Intelligent simulation and restoration system and method for traffic accidents based on internet of vehicles

Abstract

The application is based on a traffic accident intelligent simulation restoration system and method based on Internet of Vehicles, comprising: acquiring vehicle-mounted operation data of each participating vehicle before and after the collision and generating associated data sets; aligning a unified space-time reference system, determining a collision vehicle sequence; identifying independent collision events between adjacent vehicle pairs and splitting into multiple collision stage segments including at least a first active collision stage and a passive transmission collision stage; determining a cross-stage transmission of the collision role of the same vehicle between adjacent collision stage segments, respectively connecting the collision role nodes of the vehicle in different segments which do not merge with each other, and generating a layered collision responsibility network; and performing digital reconstruction of the collision process based on the layered collision responsibility network and outputting the restoration result. The application solves the problems of the prior art that the responsibility attributes of multiple vehicle serial collisions are merged as a whole and the differentiated roles of the intermediate vehicles in different collision stages cannot be expressed in layers.

Description

Technical Field

[0001] This disclosure relates to the field of intelligent transportation and vehicle-to-everything (V2X) data processing technology, and more specifically, to an intelligent simulation and reconstruction system and method for traffic accidents based on the V2X. Background Technology

[0002] In multi-vehicle sequential collision events (typically multi-vehicle rear-end collisions on highways or urban expressways), a single accident often involves multiple collision processes that are closely linked in time but have different mechanical properties. Taking a four-vehicle rear-end collision as an example: the third vehicle first actively impacts the second vehicle (initial active collision); the second vehicle, after being pushed, then impacts the first vehicle (passive pushing collision); the fourth vehicle rear-ends the third vehicle and continues to push against the vehicles in front through the third vehicle (inertial collide collision); after the initial collision, the vehicles make contact again due to elastic rebound or the back pressure of the collision queue (secondary rebound collision). In this process, the vehicles located in the middle of the collision queue (such as the second and third vehicles) play drastically different roles in different collision stages—they may be passive victims pushed from behind, responsible parties who actively rear-end the vehicles in front, or participants in the secondary contact caused by the rebound after the initial collision.

[0003] However, existing technical solutions generally suffer from the following core defects:

[0004] (1) Collision stage merging leads to the indivisibility of intermediate vehicle roles. Existing solutions treat multi-vehicle serial collisions as a single event, merging multiple consecutive collisions into a single accident segment. The consequence of this "overall merging" is that the differentiated roles of intermediate vehicles in different collision stages, such as active collision, passive pushing, inertial transmission, and secondary rebound, are forcibly categorized into the same general responsibility label. When it is necessary to inquire whether "the intermediate vehicle actively rear-ended or was passively pushed in the second collision," existing solutions cannot provide a staged, traceable answer—because their output responsibility attribute is a holistic conclusion, rather than a judgment result unfolded in a chronological order.

[0005] (2) Lack of analytical mechanism for cross-stage role transfer and inheritance. Even in some schemes that attempt to distinguish collision stages, the analysis is "fragmented"—each stage is handled independently, without establishing a role transfer relationship between adjacent stages for the same vehicle. This means that the existing motion state and force attributes of a vehicle in the previous stage cannot be transferred and continued to the role determination in the next stage. For the intermediate vehicle, whether its role transition from "being hit by (the following vehicle)" to "hitting (the preceding vehicle)" has causal continuity (i.e., role inheritance) cannot be answered by the existing fragmented analysis, resulting in the chain of responsibility breaking at the stage boundary.

[0006] (3) Lack of robust design, data gaps directly lead to analysis interruption. In actual vehicle data collection for the Internet of Vehicles, there are data gaps such as signal loss, inconsistent sampling times, and sensor failures. When faced with missing data segments, existing solutions usually adopt the approach of directly skipping or simple interpolation, lacking a robust design that uses "data integrity" as a model input factor and adaptively adjusts the decision confidence level. This results in the entire analysis process being forced to stop or outputting unreliable conclusions when data gaps occur within the collision phase.

[0007] In summary, existing technologies cannot achieve the technical path of "breaking down multi-vehicle serial collisions into independent collision stage segments → establishing cross-stage role transfer relationships → attaching non-merging collision role nodes to the same vehicle at different stages". This is a core technical challenge that urgently needs to be overcome in this field. Summary of the Invention

[0008] This invention provides an intelligent simulation and reconstruction method for traffic accidents based on the Internet of Vehicles, comprising:

[0009] S1: Obtain the on-board operation data of each participating vehicle in the first time window before and after the collision in a multi-vehicle serial collision event, and generate a multi-vehicle collision association dataset.

[0010] S2: Based on the multi-vehicle collision association dataset, the on-board operation data of each participating vehicle are aligned under a unified spatiotemporal reference system, and the sequence position relationship of each participating vehicle in the collision vehicle sequence is determined to generate a collision vehicle spatiotemporal alignment sequence.

[0011] S3: Based on the spatiotemporal alignment sequence of the collision vehicles, identify independent collision events between each pair of adjacent vehicles, and divide the independent collision events into multiple collision stage segments, including at least the first active collision stage and the passive transmission collision stage, according to the collision triggering sequence and collision mechanical characteristics, to generate a multi-stage collision event sequence.

[0012] S4: For each collision stage segment in the multi-stage collision event sequence, extract the collision state features of the vehicle and the vehicle being collided with, and determine the cross-stage transfer of the collision role of the same vehicle between adjacent collision stage segments. Assign the corresponding collision role attributes to the same vehicle in different collision stage segments to generate a hierarchical collision responsibility network. In the hierarchical collision responsibility network, the same intermediate vehicle is associated with multiple non-merged collision role nodes, and each collision role node corresponds to a different collision stage segment and a different collision target vehicle.

[0013] S5: Based on the hierarchical collision responsibility network, the collision process is digitally reconstructed according to the time hierarchy of the collision stage segments, and the result of the full process restoration of the accident containing the collision role attributes of each vehicle in different collision stage segments is output.

[0014] Furthermore, the vehicle operation data includes a multi-source trigger timestamp set from multiple independent triggering devices; generating the multi-vehicle collision association dataset includes: fusing the multi-source trigger timestamp set into a unified collision center time through a weighted fusion model, and using the collision center time as a benchmark, determining the start and end boundaries of the first time window based on the collision extension reference duration and a preset extension coefficient.

[0015] Furthermore, the alignment process under the unified spatiotemporal reference system includes: calculating the time offset value relative to the collision center time for the timestamp of the vehicle terminal of each participating vehicle; uniformly correcting the state sampling time of each participating vehicle based on the time offset value; and projecting the position data of each participating vehicle onto the local road coordinate system along the direction of the collision segment, and introducing a lateral offset penalty to determine the queue order of each participating vehicle in the collision vehicle sequence.

[0016] Furthermore, the identification of independent collision events between adjacent vehicle pairs includes: for each adjacent vehicle pair, a collision occurrence confidence level is generated through a multi-factor fusion judgment model based on the longitudinal distance between vehicles, relative closing velocity, relative longitudinal acceleration difference, collision sensor triggering state, and data integrity index; when the collision occurrence confidence level exceeds a preset candidate threshold and the distance between vehicles enters the contact critical range, it is determined as the starting point of the collision event; when the collision occurrence confidence level continues to decay and the relative velocity changes from closed to separated, it is determined as the ending point of the collision event; for multiple discrete contact features of the same adjacent vehicle pair within a preset short time window, if the interval between two adjacent contacts is less than a preset minimum separation threshold, they are merged into the same independent collision event; otherwise, they are determined as different independent collision events.

[0017] Furthermore, the passive transmission collision phase is further divided into a passive pushing collision phase, an inertial linkage collision phase, and a secondary rebound collision phase. The passive pushing collision phase refers to the collision process in which the middle vehicle is pushed forward and contacts the vehicle in front after being hit by the vehicle behind. The inertial linkage collision phase refers to the linkage contact process in which the middle vehicle has already contacted the vehicle in front, and the vehicle behind continues to apply thrust to maintain contact. The secondary rebound collision phase refers to the collision process in which the vehicles deflect or separate after the collision, and then contact each other again due to elastic rebound or queue backpressure.

