Vehicle rear collision warning method, device and vehicle

Abstract

The application relates to a vehicle rear collision warning method and device and a vehicle. The method comprises the following steps: adopting a space-time feature fusion model to perform multi-scale feature processing on multi-source perception data of the vehicle to obtain multi-scale risk representation features of the vehicle; based on a behavior intention prediction model and the multi-source perception data, classifying the behavior intention of a rear target corresponding to the vehicle, determining the behavior intention of the rear target; according to the behavior intention of the rear target, the multi-scale risk representation features and the multi-source perception data, identifying the collision risk of the vehicle to obtain a rear collision risk quantitative value of the vehicle; and based on the rear collision risk quantitative value, performing rear collision warning on the vehicle. The above method is used to extract risk representation features of the vehicle from space-time multi-dimensions, effectively capture rear vehicle behavior mutation and long-distance weak risk signals, realize rear collision warning on this basis, and can reduce the false negative rate and the false positive rate, improve the accuracy of the warning.

Description

Technical Field

[0001] This application relates to the field of vehicle technology, and in particular to a vehicle rear collision warning method, device, and vehicle. Background Technology

[0002] With the development of intelligent driving technology, rear collision warning (RCW), as a core function of intelligent driving systems, plays a crucial role in preventing high-speed rear-end collisions and reducing the risk of injury or death to vulnerable road users in rear-end collisions on urban roads.

[0003] In related technologies, the main approach is based on the strong assumption that the target behind the vehicle continues to maintain its current state of motion. By collecting basic kinematic parameters such as target distance and speed, the time to collision (TTC) is calculated, and a warning is triggered by comparing the collision time with a fixed threshold.

[0004] However, the relevant technologies suffer from low accuracy in providing early warnings. Summary of the Invention

[0005] Therefore, it is necessary to provide a vehicle rear collision warning method, device, and vehicle to address the aforementioned technical problems.

[0006] In a first aspect, embodiments of this application provide a vehicle rear collision warning method, the method comprising:

[0007] A spatiotemporal feature fusion model is used to perform multi-scale feature processing on the multi-source perception data of vehicles to obtain multi-scale risk characterization features of vehicles.

[0008] Based on the behavioral intent prediction model and multi-source perception data, the behavioral intent of the target behind the vehicle is classified and the behavioral intent of the target behind the vehicle is determined.

[0009] Based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data, the collision risk of the vehicle is identified, and the quantitative value of the vehicle's rear collision risk is obtained.

[0010] Based on the quantified value of rear collision risk, a rear collision warning is issued for the vehicle.

[0011] In one embodiment, the spatiotemporal feature fusion model includes a spatial feature pyramid branch and a temporal feature pyramid branch; using the spatiotemporal feature fusion model, multi-scale feature processing is performed on the multi-source perception data of the vehicle to obtain multi-scale risk characterization features of the vehicle, including:

[0012] Multi-source sensing data is input into the spatial feature pyramid branch for feature extraction to obtain multi-scale spatial fusion features;

[0013] Multi-source sensing data is input into the temporal feature pyramid branch for feature extraction to obtain full temporal behavior fusion features;

[0014] Multi-scale spatial fusion features and full-time behavioral fusion features are fused to obtain multi-scale risk characterization features of vehicles.

[0015] In one embodiment, based on a behavioral intent prediction model and multi-source perception data, the behavioral intent of the target behind the vehicle is classified to determine the behavioral intent of the target behind the vehicle, including:

[0016] Multidimensional feature extraction is performed based on multi-source perception data to obtain multiple rear collision risk characterization features of the vehicle;

[0017] Multi-scale feature fusion is performed based on the characteristics of each rear collision risk to determine the multi-scale feature matrix of the vehicle.

[0018] Based on the behavioral intent prediction model and multi-scale feature matrix, causal reasoning is performed on the behavioral intent of the target behind, and the behavioral intent of the target behind is determined.

[0019] In one embodiment, multi-dimensional feature extraction is performed based on multi-source perception data to determine multiple rear-collision risk characterization features of the vehicle, including:

[0020] The multi-source sensing data is preprocessed to obtain preprocessed data; the preprocessing includes invalid removal, spatiotemporal alignment, trajectory tracking and smoothing filtering.

