat mechanism calibration data, actuator fault status, in-vehicle geometric envelope, human body model, and seat belt geometry information, generating a set of currently allowed search and execution seat and restraint device states in each optimization cycle. The damage prediction module receives occupant status, digital human body parameters, candidate collision scenarios, and candidate seat states, outputting multi-site damage risk, constraint anomaly probability, and prediction uncertainty. The multi-objective attitude optimization module searches only within the safe feasible region, obtaining target state, segmented trajectory, predicted gain, and confidence level.

[0097] The collaborative control module converts the trajectory into control commands for the seat back, slide rails, height adjustment, seat cushion, headrest, leg rest, side wings, and reversible seatbelt pretensioners, and exchanges current states and action windows with the airbag controller, brake controller, or autonomous driving domain controller. The closed-loop feedback and degradation module continuously compares the target state with the measured state, updates the remaining time, occupant posture, and collision risk; when the replanning conditions are met, it feeds back the current state to the safe feasible domain generation module and optimization module; when the fault or exit conditions are met, it executes stop, hold, rollback, or conservative strategies.

[0098] To balance computing power and safety levels, sensing, digital human body, and optimization can be implemented on a higher-powered controller, while final action authorization, mechanical limits, and actuator fault verification can be placed in a control unit with independent safety protection. Even if the upper-level optimization output is abnormal, the lower-level authorization unit will not issue the corresponding action if the target or trajectory exceeds the calibrated mechanical, temporal, human body gap, or seat belt geometric boundaries.

[0099] This embodiment addresses the challenge of unifying data semantics and control states when occupant perception, damage prediction, seat optimization, and restraint device execution are handled by different functional modules, by clearly defining the input, output, and feedback directions of each module. Its technical advantage lies in establishing a complete system connection from occupant state and collision scenario input, to the safe feasible domain and damage assessment, and then to collaborative execution and closed-loop degradation, facilitating deployment according to functional safety requirements on different vehicle electronic and electrical architectures.

Claims

1. A vehicle seat active protection method based on occupant posture injury prediction, characterized in that, include: Acquire occupant multimodal perception data, seat status, and constraint device status on vehicle seats; perform time synchronization, coordinate transformation, anomaly handling, and reliability fusion on the occupant multimodal perception data to obtain occupant status including occupant three-dimensional key points, posture angles, human body parameters, contact relationships, and state uncertainties. At least one candidate collision scenario is generated based on the motion information of the potential collision object and the vehicle's motion state. The candidate collision scenario includes the collision probability, the estimated remaining time of the collision, the collision direction, the collision intensity, and the estimated deceleration pulse. When the active adjustment triggering conditions are met, a safe feasible domain for the seat and restraint device is generated based on the current state of the seat actuator, the estimated remaining time of the collision, the seat mechanical constraints, the human body and the vehicle interior environment gap constraints, and the seat belt geometric constraints. The occupant state, the candidate collision scenario, and the candidate seat and restraint device states within the safe and feasible domain are input into the damage prediction model to obtain the occupant multi-part damage risk and prediction uncertainty corresponding to each candidate state. Based on the aforementioned multi-part damage risk and combined with action time, action amplitude, occupant disturbance and prediction uncertainty, the target seat and restraint device state and segmented adjustment trajectory are solved within the safe and feasible domain, and the accessibility, safety and predicted benefits of the segmented adjustment trajectory are verified. After verification, a collaborative control command is issued to the seat actuator and reversible constraint device. During execution, the occupant status and candidate collision scenarios are continuously updated, and a safe and feasible domain is regenerated based on the update results to maintain, replan, stop, or downgrade the segmented adjustment trajectory.

2. The vehicle seat active protection method according to claim 1, characterized in that, The occupant multimodal perception data includes in-vehicle images, depth data, millimeter-wave human body echoes, seat pressure distribution, seat belt pull-out and buckle status, and seat position in each degree of freedom. The credibility fusion includes: determining the credibility of each perception branch based on occlusion ratio, image brightness, depth point cloud integrity, millimeter-wave signal-to-noise ratio, pressure distribution stability, and sensor self-diagnosis results; converting the occupant key points output by each perception branch to a unified vehicle coordinate system and then fusing them according to the credibility; and using human bone segment length constraints and joint range of motion constraints to correct key points that violate human kinematic constraints, so as to output continuous occupant three-dimensional posture and its uncertainty.

3. The active protection method for vehicle seats according to claim 2, characterized in that, Based on seat pressure load, occupant visual or depth profile, distance between key points on the human body, and optional vehicle account information, the occupant's height, weight, shoulder width, hip width, sitting height, leg length, and body type are estimated. A baseline human body model is determined from a preset digital human body primitive library. A parameterized digital human body model is generated according to bone segment scaling coefficient, body segment mass distribution coefficient, and soft tissue correction coefficient. When a switch in occupant session is detected or a change in body type estimation meets the update conditions, the individualized parameters of the previous occupant session are cleared and the parameterized digital human body model is rebuilt.

4. The active protection method for vehicle seats according to claim 1, characterized in that, The candidate collision scenarios are constructed based on the collision direction, relative velocity, overlap rate, potential collision object category, expected collision time, and expected deceleration pulse. When the collision probability reaches a first threshold, the scenario enters a preparatory state and pre-calculates candidate safe postures. When the collision probability reaches a second threshold higher than the first threshold and the expected remaining collision time is within the executable range, the scenario enters an active adjustment state. For cases where there is uncertainty in the collision direction, collision intensity, or collision time, multiple candidate collision scenarios with probability weights are generated around their predicted distribution.