[0018] Furthermore, the cross-stage transfer determination of the collision role of the same vehicle between adjacent collision stage segments includes: generating a role confidence score through a multi-factor transfer determination model based on the time interval between the end time of the previous collision stage segment and the start time of the next collision stage segment, the predictability of the vehicle's position between the two times, the speed continuity and data integrity indicators; when the role confidence score meets the transfer condition, at least some of the collision attribute states determined for the vehicle in the previous collision stage segment are carried over to the next collision stage segment.

[0019] Furthermore, the hierarchical collision responsibility network includes vehicle entity nodes, collision event nodes, and collision role nodes. Each vehicle entity node is connected to the collision role nodes it assumes within the corresponding collision phase segment via directed edges. Each collision role node is connected to the corresponding collision event node via directed edges. Adjacent collision event nodes are connected via time-sequence edges to represent the hierarchical relationship of collision propagation. The collision role nodes include at least the following types: active collision attribute nodes representing a vehicle actively chasing and causing a collision with the vehicle in front; passive transmission attribute nodes representing a vehicle transmitting impact forward after being pushed forward by a collision from a vehicle behind; and secondary contact attribute nodes representing a vehicle making contact with other vehicles again due to rebound or deflection after a collision.

[0020] Furthermore, the digital reconstruction of the collision process includes: for each collision stage segment, selecting at least some of the following moments as key state snapshots: the closing moment before the collision, the collision start moment, the moment of maximum speed change, the moment of passive transmission formation, the moment of secondary contact, and the moment of final stopping; the selection of key state snapshots is based on key state snapshot intensity indicators, which are jointly generated based on the speed change amplitude, the vehicle spacing change amplitude, the attitude change amplitude, the collision occurrence confidence, and the role confidence.

[0021] Furthermore, the accident reconstruction results include at least one of the following output formats: a digital reconstruction report displayed according to the collision stage time hierarchy, a dynamic collision process timeline, staged reconstruction images, and a standardized collision scene dataset for training autonomous driving models; wherein, the standardized collision scene dataset retains data integrity indicators and the confidence level of each collision role node.

[0022] This invention also provides an intelligent simulation and reconstruction system for traffic accidents based on the Internet of Vehicles, comprising:

[0023] The data acquisition module is used to acquire the on-board operation data of each participating vehicle in a multi-vehicle serial collision event within the first time window before and after the collision, and generate a multi-vehicle collision association dataset.

[0024] The spatiotemporal alignment module is used to perform alignment processing on the on-board operation data of each participating vehicle under a unified spatiotemporal reference system based on the multi-vehicle collision association dataset, and to determine the sequence position relationship of each participating vehicle in the collision vehicle sequence, thereby generating a spatiotemporal alignment sequence of collision vehicles.

[0025] The collision phase segmentation module is used to identify independent collision events between adjacent vehicle pairs based on the spatiotemporal alignment sequence of the collision vehicles, and to segment the independent collision events into multiple collision phase segments, including at least the first active collision phase and the passively transmitted collision phase, according to the collision triggering sequence and collision mechanical characteristics, thereby generating a multi-stage collision event sequence.

[0026] The role hierarchical inheritance module is used to extract the collision state features of the acting vehicle and the collided vehicle for each collision stage segment in the multi-stage collision event sequence, and to determine the cross-stage transfer of the collision role of the same vehicle between adjacent collision stage segments. The module attaches the corresponding collision role attributes to the same vehicle in different collision stage segments to generate a hierarchical collision responsibility network. In the hierarchical collision responsibility network, the same intermediate vehicle is associated with multiple non-merged collision role nodes, and each collision role node corresponds to a different collision stage segment and a different collision object vehicle.

[0027] The reconstruction output module is used to digitally reconstruct the collision process based on the hierarchical collision responsibility network and according to the time hierarchy of the collision stage segments, and output the full process reconstruction result of the accident, which includes the collision role attributes of each vehicle in different collision stage segments.

[0028] The technical effects of this invention are as follows:

[0029] (1) Addressing the problem of "merging collision stages leading to indivisible intermediate vehicle roles"—This invention decomposes multi-vehicle serial collisions into independent collision stage segments such as the first active collision, passive pushing collision, inertial linkage collision, and secondary rebound collision by splitting the collision stage and determining cross-stage role inheritance. Furthermore, it attaches non-merging collision role nodes (active collision attribute, passive transmission attribute, secondary contact attribute, etc.) to the same vehicle in different stage segments. Thus, the same intermediate vehicle can be simultaneously associated with multiple collision role nodes corresponding to different collision stage segments and different collision target vehicles, fundamentally solving the core defect of existing technologies that "package" the responsibility attributes of intermediate vehicles as a whole, making it impossible to trace them stage by stage.

[0030] (2) To address the problem of "lack of an analysis mechanism for cross-stage role transfer and inheritance"—this invention introduces a multi-factor transfer judgment model based on time continuity, position predictability, speed continuity, and data integrity indicators. This model generates role confidence scores between adjacent collision stage segments, conditionally extending the collision attribute state of a vehicle in the previous stage to the next stage. This ensures that the causal continuity of the role transition process of the intermediate vehicle from "being hit → hitting the vehicle in front" is quantitatively determined, rather than being analyzed in a fragmented and independent manner, thus maintaining the continuous connection of the collision responsibility chain at the stage boundaries.

[0031] (3) To address the issue of "missing robust design, data gaps directly leading to analysis interruption"—each judgment model in this invention (collision occurrence confidence, role confidence, key state snapshot strength, etc.) embeds data integrity indicators as input factors. When a data source is missing, its impact is reduced through adaptive weight adjustment, rather than simply skipping or interrupting the processing flow. At the same time, missing data segments with continuous missing duration exceeding the tolerance threshold are marked with missing tags for downstream processes to make conservative judgments, rather than ignoring the risk of missing data and forcibly outputting unreliable conclusions.

[0032] In addition, the hierarchical collision responsibility network and standardized collision scenario dataset produced by this invention can serve the traffic management department in the phased responsibility assessment (outputting digital reconstruction reports and phased reconstruction images), and can also be directly used for collision avoidance and decision planning algorithm training of autonomous driving models (preserving the confidence of role nodes and data integrity annotation), thus having dual application value. Attached Figure Description

[0033] Figure 1 The implementation flowchart of Example 1 is shown. Detailed Implementation

[0034] Example 1: This example corresponds to the complete implementation process of the method described in this invention. See [link to example]. Figure 1 The process consists of four steps: data acquisition and associated dataset generation (S1), unified spatiotemporal alignment and sequence sorting (S2), collision phase splitting (S3), role hierarchical inheritance (S4), and digital reconstruction of the collision process and result output (S5).

[0035] S1: Data Acquisition and Generation of Multi-Vehicle Collision Association Datasets

[0036] S11: Multi-source trigger timestamp acquisition and collision center time fusion

[0037] In S1, the vehicle operation data includes a set of multi-source trigger timestamps originating from multiple independent triggering devices. In one specific embodiment, the multi-source trigger timestamps include at least the following three categories:

[0038] The first category is traffic accident alarm time, which originates from traffic management platforms, eCall emergency call systems, or manual alarm recording systems. The accuracy of this alarm time is affected by factors such as manual triggering delays and platform forwarding delays, and it usually lags behind the actual physical collision time by several seconds or even tens of seconds.

[0039] The second category is the initial trigger time of the vehicle-mounted collision sensor, which originates from the acceleration threshold trigger signal in the vehicle-mounted event data recorder (EDR) or vehicle-mounted telematics terminal (T-BOX) installed in each vehicle. The trigger condition for this collision sensor trigger time is that the longitudinal or lateral acceleration detected by the vehicle-mounted acceleration sensor exceeds the preset collision trigger threshold, serving as a reliable reference for the moment approaching physical impact.