[0021] Feature extraction is performed on the preprocessed data to obtain multiple rear collision risk characterization features of the vehicle.

[0022] In one embodiment, multi-scale feature fusion is performed based on the characteristics of each rear collision risk to determine the vehicle's multi-scale feature matrix, including:

[0023] Multi-scale pyramid branch mapping is performed based on the characteristics of each post-collision risk to obtain the embedding features of each scale pyramid branch.

[0024] Multi-scale feature fusion is performed on each embedded feature to obtain the multi-scale feature matrix of the vehicle.

[0025] In one embodiment, based on a behavioral intent prediction model and a multi-scale feature matrix, causal reasoning is performed on the behavioral intent of the target to determine the behavioral intent of the target, including:

[0026] The multi-scale feature matrix is ​​input into the behavioral intent prediction model, and the posterior probability of the target's upcoming behavioral intent is output.

[0027] The behavioral intent corresponding to the largest posterior probability among all posterior probabilities is determined as the behavioral intent of the target.

[0028] In one embodiment, the collision risk of a vehicle is identified based on the behavioral intent of the rear target, multi-scale risk characterization features, and multi-source perception data to obtain a quantitative value of the vehicle's rear collision risk, including:

[0029] Based on multi-scale risk characterization features and multi-source perception data, a global traffic flow topology map of vehicles is constructed.

[0030] The global traffic flow topology is input into the graph pyramid network model, which outputs multi-scale fused features of vehicles.

[0031] Based on multi-scale fusion characteristics, normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and lane matching degree between the vehicle and the upstream target, the risk transmission coefficient of the risk from the upstream target to the vehicle is determined.

[0032] Based on the causal reasoning model and risk transmission coefficient, the rear collision risk of a vehicle is quantified to obtain a quantitative value of the rear collision risk.

[0033] In one embodiment, the rear-collision risk of a vehicle is quantified based on a causal inference model and a risk transmission coefficient to obtain a quantified value of the rear-collision risk, including:

[0034] Based on the risk transmission coefficient and the causal probability of collision between the upstream target and the vehicle in front, the indirect chain collision causal probability of the upstream target is determined.

[0035] The causal probability of indirect chain collisions, the causal effect weight of the behavioral intention of the rear target on the collision risk, and the prior probability are input into the causal inference model to perform causal association inference and determine the direct collision causal probability of the rear target on the vehicle.

[0036] The maximum value between the causal probability of indirect chain collisions and the causal probability of direct collisions is determined as the quantified value of post-collision risk.

[0037] In one embodiment, a rear collision warning is issued to the vehicle based on a rear collision risk quantification value, including:

[0038] The rear collision risk level of a vehicle is determined based on the rear collision risk quantification value.

[0039] Based on the rear collision risk level, a corresponding warning strategy is obtained, and a rear collision warning is issued to the vehicle according to the warning strategy.

[0040] In one embodiment, a rear collision warning is issued to the vehicle according to a warning strategy, including:

[0041] The vehicle's driving scenario data is determined based on multi-source perception data, and the vehicle's driving style data is determined based on the vehicle's historical driving information.

[0042] Based on vehicle driving scenario data and driving style data, the warning strategy is adjusted, and the adjusted warning strategy is obtained.

[0043] According to the adjusted warning strategy, a rear collision warning will be issued for the vehicle.

[0044] In one embodiment, the method further includes:

[0045] Based on the vehicle's rear collision risk level and multi-source perception data, a closed-loop control strategy is obtained for the vehicle's environment.

[0046] Based on a closed-loop control strategy that considers the vehicle's environment, the vehicle is subjected to safety control.

[0047] Secondly, embodiments of this application provide a vehicle rear collision warning device, the device comprising:

[0048] The feature processing module is used to perform multi-scale feature processing on the multi-source perception data of the vehicle using a spatiotemporal feature fusion model to obtain the multi-scale risk characterization features of the vehicle.