5. The active protection method for vehicle seats according to claim 3, characterized in that, The damage prediction model is trained using offline high-precision collision simulation data, trolley test data, and / or multibody dynamics simulation data. The inputs include at least the occupant state, the body shape parameters of the parameterized digital human body model, candidate collision scenarios, seat and restraint device states, and input uncertainties. The outputs include at least the damage indicators or damage probabilities corresponding to the head, neck, chest, abdomen, pelvis, and lower limbs, as well as the restraint anomalies of seat belt slippage, lap belt upward movement, occupant descent, or non-ideal airbag contact. The damage outputs are normalized to form a comprehensive damage risk, and a conservative margin is added to the prediction uncertainty or inputs that exceed the coverage of the training data.

6. The vehicle seat active protection method according to claim 1, characterized in that, The safe feasible domain is generated by the intersection of the following constraints: mechanical feasibility constraints formed by the mechanical limits of each degree of freedom of the seat, the maximum permissible speed, the maximum permissible acceleration, the motor current, the temperature, and the mechanical interference relationship of multiple degrees of freedom; time reachability constraints formed by deducting communication delay, calculation delay, actuator setup time, and safety margin from the estimated remaining collision time; human body clearance constraints formed by the parametric digital human body outer envelope and the predicted clearance between the steering wheel, dashboard, center console, door, body pillar, roof, or adjacent seats; and seat belt geometric constraints formed by the position of the shoulder belt relative to the clavicle area, the position of the lap belt relative to the iliac crest area, belt torsion, retractor margin, and buckle force direction.

7. The active protection method for vehicle seats according to claim 1, characterized in that, The target seat and restraint device state and segmented adjustment trajectory are solved using a hierarchical optimization method, including: selecting baseline candidate states from a safety attitude library based on the collision direction, occupant body type, and current posture; locally optimizing the available degrees of freedom in the backrest angle, seat cushion tilt angle, slide rail position, seat height, headrest position, leg rest position, seat rotation angle, seat belt reversible pretension, and side wing support near the baseline candidate states; for multiple candidate collision scenarios, a robust target is jointly determined by the damage risk weighted by the probability of each scenario and the damage risk of the worst scenario, and the seat degrees of freedom involved in the adjustment are dynamically reduced based on the estimated remaining collision time; when the predicted damage benefit is lower than the minimum predicted benefit threshold, the corresponding seat action is canceled or switched to a preset conservative protection strategy.

8. The active protection method for vehicle seats according to claim 1, characterized in that, The coordinated control commands include the target position, speed limit, and action sequence of the seat back, slide rail, lift, seat cushion, headrest, leg rest, or side wing actuators, as well as the pretensioning target of the seat belt reversible pretensioner. During execution, seat position feedback, motor current, occupant posture, seat belt status, and collision risk are collected. The target state is compared with the measured state. When execution deviation, occupant active resistance, changes in the collision scenario, or changes in human body clearance are detected to meet the replanning conditions, the safe and feasible domain is regenerated based on the current measured state, and the subsequent adjustment trajectory is updated. At the same time, the updated occupant posture and seat status are provided to the airbag controller.

9. The active protection method for vehicle seats according to claim 8, characterized in that, The active protection process is managed according to the following states: normal, prepared, active adjustment, collision hold, risk elimination, and fault degradation. When the credibility of perception decreases, the weight of low-credibility perception branches is reduced. When multimodal perception is insufficient to support fine attitude estimation, large-scale seat adjustment is stopped and the reversible seat belt pretensioning or headrest action verified by the vehicle model is retained. When abnormal seat position feedback or actuator stall occurs, the corresponding degree of freedom is prohibited from continuing to move. After the collision risk decreases to the exit threshold and continues for a predetermined time, active adjustment is stopped, the current safe state is maintained, or the seat is restored to the pre-event state according to a low-speed trajectory after occupant confirmation.

10. A vehicle seat active protection system based on occupant posture injury prediction, characterized in that, The system includes a multimodal occupant perception module, a spatiotemporal synchronization and credibility assessment module, an occupant posture reconstruction module, a human body parameter and digital human body module, a collision risk prediction module, a damage prediction module, a safe feasible domain generation module, a multi-objective posture optimization module, a cooperative control module, and a closed-loop feedback and degradation module. The multimodal occupant perception module acquires occupant multimodal perception data, seat status, and restraint device status. The spatiotemporal synchronization and credibility assessment module and the occupant posture reconstruction module form an occupant status including three-dimensional key points, posture angles, contact relationships, and state uncertainties. The human body parameter and digital human body module generates a parameterized digital human body model. The collision risk prediction module generates candidate collision scenarios. The safe feasible domain generation module generates a safe feasible domain based on the estimated remaining collision time, seat actuation capability, human body clearance, and seatbelt geometry. The damage prediction module is used to predict the risk of damage to multiple parts of the occupant and the prediction uncertainty corresponding to different candidate states within the safe and feasible domain; the multi-objective attitude optimization module is used to solve the state of the target seat and restraint device and the segmented adjustment trajectory within the safe and feasible domain; the cooperative control module is used to issue cooperative control commands to the seat actuator and the reversible restraint device; the closed-loop feedback and degradation module is used to maintain, replan, stop or degrade the segmented adjustment trajectory according to the changes in occupant state, execution deviation and collision risk during the execution process.

Images