[0040] The third category is the vehicle airbag deployment time, which originates from the ignition trigger signal of the airbag control unit (ACU). This airbag deployment time typically occurs within a very short time (milliseconds) after the collision sensor is triggered, and is one of the most accurate markers of the moment of physical impact, but it only occurs when the severity of the collision reaches the condition for airbag deployment.

[0041] The aforementioned multi-source trigger timestamp set is fused into a unified collision center time using a weighted fusion model. Specifically: for each participating vehicle, when all three types of trigger times exist, a weighted average is used to obtain the candidate collision center time value for that vehicle. Since the three time sources differ fundamentally in their accuracy of indicating the physical impact time—airbag trigger time has the highest accuracy, followed by collision sensor trigger time, and traffic accident alarm time has the lowest accuracy due to human and system delays—the weights of each time source are allocated as follows during weighted fusion: airbag trigger time has the highest weight, collision sensor trigger time has a medium weight, and traffic accident alarm time has the lowest weight, thus ensuring that the fusion result closely matches the actual physical impact time.

[0042] Preferably, when a certain type of time source is missing (e.g., a minor collision does not trigger the airbags, or older vehicles are not equipped with EDR, resulting in unavailable collision sensor data), the weight corresponding to the missing time source is automatically set to zero, and the remaining available time sources continue to participate in the weighted calculation. This automatic weight reduction mechanism ensures that in a real-world connected vehicle environment where data sources are uncertain, the system can always generate a collision center time reference, rather than interrupting the entire processing flow due to the absence of a single data source.

[0043] For all participating vehicles, the mean or median of the candidate collision center times for each vehicle is taken as the unified collision center time. In practice, outliers that deviate significantly from the median can be removed first, and then the remaining values ​​can be averaged.

[0044] S12: Dynamic determination of the first time window

[0045] Based on the collision center time obtained by fusion, the pre-collision truncation time and the post-collision truncation time are determined respectively, and the two together define the start and end boundaries of the first time window.

[0046] The pre-collision capture time is determined based on the product of the collision expansion reference time and the preset pre-collision expansion coefficient. It is also subject to the dual constraints of the minimum and maximum capture time of the pre-collision window. That is, the capture time is not lower than the minimum capture time lower limit to ensure that key driving behavior data is fully covered, while not exceeding the maximum capture time upper limit to avoid including normal driving data or passing vehicle data that are unrelated to the collision in the analysis scope due to an excessively long collection window.

[0047] The post-collision capture duration uses the same limiting logic as before the collision: it is determined by multiplying the collision expansion reference duration with the preset post-collision expansion coefficient as the base value, and then limiting it by both the minimum capture duration lower limit and the maximum capture duration upper limit of the post-collision window.

[0048] It is important to note that the pre-collision expansion coefficient is greater than the post-collision expansion coefficient. The core rationale for this design is that the determination of responsibility in a chain-reaction rear-end collision relies heavily on a comprehensive analysis of the braking behavior of each vehicle before the collision, the evolution of following distance, and the trajectory of speed changes—this is crucial physical evidence for determining whether the middle vehicle was "actively rear-ending" or "passively pushed." Post-collision data is primarily used to confirm the direction of the push and the final stopping state. The asymmetry in information density before and after the collision dictates that the duration of the pre-collision data acquisition window should be longer than the post-collision window.

[0049] The first time window is a time interval on the time axis consisting of the time segment before the collision center moment extending forward and the time segment after the collision extending backward.

[0050] S13: Collection range of vehicle operation data

[0051] Within the defined first time window, operational status data is collected for each participating vehicle. The sources of vehicle operational data include at least one of the following on-board data acquisition devices installed in each vehicle: on-board telematics terminal (T-BOX), on-board diagnostic system (accessed via OBD-II interface), and event data recorder (EDR).

[0052] The collected data items cover at least the following categories: For location and motion data, this includes vehicle positioning trajectory (such as latitude, longitude, or GPS coordinates), longitudinal speed, longitudinal acceleration, and lateral acceleration; for attitude data, this includes heading angle and yaw rate; for control status data, this includes brake pedal opening, accelerator pedal opening, steering angle, and gear position; for vehicle status data, this includes headlight status (brake lights, turn signals) and collision trigger status; and for metadata, this includes the vehicle terminal timestamp and vehicle dimensions (length, width, and height). The sampling frequency for each data item may vary depending on the actual configuration of the vehicle terminal, typically ranging from several to tens of times per second.

[0053] S14: Establishment of a multi-vehicle collision correlation dataset

[0054] Vehicles that collide consecutively in the same direction of travel, lane, or adjacent lanes are selected as candidate vehicles. A correlation metric is generated between vehicle pairs using a correlation determination model. This correlation metric comprehensively evaluates the following factors:

[0055] First, the longitudinal distance between the two vehicles. In the local road coordinate system, the actual distance between the two vehicles is measured along the longitudinal direction of the road; the smaller the distance, the higher the probability that the two vehicles are in the same collision queue when a collision occurs, and the stronger the correlation.

[0056] Secondly, the degree of difference in the heading angles of the two vehicles. The smaller the difference in the heading angles of the two vehicles (usually not exceeding the reasonable deflection range of the road's direction of travel), the higher the probability of the two vehicles traveling in the same direction and the stronger their correlation.

[0057] Third, the difference in relative speed between the two vehicles. An abnormal change in the speed difference between the two vehicles near the moment of impact center, consistent with the propagation of the impact (e.g., a sudden drop in the speed of the rear vehicle and a sudden increase in the speed of the front vehicle), is a strong correlation signal that there is a transmission of collision mechanics between the two vehicles.

[0058] Fourth, the proximity of the collision trigger times of the two vehicles. The smaller the difference in the trigger times of the collision sensors of the two vehicles, the more likely the two vehicles are to be affected successively in the same collision sequence, and the stronger the correlation.

[0059] When the correlation metric exceeds a preset correlation threshold, the two vehicles are determined to belong to the same collision sequence, are included in the same multi-vehicle collision correlation dataset, and are assigned a unified accident number. Bystander vehicles, oncoming vehicles, and vehicles outside the median strip that do not meet the correlation criteria but are spatially close are not included in this dataset. Each multi-vehicle collision correlation dataset contains at least two vehicles.

[0060] S2: Unified Spatiotemporal Alignment and Collision Vehicle Sequence Ordering

[0061] S21: Time Dimension Alignment

[0062] For the timestamps of the onboard terminals of each participating vehicle, a time offset value relative to the unified collision center time is calculated. This time offset value is determined based on the relative relationship between the vehicle's available trigger times—including airbag trigger time, collision sensor trigger time, and initial braking trigger time (i.e., the moment when the brake pedal opening first changes from zero to non-zero)—and the unified collision center time. When a trigger time is missing because the collision did not meet the triggering conditions, the contribution weight of the corresponding trigger time is set to zero.

[0063] After the offset calculation is completed, the sampling times of each participating vehicle are uniformly corrected to the offset relative to the unified collision center time. This is done by subtracting the corresponding time offset value and the unified collision center time from the original timestamp of each vehicle's on-board terminal, resulting in the corrected relative accident time. After the above correction, the state data of all participating vehicles are uniformly aligned to the relative accident time axis with the collision center time as the origin. The state data of each vehicle, such as speed, position, and acceleration, can be compared laterally moment by moment under the same relative accident time coordinate.

[0064] S22: Spatial Dimension Alignment

[0065] Project the location data (latitude and longitude coordinates or global planar coordinates) of each participating vehicle onto the local road coordinate system along the direction of the collision segment.

[0066] Specifically, the longitudinal coordinate axis is the lane direction of the accident section (the positive direction is consistent with the vehicle's travel direction), and the lateral coordinate axis is the direction perpendicular to the lane direction. For the original position coordinates of each vehicle, the coordinates are rotated based on the road travel direction angle (i.e., the azimuth angle of the road section in the global coordinate system, which can be determined by high-precision map data or the direction of the line connecting the start and end points of the road section), and the two-dimensional position coordinates in the global coordinate system are projected into the cumulative position along the longitudinal direction of the road and the lateral position deviating from the center line of the target lane.