[0049] The classification module is used to classify the behavioral intentions of the vehicle's rear targets based on the behavioral intention prediction model and multi-source perception data, and to determine the behavioral intentions of the rear targets.

[0050] The risk identification module is used to identify the collision risk of a vehicle based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data, and obtain a quantitative value of the vehicle's rear collision risk.

[0051] The warning module is used to provide rear collision warnings to vehicles based on the quantified value of rear collision risk.

[0052] Thirdly, embodiments of this application also provide a vehicle, which includes a memory and an on-board chip. The memory stores a computer program, and the on-board chip executes the computer program to implement the steps of the method in any of the embodiments of the first aspect described above.

[0053] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by an on-board chip, implements the steps of the method in any of the embodiments of the first aspect described above.

[0054] Fifthly, embodiments of this application also provide a computer program product, which includes a computer program that, when executed by an on-board chip, implements the steps of the method in any of the embodiments of the first aspect described above.

[0055] The vehicle rear collision warning method, device, and vehicle provided in this application include: employing a spatiotemporal feature fusion model to perform multi-scale feature processing on multi-source perception data of the vehicle to obtain multi-scale risk characterization features of the vehicle; classifying the behavioral intentions of the target behind the vehicle based on a behavioral intention prediction model and multi-source perception data to determine the behavioral intentions of the target behind the vehicle; identifying the collision risk of the vehicle based on the behavioral intentions of the target behind the vehicle, the multi-scale risk characterization features, and the multi-source perception data to obtain a quantitative value of the vehicle's rear collision risk; and issuing a rear collision warning based on the quantitative value of the rear collision risk. The above method does not require strong assumptions, can proactively avoid risks, seize the golden time for risk avoidance, avoid severe warning lag, and improve the timeliness of warnings. Furthermore, this method no longer relies on a single collision time and a fixed threshold, but extracts the risk characterization features of the vehicle from multiple spatiotemporal dimensions, effectively capturing sudden changes in the behavior of following vehicles and weak risk signals at long distances. Based on this, a rear collision warning is achieved, which can reduce the false alarm rate and the false alarm rate, and improve the accuracy of the warning. At the same time, the above method can identify risk precursors such as rapid acceleration, forced changes, and loss of vehicle control in advance, avoiding the delayed triggering of warnings in traditional solutions, and increasing... The strong and timely early warning provides ample time for evasive maneuvers. Furthermore, the aforementioned methods can address both localized, near-field risks and global cascading risks, are not limited by scenario, and are adaptable to complex conditions such as highways, urban areas, obstructed views, and severe weather, resulting in stronger robustness of rear-collision warnings. Moreover, these methods can incorporate the behavioral intent of following vehicles to trigger rear-collision warnings, distinguishing between normal following and high-risk following behaviors. This makes the warning basis more comprehensive and interpretable, effectively reducing the rate at which users actively disable the rear-collision warning function and increasing the practical significance of mass production and implementation of the rear-collision warning function. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a vehicle rear collision warning method in one embodiment of this application;

[0057] Figure 2 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0058] Figure 3 This is a structural diagram of a spatiotemporal feature fusion model in one embodiment of this application;

[0059] Figure 4 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0060] Figure 5 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0061] Figure 6This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0062] Figure 7 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0063] Figure 8 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0064] Figure 9 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0065] Figure 10 This is a flowchart illustrating the process architecture for determining the pyramid network model structure and risk transmission coefficient in one embodiment of this application.

[0066] Figure 11 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0067] Figure 12 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0068] Figure 13 This is a flowchart illustrating a vehicle rear collision warning method in another embodiment of this application;

[0069] Figure 14 This is a structural block diagram of a vehicle rear collision warning device in one embodiment of this application;

[0070] Figure 15 This is an internal structural diagram of a computer device in one embodiment of this application. Detailed Implementation

[0071] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0072] In the field of vehicle safety, rear collision warning is crucial and a core technology for preventing high-speed rear-end collisions and rear-end injuries to vulnerable road users in urban areas. Related technologies primarily rely on the strong assumption that targets behind the vehicle maintain their current state of motion. They calculate the collision time by collecting basic kinematic parameters such as target distance and speed, and trigger a warning by comparing the collision time with a fixed threshold. However, most rear-end collisions in real-world scenarios stem from sudden behavioral changes by targets behind the vehicle (forced lane changes, loss of vehicle control, drunk driving, or fatigue-related abnormal operations, etc.), rather than deterministic collisions with uniform speed or acceleration. The strong assumptions in these technologies are completely inconsistent with the actual accident mechanisms, triggering warnings only when a collision is unavoidable, thus losing the golden window for proactive risk avoidance. Therefore, these technologies suffer from low accuracy in rear collision warning. To address this, this application provides a vehicle rear collision warning method that improves the accuracy of rear collision warnings.

[0073] The vehicle rear collision warning method provided in this application can be applied to any type of vehicle, which can support intelligent driving functions or driving assistance functions. The vehicle includes an on-board chip, which can also be a domain controller. The following embodiments will specifically describe the specific process of the vehicle rear collision warning method, and will use the on-board chip as the execution subject to describe the specific process of the vehicle rear collision warning method.

[0074] like Figure 1 The diagram shown is a flowchart of a vehicle rear collision warning method provided in an embodiment of this application. The method may include the following steps:

[0075] Step S100: Using a spatiotemporal feature fusion model, multi-scale feature processing is performed on the multi-source perception data of the vehicle to obtain the multi-scale risk characterization features of the vehicle.

[0076] Specifically, the vehicle chip can acquire a pre-trained spatiotemporal feature fusion model and use the spatiotemporal feature fusion model to perform multi-scale feature processing on the multi-source perception data of the vehicle over a period of time to obtain the multi-scale risk characterization features of the vehicle.

[0077] Optionally, the aforementioned spatiotemporal feature fusion model can be implemented by at least one of convolutional neural network models, fully connected neural network models, long short-term memory neural network models, residual neural network models, etc. In this embodiment, the aforementioned spatiotemporal feature fusion model can be a temporal-spatial-bidirectional feature pyramid network (T-BiFPN) model, used to extract multi-scale, full-time-series risk features of rear targets, i.e., multi-scale risk representation features. The aforementioned multi-source perception data is time-series data. It should be noted that the spatiotemporal feature fusion model is trained based on a training dataset; the input variables of this training dataset need to include continuous rear-view frame images, radar target kinematic data, lane line semantic information, and target trajectory time-series data. Therefore, this training dataset can include continuous video streams collected by real-vehicle road test rear-view cameras in high-speed driving, urban road driving, and low-speed parking scenarios, point cloud data collected by vehicle-mounted radar, and global traffic flow data of the vehicle's environment.

[0078] In practical applications, vehicles can be in any driving environment, such as high-speed driving, urban road driving, and low-speed parking. The multi-source perception data of the vehicle can include road test data of the vehicle, pre-acquired publicly available accident datasets, and rear-collision related working condition data collected by the simulation platform, covering all working conditions of the vehicle and the target behind it, including driving, standing, and reversing.

[0079] Optionally, the vehicle's road test data may include point cloud data collected by onboard radars (such as infrared radar, millimeter-wave radar, etc.) and continuous video streams collected in real time by real-time road test rearview cameras installed on the vehicle; publicly available accident datasets may include data on the speed, location, collision pattern, road conditions, and causes of real rear-end collisions; rear-collision related condition data may include extreme condition data such as the behavior of virtual targets behind the vehicle, chain collisions, heavy rain, backlighting, and obstruction. The aforementioned multi-scale risk characterization features of the vehicle can be understood as a multi-dimensional feature set used to comprehensively, finely, and interpretably quantify the rear-collision risk of the vehicle by extracting and fusing features layer by layer from three dimensions: time, space, and interaction. This covers instantaneous, short-term, long-term, local, global, and cluster risks, supporting causal inference. It should be noted that the aforementioned multi-source perception data is time-series perception data, which may include radar data collected by onboard radars, image frame data collected by cameras, and global traffic flow data.

[0080] Step S200: Based on the behavioral intent prediction model and multi-source perception data, classify the behavioral intent of the target behind the vehicle and determine the behavioral intent of the target behind the vehicle.