[0067] After spatial projection, the cumulative position of each vehicle along the longitudinal direction of the road can be directly used to determine the order relationship between vehicles in front and behind, while the lateral distance from the center line of the target lane is used to help identify vehicles located in adjacent lanes or non-target lanes.

[0068] S23: Sequence order of collision vehicles

[0069] In the local road coordinate system, a penalty term based on lateral distance is introduced for the cumulative position of each participating vehicle along the longitudinal direction of the road to generate a comprehensive ranking index. This penalty term is the product of the lateral distance and a preset penalty coefficient. Its function is to: when a vehicle deviates significantly from the centerline of the target lane (i.e., is located in an adjacent lane), the vehicle's ranking index is penalized and lowered, pushing it down in the order of collisions, thereby excluding bystander vehicles located in adjacent lanes but not actually participating in the same collision queue. Similarly, vehicles whose heading angle deviates from the road's driving direction angle beyond a preset reasonable range are also excluded to filter oncoming or side-collision vehicles, ensuring that the collision sequence ranking result only reflects vehicles that actually participated in the collision queue in the same driving direction.

[0070] By arranging all participating vehicles in descending or ascending order according to the comprehensive ranking index, the collision queue order from the first vehicle to the last vehicle (or vice versa) can be determined.

[0071] S24: Data Missing Handling

[0072] In the aligned spatiotemporal alignment sequence of the collision vehicles, for state data segments whose continuous missing duration exceeds the preset missing tolerance threshold, the system does not directly delete the data segment, but generates a missing marker for it and records information such as the missing start time, missing duration, and missing data item list.

[0073] This missing marker will serve as the input for the data integrity factor in all subsequent decision models. When calculating the collision occurrence confidence, role confidence, and key state snapshot intensity index, if a missing marker exists within the corresponding time interval involved in the decision model, the data integrity factor is adaptively lowered, thereby reducing the reliability of the confidence output for that interval. This "no discarding, but lowering the weight" design enables the system to continuously perform analysis and processing without interruption when facing the data missing problem commonly found in real-world vehicle networking environments, while ensuring the conservatism and credibility of the analysis conclusions through adaptive reduction of confidence.

[0074] S3: Collision Phase Breakdown

[0075] S3 specifically includes two sub-processes: independent collision event identification and collision phase classification.

[0076] S31: Confidence determination of multi-factor fusion collision occurrence

[0077] For each pair of adjacent vehicles (the preceding and following vehicles) in the spatiotemporal alignment sequence of the collision vehicles, a collision confidence score is generated at each uniform sampling time using a multi-factor fusion model. This model integrates factors from the following five dimensions, quantifying the assessment of "whether a collision has occurred at the current moment" from different physical dimensions:

[0078] The first factor is the contact proximity factor, which is based on a comparison between the actual longitudinal distance between the two vehicles (after deducting the influence of the vehicle lengths) and the contact critical scale. When the two vehicles have made physical contact (the distance approaches zero or is negative), this factor reaches its maximum value, indicating a collision has occurred. When the distance between the two vehicles significantly exceeds the contact critical range, this factor approaches its minimum value, indicating that the two vehicles have not yet made contact. Within the intermediate transition range, this factor monotonically increases as the distance decreases. To compensate for the inherent positioning error of the vehicle positioning system, a buffer tolerance mechanism based on the vehicle's external dimensions can be introduced—when the distance between the two vehicles is below this buffer scale, it is considered that they have entered the contact critical range.

[0079] The second factor is the relative closing speed factor, based on the difference between the speeds of the following vehicle and the preceding vehicle. When the speed of the following vehicle is faster than the speed of the preceding vehicle (i.e., the two vehicles are closing the gap), this factor increases as the speed difference increases; when the speed of the following vehicle is slower than or equal to the speed of the preceding vehicle (the closing condition is not met), this factor takes its minimum value. This factor can be further transformed through continuous and smooth normalization to make it change continuously within its value range, thereby avoiding the judgment jumps caused by hard thresholds.

[0080] The third factor is the acceleration difference factor, which is based on the difference between the longitudinal acceleration of the following vehicle and the longitudinal acceleration of the preceding vehicle. This factor pays special attention to differentiated scenarios where the following vehicle does not decelerate or even accelerates, while the preceding vehicle has decelerated. This acceleration difference pattern indicates that the driver of the following vehicle did not respond in a timely manner to the deceleration behavior of the preceding vehicle, and is a strong indicator signal for judging "active rear-end collision" (rather than passive pushing).

[0081] The fourth factor is the collision sensor triggering factor, which directly utilizes the physical triggering signals of the vehicle's collision sensors. If the collision sensor of the following vehicle has been triggered within a preset proximity time window, this factor takes the highest value, indicating hardware-level evidence of a physical collision; if only the collision sensor of the preceding vehicle has been triggered (the preceding vehicle's triggering may originate from the pushing and transferring of a vehicle further ahead rather than a direct impact from the following vehicle), this factor takes a partial value; if neither has been triggered, this factor takes the minimum value. In some implementations, this factor can be further quantized with finer granularity based on the intensity of the collision sensor triggering signal.

[0082] The fifth factor is the data integrity factor, calculated based on the degree of data loss in the data segments before and after the current sampling time and the continuity of the sampling cycle. This factor is designed as a global modulation factor, rather than a simple addition or multiplication of the aforementioned four physical dimension factors. Its role is to use data integrity as a "reliability discount factor": it remains at its full value when the data is complete, without affecting the confidence output; within the time interval where missing markers exist, this factor is lowered, thereby proportionally reducing the collision confidence of the output at that moment. This design reflects the core idea of ​​"conservative stance when data is missing".

[0083] In the aforementioned five-factor architecture, the first four physical dimensions (contact proximity, relative closing velocity, acceleration difference, and collision sensor triggering) are fused using a weighted summation method, each assigned a reasonable fusion weight to reflect the relative importance of each physical dimension in collision determination. The fused result is then globally modulated by the fifth factor (data integrity factor)—maintaining the full value when data is complete and proportionally reducing it when data is missing—ultimately generating the collision occurrence confidence at that moment.

[0084] S32: Dynamic Recognition of Collision Event Boundaries

[0085] The collision confidence time-series curves of each pair of adjacent vehicles are scanned along the time axis, and the start and end boundaries of the collision event are determined by a joint judgment of confidence threshold and kinematic state.

[0086] The determination of the starting point of a collision event requires the simultaneous fulfillment of two conditions: First, the confidence level of the collision exceeds the preset candidate threshold for the first time, indicating that the probability of a collision has reached a confidence level worthy of being marked; Second, the longitudinal distance between the two vehicles has entered the critical contact range, or the value of the collision confidence level itself has reached the level indicating a highly credible contact state, thereby avoiding false alarms caused by positioning errors.

[0087] The determination of the end point of a collision event also requires the simultaneous fulfillment of two conditions: First, the confidence level of the collision has continuously decreased from the peak value within the event to below the preset candidate threshold, indicating that the probability of a collision has dropped to a level that is not worth continuing to mark; Second, the relative speed of the two vehicles has changed from a "closed" state where the rear vehicle is faster than the front vehicle to a "separated or equal speed" state where the rear vehicle is slower than or equal to the front vehicle, and this interval state needs to be maintained for a preset minimum duration to avoid misjudgment due to instantaneous fluctuations in the speed signal.

[0088] S33: Merging and Splitting of Short-Term Multiple Discrete Contacts

[0089] In real-world multi-vehicle collision scenarios, the same pair of adjacent vehicles may experience multiple discrete contact characteristics within a very short period of time—typically, the elastic rebound of the vehicle body after a collision causes brief, minute separations (on a timescale of hundreds of milliseconds), followed by renewed contact. Collision phase separation requires correctly distinguishing between this "repeated elastic oscillations within the same collision process" and "truly independent subsequent collisions."