[0081] In practical applications, the vehicle-mounted chip can input multi-source perception data into a behavior intention prediction model to classify the behavior intentions of the targets behind the vehicle and output the behavior intentions of the targets behind. Optionally, the aforementioned behavior intention prediction model can be implemented by at least one of the following: convolutional neural network model, attention mechanism, fully connected neural network model, long short-term memory neural network model, residual neural network model, etc.

[0082] Among them, the aforementioned rear targets can include all types of traffic participants such as passenger cars, commercial vehicles, non-motorized vehicles, and pedestrians directly behind the vehicle and to the side and rear where there is a possibility of direct collision with the vehicle; the behavioral intention of the rear targets can be understood as the driving action, driving purpose, and planning trend that the vehicle currently wants to perform.

[0083] It should be noted that the execution order of the above steps S100 and S200 can be processed in parallel or sequentially (and the order is not fixed). This application embodiment does not limit this.

[0084] Step S300: Based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data, the collision risk of the vehicle is identified, and the quantitative value of the rear collision risk of the vehicle is obtained.

[0085] Specifically, the vehicle chip can use a risk identification algorithm to identify the collision risk of the vehicle based on the behavioral intent of the target behind, multi-scale risk representation features, and multi-source perception data, and obtain a quantitative value of the vehicle's rear collision risk.

[0086] Optionally, the aforementioned risk identification algorithm can be a traditional rule-based and statistical method, a machine learning method, a deep learning method, etc.

[0087] Among them, the rear collision risk of the aforementioned vehicles can be understood as the potential danger of being rear-ended, scraped, or involved in a chain collision by a target behind (vehicle directly behind and / or vehicle to the side) when the vehicle is in motion or stationary. This includes the risk of direct rear-end collision, the risk of collision with a vehicle changing lanes from the side, and the risk of multi-vehicle chain collisions at high speeds.

[0088] Step S400: Based on the rear collision risk quantification value, a rear collision warning is issued to the vehicle.

[0089] In practical applications, the onboard chip can acquire a pre-trained algorithm model, then input the quantified rear collision risk value of the vehicle into the algorithm model for processing, and output a rear collision warning strategy. The rear collision warning strategy is then applied to the vehicle to issue a rear collision warning. Optionally, the rear collision warning strategy may include rear collision risk behaviors, rear collision risk mitigation measures, and rear collision risk warning information.

[0090] In addition, the onboard chip can perform mapping, matching and / or analysis processing on the quantified value of rear collision risk, obtain the rear collision warning strategy based on the processing results, and then issue a rear collision warning to the vehicle according to the rear collision warning strategy.

[0091] The technical solution in this application employs a spatiotemporal feature fusion model to perform multi-scale feature processing on the vehicle's multi-source perception data, obtaining multi-scale risk representation features of the vehicle. Based on a behavioral intent prediction model and multi-source perception data, the behavioral intent of the target behind the vehicle is classified to determine the behavioral intent of the target behind the vehicle. Based on the behavioral intent of the target behind the vehicle, the multi-scale risk representation features, and the multi-source perception data, the collision risk of the vehicle is identified, obtaining a quantitative value of the vehicle's rear collision risk. A rear collision warning is then issued based on the quantitative value of the rear collision risk. This method does not require strong assumptions, can proactively avoid risks, seizes the golden time for risk avoidance, avoids severe warning delays, and improves the timeliness of warnings. Furthermore, this method no longer relies on a single collision time or a fixed threshold, but extracts the risk representation features of the vehicle from multiple spatiotemporal dimensions, effectively capturing the behavior of following vehicles. By leveraging sudden changes and weak risk signals at long distances to implement rear collision warnings, the false alarm and missed alarm rates can be reduced, improving the accuracy of warnings. Simultaneously, this method can identify early warning characteristics such as rapid acceleration, forced changes, and loss of vehicle control, avoiding the delayed triggering of warnings in traditional solutions, enhancing the timeliness of warnings and allowing sufficient time for evasive maneuvers. Furthermore, this method can address both localized near-distance risks and global cascading risks, is not limited by scenario, and can adapt to complex conditions such as highways, urban areas, obstructed views, and severe weather, making the rear collision warning more robust. Moreover, this method can incorporate the behavioral intent of following vehicles to trigger rear collision warnings, distinguishing between normal following and high-risk following behaviors, making the warning basis more comprehensive and interpretable. This effectively reduces the rate at which users actively disable the rear collision warning function, increasing the practical significance of mass production and implementation of the rear collision warning function.