[0090] To address this, the system employs the following rules: For multiple collisions involving adjacent vehicles occurring within a preset short observation time window, where the confidence level crosses the candidate threshold and meets the starting point condition, the time interval between two adjacent contact points is calculated. If this interval is less than a preset minimum separation threshold, the two contacts are determined to be repeated oscillations within the same collision process, and are merged into a single independent collision event, with the event's time span extended from the starting point of the first contact to the ending point of the last contact. If the interval reaches or exceeds the minimum separation threshold, the events are determined to be two independent collision events, and are separately segmented into stages.

[0091] S34: Classification and decomposition of the collision phase

[0092] Each identified independent collision event is divided into multiple collision stage segments, including at least an initial active collision stage and a passively transmitted collision stage, according to the collision triggering sequence and collision mechanics characteristics. In a preferred embodiment, the collision stages are specifically divided into the following four types:

[0093] (i) Initial Active Collision Phase. This phase describes the collision process in which a rear vehicle actively pursues and impacts the vehicle in front. Its characteristic conditions include: First, before there is evidence of being pushed from behind, the rear vehicle has already formed a close-range approach with the vehicle in front—that is, the speed of the rear vehicle is faster than that of the vehicle in front and the distance between the two vehicles continues to decrease; Second, the braking response of the rear vehicle is insufficient, manifested as the brake pedal opening being lower than the preset effective braking threshold, or the braking operation being initiated too late so that the distance to the collision is insufficient to complete safe braking; Third, the collision is triggered by the rear vehicle actively closing the gap, rather than being pushed forward by an impact from a vehicle further behind.

[0094] (II) Passive Push-and-Push Collision Stage. This stage describes the collision transmission process in which the middle vehicle is pushed forward and comes into contact with the vehicle in front after being hit by a vehicle behind it. Its characteristic conditions include: First, before being hit from behind, there is a safe driving distance between the middle vehicle and the vehicle in front (the distance exceeds the critical contact range); Second, after being hit from behind, the longitudinal velocity of the middle vehicle undergoes a significant positive change in a short period of time, and the magnitude of the change exceeds the preset velocity change threshold. This change has the typical mechanical characteristics of being "pushed"; Third, the velocity change follows the logic of "being hit first, then hitting a person" in the causal sequence - the time of the velocity change should be later than the time when the vehicle behind hits the middle vehicle, and this change directly causes the middle vehicle to catch up with and actually make contact with the vehicle in front.

[0095] (III) Inertial Collision Stage. This stage describes the process where, after the middle vehicle has established contact with the vehicle in front, the following vehicle continues to apply thrust, maintaining the compression state. Its characteristic conditions include: First, the contact state between the middle vehicle and the vehicle in front has already been established in the previous stage (usually the passive pushing collision stage), and the longitudinal distance between the two vehicles is within the critical contact range; second, the speed of the following vehicle is still not lower than the speed of the middle vehicle, meaning that the rear pushing force continues to exist rather than diminishing and disappearing; third, the middle vehicle passively maintains the state of compression against the vehicle in front under the continuous thrust from behind, rather than autonomously accelerating to catch up with the vehicle in front, and the dynamic causes of the two are completely different.

[0096] (iv) Secondary Rebound Collision Phase. This phase describes the process by which a vehicle re-engages with a previously separated vehicle due to elastic rebound, deflection, or backpressure from the rear after a collision event. Its characteristic conditions include: first, after the collision event, the longitudinal distance between the vehicle and the previously contacted vehicle has increased beyond the contact critical range, confirming physical separation; second, after separation, due to one or more factors such as the elastic restoring force of the vehicle body deformation, attitude changes caused by lateral deflection (abrupt change in yaw rate), or the re-exertion of pressure from the rear, the longitudinal distance shrinks back to within the contact critical range, and a peak in the confidence level of collision reappears near this moment.

[0097] After completing the above classification, all collision phase segments are arranged according to the collision triggering sequence, and each segment is labeled with its start and end time, participating vehicle identification, and phase type to generate a multi-stage collision event sequence.

[0098] S4: Role Layer Inheritance and Layer Collision Responsibility Network Generation

[0099] S41: Extraction of Collision State Features

[0100] Each collision stage segment is used as an independent analysis unit. Multiple collision state features of the vehicle that exerts the collision force (i.e., the party that applies the collision force) and the vehicle that is hit (i.e., the party that bears the collision force) are extracted as the basic data for subsequent role determination and responsibility network construction.

[0101] In one specific embodiment, the collision state characteristics include at least the following seven items: First, the identification of the acting vehicle and the identified vehicle, directly determined from the collision phase segment. Second, the collision start time and collision end time, corresponding to the time boundaries of this phase segment. Third, the braking state of the acting vehicle before the collision, specifically including the maximum brake pedal opening of the acting vehicle within a preset time period before the collision begins, and the response delay duration from the moment braking should occur to the moment braking actually occurs. Fourth, the relative speed before the collision, reflecting the magnitude and direction of the closing velocity amplitude of the following vehicle relative to the preceding vehicle within a time window before the collision start time. Fifth, the direction and magnitude of the speed change after the collision, reflecting the magnitude and sign of the speed changes of the acting vehicle and the affected vehicle during the period from the start to the end of the collision—positive values ​​usually indicate acceleration or being pushed, and negative values ​​usually indicate deceleration or being obstructed—and calculating the speed change amplitude accordingly. Sixth, the vehicle attitude change state before and after the collision, including the change in yaw rate and the change in heading angle, used to determine whether sideslip or yaw occurs during the collision. Seventh, collision confirmation strength, which is the peak or average of the confidence level of collision occurrence within the segment of this stage, reflects the degree of certainty of the collision event occurring in this stage.

[0102] S42: Cross-Stage Role Transfer Judgment

[0103] A cross-stage transferability determination is performed on the collision role of the same vehicle between adjacent collision phase segments (the preceding stage and the immediately following stage) to determine whether the collision attribute state determined in the preceding stage is physically causally continuous in the following stage. This determination is based on the following four dimensions, and a role confidence score is generated through a multi-factor transferability determination model:

[0104] The first dimension is temporal continuity. It is evaluated based on the time interval between the end of one phase and the start of the next – the shorter the interval, the stronger the temporal continuity. When the interval exceeds a preset upper limit, the temporal continuity factor approaches its minimum. This usually means that the collision sequence has been interrupted (e.g., vehicle relocation at the scene after an accident has been handled, or an independent event unrelated to the collision has occurred between the two phases), and role transfer relationships across phases should not be established.

[0105] The second dimension is position predictability. Using the vehicle's position and velocity at the end of the previous stage as initial conditions for kinematic extrapolation, and based on reasonable kinematic models such as uniform speed or uniform deceleration, the expected position of the vehicle at the start of the next stage is calculated, along with the deviation distance between this expected position and the vehicle's actual position at that moment. A smaller deviation distance indicates higher position predictability, suggesting that the vehicle's trajectory between the two stages is reasonably predictable, supporting the establishment of a role transfer relationship. If the deviation distance exceeds the reasonable kinematic range (i.e., exceeds the area the vehicle might reach after considering maximum braking deceleration and maximum steering rate), it indicates that the vehicle may have experienced external intervention unrelated to the collision during the transition (e.g., being moved by a tow truck) or that there has been a significant jump in positioning data; in this case, a transfer relationship should not be established.

[0106] The third dimension is speed continuity. This compares the speed at the end of the previous phase with the speed at the beginning of the next phase—the smaller the change in speed, the stronger the continuity. Of particular note is whether there are any abrupt speed changes during the transition period that are inconsistent with the direction of the collision force: for example, a significant decrease in speed (a sharp deceleration) may indicate that the collision has ended, the driver has intervened, and independently applied the brakes; conversely, if the direction and magnitude of the speed change exhibit a reasonable continuation consistent with the direction of the collision's pushing force, the speed continuity factor remains high.

[0107] The fourth dimension is data completeness. If there are missing markers in the transition interval between two adjacent stages, the reliability of the data in that interval decreases, the data completeness factor is lowered, and the result of the transmission judgment is more conservative.

[0108] Among the four dimensions mentioned above, temporal continuity, location predictability, and velocity continuity are fused by weighted summation, with each assigned a fusion weight that reflects its relative importance. The fused result is globally modulated by a data completeness factor—maintaining the full value when data is complete in the transition interval and proportionally reducing it when data is missing—ultimately generating role confidence.

[0109] When the character confidence level reaches or exceeds the preset transmission condition, the system determines that there is a character inheritance relationship between the two adjacent stages, and conditionally carries over at least some of the collision attribute states determined in the previous stage to the next stage. Conversely, if the character confidence level does not reach the preset transmission condition, it is determined that there is no character inheritance relationship, and the character determination in the next stage is carried out independently, unaffected by the character attribute states in the previous stage.

[0110] S43: Collision Character Attribute Connection

[0111] Based on the results of cross-stage transfer determination and collision state characteristics, non-merging collision role nodes are attached to the same vehicle in different collision stage segments. In this embodiment, collision role nodes include at least the following three types:

[0112] (i) Active Collision Attribute Node. This node indicates that the vehicle actively pursues and collides with the vehicle in front without being pushed by an external force from a vehicle behind. The criteria for determining this include: First, before the physical evidence of a rear-end collision occurs, the vehicle has already formed a close-range pursuit with the vehicle in front—that is, the speed of the rear vehicle is faster than the speed of the front vehicle and the distance between the vehicles is continuously decreasing; Second, the vehicle's braking status indicates that it did not take effective braking avoidance actions before the collision, such as the brake pedal opening being lower than the effective braking threshold, or the time between the braking operation and the collision being too short to complete effective deceleration before the collision.

[0113] (ii) Passive Transmission Attribute Node. This represents the vehicle being impacted by a rear vehicle, transmitting the impact forward, and thus catching up with and making contact with the vehicle in front. The criteria for determination include: First, the speed change of the vehicle exceeds a preset speed change threshold, exhibiting typical dynamic characteristics of being "pushed"; second, the speed change occurs later in time than the moment the rear vehicle hits the vehicle—that is, the causal chain follows the logical sequence of "first being hit by the rear vehicle, then hitting the vehicle in front"; third, the speed change objectively causes the vehicle to catch up with and actually make contact with the vehicle in front.

[0114] (iii) Secondary Contact Attribute Node. This indicates that the vehicle makes contact with a previously separated vehicle again after the collision event due to non-active factors such as elastic rebound, deflection, or backpressure from the rear queue. The determination conditions include: First, after the collision event, the vehicle and the previously contacted vehicle have entered a state of physical separation (the distance exceeds the contact critical range); Second, after separation, there is a situation where the change in yaw rate exceeds the preset deflection threshold, or the distance between the two vehicles shrinks back to the contact critical range and a collision occurrence confidence peak is generated again near that moment, indicating the occurrence of secondary contact.

[0115] It is important to note that the same intermediate vehicle may play drastically different roles in different collision phases. Taking a typical four-vehicle rear-end collision scenario as an example, the third vehicle in the queue may correspond to one of three role nodes: In the segment "the third vehicle catches up with the second vehicle," if the third vehicle actively rear-ended the second vehicle before being struck by the fourth vehicle, it is marked as an active collision attribute node; in the segment "the third vehicle pushes the second vehicle after the fourth vehicle hits the third vehicle," the third vehicle is marked as a passive transmission attribute node due to the rearward thrust; in the rebound phase after the collision, if the third vehicle comes into contact with the second vehicle again due to deformation or rebound, it may be marked as a secondary contact attribute node. These three role nodes are not merged; they exist independently in the hierarchical collision responsibility network, each corresponding to different collision phases and different collision target vehicles.

[0116] S44: Construction of a Layered Collision Responsibility Network

[0117] The hierarchical collision responsibility network is constructed as a three-layer directed graph structure.

[0118] The first layer is the vehicle entity node layer. Each vehicle involved in the accident corresponds to a vehicle entity node, which serves as the unified identification entry point for that vehicle in the hierarchical collision liability network, recording basic attributes such as vehicle number, vehicle type, and external dimensions.

[0119] The second layer is the collision role node layer. Each vehicle assumes an independent collision role in each collision phase segment, corresponding to an independent collision role node. The vehicle entity node of the same vehicle can connect to multiple non-merged collision role nodes. Each collision role node records at least: role type (active collision attribute, passive transmission attribute, or secondary contact attribute), role confidence, associated collision phase segment identifier, and associated collision object vehicle identifier.

[0120] The third layer is the collision event node layer. Each identified collision phase segment corresponds to a collision event node, which records at least: the collision phase type (initial active collision phase, passive shoving collision phase, inertial cascading collision phase, or secondary rebound collision phase), start and end times, list of participating vehicles, and collision confirmation intensity.

[0121] Three types of directed edge relationships are established between the three layers of nodes: First, vehicle entity nodes are connected to the various collision role nodes that the vehicle assumes within the corresponding collision stage segment via directed edges, with the edge direction pointing from the vehicle entity node to the role node, semantically representing "this vehicle assumes this role". Second, each collision role node is connected to its corresponding collision event node via directed edges, with the edge direction pointing from the role node to the event node. These directed edges are accompanied by association weights to quantify the certainty of the role determination. These weights comprehensively reflect indicators such as role confidence, collision confirmation strength, and data completeness. Third, adjacent collision event nodes are connected via temporal order edges, with the edge direction pointing from the earlier-occurring event node to the later-occurring event node, semantically representing the temporal hierarchy between different collision stages in the entire collision accident and the direction of collision impact propagation.

[0122] This three-layer directed graph structure endows the hierarchical collision responsibility network with the following key capabilities: it can quickly locate all collision role nodes associated with any participating vehicle through vehicle entity nodes; it can trace the complete physical causal chain of collision propagation along the time axis through the temporal sequence edges between collision event nodes; it can enable the different roles played by the same vehicle in different collision stages to coexist in the same network without conflict through the non-merging characteristic of collision role nodes; and it can quantify the degree of certainty of each role determination through the weight information attached to the directed edges, providing a quantifiable confidence reference for downstream responsibility assessment.

[0123] S5: Digital Reconstruction of Collision Process and Output of Results

[0124] S51: Intelligent Selection of Critical State Snapshots

[0125] For each collision phase segment, instead of passive sampling with uniform time intervals, the system actively selects the key moments that are most valuable for collision mechanics reconstruction and role determination as key state snapshots based on the key state snapshot intensity index.

[0126] The critical state snapshot intensity index comprehensively measures the following five factors, which are uniformly quantified through weighted fusion: The first factor is the magnitude of velocity mutation, reflecting the severity of longitudinal velocity changes in each participating vehicle at that moment; the second factor is the magnitude of vehicle spacing mutation, reflecting the abrupt component of collision contact or physical separation at that moment; the third factor is the magnitude of attitude mutation, calculated based on changes in yaw rate and heading angle, reflecting the degree of vehicle deflection or sideslip caused by the collision impact; the fourth factor is the peak confidence level of collision occurrence, reflecting the certainty of the collision event occurring at that moment or in the adjacent interval; and the fifth factor is the confidence level of role determination, reflecting the certainty of role determination at that moment or in the adjacent interval. Each factor is assigned a fusion weight commensurate with its contribution to the information value in collision mechanics reconstruction, resulting in relatively higher weight allocations for factors such as velocity mutation and collision confidence level, which directly indicate the moment of collision impact.

[0127] The system calculates the intensity index value for each sampling moment within each collision stage segment, arranges them in descending order, and selects several moments at the top of the ranking as key state snapshots. In specific implementation, the number of selected moments can be determined by the system based on preset configuration, or it can be automatically truncated by the system based on the lower limit threshold of the intensity index—that is, only moments when the intensity index reaches or exceeds the preset lower limit are selected as snapshots.

[0128] Preferably, the candidate time types for critical state snapshots include at least: pre-collision closure moment (the moment when the confidence of the collision begins to rise significantly), collision initiation moment (the moment when the collision event is marked as the starting point), maximum speed change moment (the moment when the rate of speed change reaches its peak), passive transmission formation moment (the moment when the middle vehicle first makes physical contact with the vehicle in front after being pushed from behind), secondary contact moment (the moment when the distance between the vehicles shrinks back to the contact critical range after physical separation), and final stopping moment (the moment when the speed of each participating vehicle returns to zero).

[0129] Each critical state snapshot records at least: the position data, speed data, braking status, collision phase label, and the collision role attribute associated with the vehicle at that moment for each participating vehicle.

[0130] S52: Generation of motion frame sequences

[0131] Based on the time-level information recorded in the hierarchical collision responsibility network, the start and end times of each collision event node are extracted and arranged chronologically to form a phased reconstruction event sequence. For each collision phase segment, based on the vehicle motion state observation data at each sampling time within that phase, combined with the kinematic difference method and vehicle dynamics constraints, a continuous frame-by-frame motion state description of each vehicle within that phase is generated, including at least information such as vehicle position, velocity, acceleration, brake pedal state, collision contact relationship, and force direction.

[0132] S53: Output format of accident full-process reconstruction results

[0133] The results of the full accident reconstruction should include at least one or more of the following output formats:

[0134] Firstly, a digital reconstruction report. This report is organized chronologically by collision phase and includes at least: basic information for each collision phase (identification of participating vehicles, start and end times, phase type), the collision role attributes of each vehicle in each phase and their corresponding confidence values, the complete causal chain of the collision impact propagation in the vehicle platoon, and a summary description of role changes across phases. The report format can be a portable document format or Hypertext Markup Language, facilitating direct access by traffic management departments.

[0135] Secondly, there is a dynamic collision process timeline. This timeline uses the temporal order edges between collision event nodes in the hierarchical collision responsibility network as a guide, presenting the temporal evolution of the collision process in an interactive interface. The timeline unfolds stage by stage along the order of collision triggering, marking each vehicle involved in that stage and its corresponding collision role attributes in each collision stage segment. Users can drag and drop along the timeline to view the position, speed, and role label information of each vehicle at any given moment.

[0136] Thirdly, the images are reconstructed in stages. Based on the aforementioned continuous motion frame sequence and key state snapshots, multi-view accident reconstruction animations are generated using computer graphics compositing technology. In the animation, vehicles with different collision roles are distinguished by differentiated visual markers (such as color, outline, or label)—the active collision party and the passive transmission party can be given their own independent visual identifiers—making the dynamic changes of each vehicle role during the collision visible and intuitively perceptible.

[0137] Fourth, a standardized collision scenario dataset. The reconstruction results are output in a structured data format (e.g., JSON, XML, or ROSbag), specifically for constructing training scenarios for collision avoidance and decision planning algorithms in autonomous driving models. A key feature of this dataset is that it does not discard uncertain information. The dataset fully preserves the confidence level of each collision role node and the data integrity indicators of each stage. This means that labels indicating "active collision" with high confidence levels are clearly distinguished from labels indicating "active collision" with low confidence levels. Downstream responsible assessment personnel or autonomous driving training algorithms can autonomously decide whether to adopt or exclude corresponding data based on confidence levels, rather than passively accepting a result that has lost its uncertain information after forced binarization.

[0138] Example 2: This example corresponds to the implementation of the system described in this invention. The system includes a data acquisition module, a spatiotemporal alignment module, a collision phase splitting module, a role hierarchical inheritance module, and a restoration output module. Each module carries the complete processing logic of the corresponding method steps in Example 1.

[0139] This invention provides an intelligent simulation and reconstruction system for traffic accidents based on the Internet of Vehicles. The system includes a data acquisition module, a spatiotemporal alignment module, a collision stage splitting module, a role hierarchical inheritance module, and a reconstruction output module. Each module carries the complete processing logic of the corresponding step in the method.

[0140] The data acquisition module, corresponding to S1, is used to achieve: weighted fusion of multi-source trigger timestamps (differentiated weighted average of traffic accident alarm time, collision sensor trigger time, and airbag trigger time, and automatic zeroing and weight reduction processing of corresponding weights when data is missing); dynamic determination of the first time window (based on the collision center time, and based on the collision extension reference duration, pre-collision extension coefficient, post-collision extension coefficient, and preset minimum and maximum upper / lower limits of the window, bidirectional amplitude limiting calculation of the pre-collision and post-collision interception durations); collection and aggregation of vehicle operation data and establishment of multi-vehicle collision association datasets.

[0141] The spatiotemporal alignment module corresponds to S2 and is used to: calculate the time bias values ​​of each participating vehicle and uniformly correct the sampling time (converting the timestamps of each vehicle's on-board terminal to the relative accident time with the collision center time as the origin); project the position data of each vehicle onto the local road coordinate system along the accident section (decomposing the global position coordinates into the cumulative position along the longitudinal direction of the road and the lateral position deviating from the lane centerline); sort the collision vehicle sequence (introducing a penalty term based on lateral distance to generate a comprehensive ranking index to exclude bystander vehicles in adjacent lanes and oncoming vehicles); and generate data missing markers and input the data integrity factor to each downstream judgment model.

[0142] The collision phase splitting module corresponds to the collision phase splitting part in S3. It is used to: generate a collision occurrence confidence score based on the weighted fusion of five dimensions: contact proximity, relative closing velocity, acceleration difference, collision sensor triggering state, and data integrity; dynamically identify the start and end boundaries of a collision event by jointly determining the collision occurrence confidence score and the relative velocity direction; merge multiple discrete contacts of the same adjacent vehicle pair that occur within a short window based on a minimum separation threshold (if the contact is below the threshold, it is merged into the same collision; if the contact reaches or exceeds the threshold, it is split into an independent collision event); and classify each independent collision event into the corresponding type among the first active collision phase, passive shoving collision phase, inertial cascading collision phase, and secondary rebound collision phase according to the collision triggering sequence and collision mechanical characteristics.

[0143] The role hierarchical inheritance module corresponds to the role hierarchical inheritance and network generation part in S4. It is used to: extract multiple collision state features of the active vehicle and the vehicle being collided for each collision stage segment; generate role confidence based on the weighted fusion of four dimensions: time continuity, position predictability, speed continuity, and data completeness, and determine whether there is a cross-stage inheritance relationship of collision roles for the same vehicle between adjacent collision stage segments; attach non-merging active collision attribute nodes, passive transmission attribute nodes, or secondary contact attribute nodes to the same vehicle in different collision stage segments; and construct a three-layer directed graph structure hierarchical collision responsibility network containing a vehicle entity node layer, a collision role node layer, and a collision event node layer.

[0144] The reconstruction output module, corresponding to S5, is used to: calculate the intensity index of key state snapshots based on a weighted fusion of five factors: speed change amplitude, vehicle spacing change amplitude, attitude change amplitude, peak confidence level of collision occurrence, and role confidence level; and select key state snapshots in descending order accordingly; generate a continuous motion frame sequence of each vehicle within each collision stage segment; and output the reconstruction results of the entire accident process in at least one of the following forms: digital reconstruction report, dynamic collision process timeline, staged reconstruction images, and standardized collision scene dataset. The standardized collision scene dataset fully preserves the confidence level and data integrity index of each collision role node.

[0145] In terms of hardware deployment, the above modules can be deployed in at least one of the following computing environments: cloud server (accessing vehicle operation data reported by each vehicle in real time or near real time through the vehicle network platform, suitable for centralized processing of regional accident data), edge computing node (deployed in roadside units to achieve near-field fast restoration locally on the accident section, reducing cloud communication latency), dedicated workstation of traffic management center, or backend server of insurance claims assessment system.

[0146] In terms of software implementation, the above modules can be implemented as computer software programs, firmware, or any combination thereof. The data acquisition module and the spatiotemporal alignment module can use a general programming language to implement the data preprocessing pipeline; the collision phase splitting module and the role hierarchical inheritance module can rely on a graph computing framework to realize the construction and maintenance of a hierarchical collision responsibility network; the restoration output module can use a front-end development framework to realize a dynamic timeline interactive interface, or use a computer graphics engine to realize high-fidelity phased restoration image synthesis.

Claims

1. A method for intelligent simulation and reconstruction of traffic accidents based on the Internet of Vehicles, characterized in that, include: S1: Obtain the on-board operation data of each participating vehicle in the first time window before and after the collision in a multi-vehicle serial collision event, and generate a multi-vehicle collision association dataset. S2: Based on the multi-vehicle collision association dataset, the on-board operation data of each participating vehicle are aligned under a unified spatiotemporal reference system, and the sequence position relationship of each participating vehicle in the collision vehicle sequence is determined to generate a collision vehicle spatiotemporal alignment sequence. S3: Based on the spatiotemporal alignment sequence of the collision vehicles, identify independent collision events between each pair of adjacent vehicles, and divide the independent collision events into multiple collision stage segments, including at least the first active collision stage and the passive transmission collision stage, according to the collision triggering sequence and collision mechanical characteristics, to generate a multi-stage collision event sequence. S4: For each collision stage segment in the multi-stage collision event sequence, extract the collision state features of the vehicle and the vehicle being collided with, and determine the cross-stage transfer of the collision role of the same vehicle between adjacent collision stage segments. Assign the corresponding collision role attributes to the same vehicle in different collision stage segments to generate a hierarchical collision responsibility network. In the hierarchical collision responsibility network, the same intermediate vehicle is associated with multiple non-merged collision role nodes, and each collision role node corresponds to a different collision stage segment and a different collision target vehicle. S5: Based on the hierarchical collision responsibility network, the collision process is digitally reconstructed according to the time hierarchy of the collision stage segments, and the result of the full process restoration of the accident containing the collision role attributes of each vehicle in different collision stage segments is output.

2. The method according to claim 1, characterized in that, The vehicle operation data includes a set of multi-source trigger timestamps from multiple independent triggering devices; generating the multi-vehicle collision association dataset includes: fusing the set of multi-source trigger timestamps into a unified collision center time through a weighted fusion model, and using the collision center time as a benchmark, determining the start and end boundaries of the first time window based on the collision extension reference duration and a preset extension coefficient.

3. The method according to claim 1, characterized in that, The alignment process under the unified spatiotemporal reference system includes: calculating the time offset value relative to the collision center time for the timestamp of the vehicle terminal of each participating vehicle; uniformly correcting the state sampling time of each participating vehicle based on the time offset value; and projecting the position data of each participating vehicle onto the local road coordinate system along the direction of the collision segment, and introducing a lateral offset penalty to determine the queue order of each participating vehicle in the collision vehicle sequence.

4. The method according to claim 1, characterized in that, The identification of independent collision events between adjacent vehicle pairs includes: for each adjacent vehicle pair, a collision occurrence confidence level is generated through a multi-factor fusion judgment model based on the longitudinal distance between vehicles, relative closing velocity, relative longitudinal acceleration difference, collision sensor triggering status, and data integrity index; when the collision occurrence confidence level exceeds a preset candidate threshold and the distance between vehicles enters the contact critical range, it is determined as the starting point of the collision event; when the collision occurrence confidence level continues to decay and the relative velocity changes from closing to separation, it is determined as the ending point of the collision event; for multiple discrete contact features of the same adjacent vehicle pair within a preset short time window, if the interval between two adjacent contacts is less than a preset minimum separation threshold, they are merged into the same independent collision event; otherwise, they are determined as different independent collision events.

5. The method according to claim 1, characterized in that, The passive transmission collision phase is further divided into a passive pushing collision phase, an inertial linkage collision phase, and a secondary rebound collision phase. The passive pushing collision phase refers to the collision process in which the middle vehicle is pushed forward and contacts the vehicle in front after being hit by the vehicle behind. The inertial linkage collision phase refers to the linkage contact process in which the middle vehicle has already contacted the vehicle in front, and the vehicle behind continues to apply thrust to maintain contact. The secondary rebound collision phase refers to the collision process in which the vehicles deflect or separate after the initial collision, and then make contact again due to elastic rebound or convoy backpressure.

6. The method according to claim 1, characterized in that, The cross-stage transfer determination of the collision role of the same vehicle between adjacent collision stage segments includes: generating a role confidence score through a multi-factor transfer determination model based on the time interval between the end time of the previous collision stage segment and the start time of the next collision stage segment, the predictability of the vehicle's position between the two times, the speed continuity and data integrity indicators; when the role confidence score meets the transfer condition, at least some of the collision attribute states determined for the vehicle in the previous collision stage segment are carried over to the next collision stage segment.

7. The method according to claim 1, characterized in that, The hierarchical collision responsibility network includes vehicle entity nodes, collision event nodes, and collision role nodes. Each vehicle entity node is connected to the collision role nodes it assumes within the corresponding collision phase segment via directed edges. Each collision role node is connected to the corresponding collision event node via directed edges. Adjacent collision event nodes are connected via time sequence edges to represent the hierarchical relationship of collision propagation. The collision role nodes include at least the following types: active collision attribute nodes representing a vehicle actively chasing and causing a collision with the vehicle in front; passive transmission attribute nodes representing a vehicle transmitting impact forward after being hit by the collision thrust of a vehicle behind; and secondary contact attribute nodes representing a vehicle making contact with other vehicles again due to rebound or deflection after the collision.

8. The method according to claim 4 or 6, characterized in that, The digital reconstruction of the collision process includes: for each collision stage segment, selecting at least some of the following moments as key state snapshots: the closing moment before the collision, the collision start moment, the moment of maximum speed change, the moment of passive transmission formation, the moment of secondary contact, and the moment of final stopping; the selection of key state snapshots is based on key state snapshot intensity indicators, which are generated jointly based on the speed change amplitude, the vehicle spacing change amplitude, the attitude change amplitude, the collision occurrence confidence, and the role confidence.

9. The method according to claim 1, characterized in that, The full-process reconstruction results of the accident include at least one of the following output formats: a digital reconstruction report displayed according to the time hierarchy of the collision stage, a dynamic collision process timeline, staged reconstruction images, and a standardized collision scene dataset for training autonomous driving models; wherein, the standardized collision scene dataset retains data integrity indicators and the confidence level of each collision role node.

10. A traffic accident intelligent simulation and reconstruction system based on the Internet of Vehicles, characterized in that, include: The data acquisition module is used to acquire the on-board operation data of each participating vehicle in a multi-vehicle serial collision event within the first time window before and after the collision, and generate a multi-vehicle collision association dataset. The spatiotemporal alignment module is used to perform alignment processing on the on-board operation data of each participating vehicle under a unified spatiotemporal reference system based on the multi-vehicle collision association dataset, and to determine the sequence position relationship of each participating vehicle in the collision vehicle sequence, thereby generating a spatiotemporal alignment sequence of collision vehicles. The collision phase segmentation module is used to identify independent collision events between adjacent vehicle pairs based on the spatiotemporal alignment sequence of the collision vehicles, and to segment the independent collision events into multiple collision phase segments, including at least the first active collision phase and the passively transmitted collision phase, according to the collision triggering sequence and collision mechanical characteristics, thereby generating a multi-stage collision event sequence. The role hierarchical inheritance module is used to extract the collision state features of the acting vehicle and the collided vehicle for each collision stage segment in the multi-stage collision event sequence, and to determine the cross-stage transfer of the collision role of the same vehicle between adjacent collision stage segments. The module attaches the corresponding collision role attributes to the same vehicle in different collision stage segments to generate a hierarchical collision responsibility network. In the hierarchical collision responsibility network, the same intermediate vehicle is associated with multiple non-merged collision role nodes, and each collision role node corresponds to a different collision stage segment and a different collision object vehicle. The reconstruction output module is used to digitally reconstruct the collision process based on the hierarchical collision responsibility network and according to the time hierarchy of the collision stage segments, and output the full process reconstruction result of the accident, which includes the collision role attributes of each vehicle in different collision stage segments.

Images