[0092] The following describes the process of multi-scale feature processing of multi-source perception data of vehicles using the aforementioned spatiotemporal feature fusion model. In one embodiment, the spatiotemporal feature fusion model includes a spatial feature pyramid branch and a temporal feature pyramid branch; as shown... Figure 2 As shown, the process in step S100 above can be implemented in the following way:

[0093] Step S101: Input the multi-source sensing data into the spatial feature pyramid branch for feature extraction to obtain multi-scale spatial fusion features.

[0094] Optionally, the above spatiotemporal feature fusion model includes a spatial feature pyramid branch, a temporal feature pyramid branch, and a fully connected layer; Figure 3This is a structural diagram of the spatiotemporal feature fusion model. The spatial feature pyramid branch is constructed using a convolutional neural network model, which effectively mines the spatial features of targets at different scales within each image frame of the video stream, enabling multi-scale spatial feature extraction. The temporal feature pyramid branch is constructed using a bidirectional long short-term memory (BiLSTM) neural network model with bidirectional learning capabilities, capturing the temporal dependence and temporal continuity of target behavior, enabling multi-scale dynamic feature extraction.

[0095] The aforementioned spatial feature pyramid branch can employ a Pyramid VisionTransformer-Lite (PVT-Lite) model and a 5-level bidirectional weighted feature pyramid structure to achieve multi-scale feature extraction and bidirectional fusion, addressing the problem that existing technologies cannot simultaneously handle fine-grained details and global semantics. In this embodiment, the aforementioned spatial feature pyramid branch may include, but is not limited to, convolutional layers and pooling layers.

[0096] Specifically, the vehicle-mounted chip can input radar data and image frame data from multi-source perception data into the Spatial Feature Pyramid Branch (BiFPN) for feature extraction, thereby obtaining multi-scale spatial fusion features.

[0097] The following is about Figure 3 The implementation process of feature extraction from the spatial feature pyramid branch described above is described below:

[0098] (1) Bottom-up feature extraction branch: The lightweight pyramid visual network (PVT-Lite) model is used as the backbone network. The input image frame data is downsampled at 5 levels to generate feature maps from high resolution to low resolution (P1-P5). Among them, P1-P2 high resolution small scale layer: fine-grained features are extracted through two 3×3 convolutional layers, focusing on millimeter-level details such as vehicle brake lights, turn signal status, wheel rotation posture, and driver facial movements; P3-P4 medium resolution medium scale layer: medium scale features are extracted through two 3×3 convolutional layers and max pooling layer, focusing on behavioral features such as vehicle lane change trajectory, following distance, and vehicle posture; P5 low resolution large scale layer: global semantic features are extracted through one 3×3 convolutional layer and global average pooling layer, focusing on scene features such as vehicle rear traffic topology, road boundary, and global traffic flow distribution.

[0099] (2) Top-down semantic transfer branch: The global semantic features of the high-level P5 are transferred to the low-level layer through upsampling, providing global semantic constraints for the fine-grained features of the low-level layer and solving the problem of feature loss of small targets at long distances;

[0100] (3) Bottom-up feature enhancement branch: The fine-grained features of the lower layer are supplemented to the higher layer by downsampling to correct the judgment bias of the global features of the higher layer;

[0101] (4) Bidirectional weighted fusion: all features at the same level (P1-P5) are weighted and fused with learnable weights. The weighted fusion can be achieved by the following formula (1):

[0102] (1)

[0103] in, This represents the weighted fused features output from the spatial feature pyramid branch, i.e., multi-scale spatial fusion features; Indicates learnable weights; Represents the input features of layer i; This represents a preset minimum constant to avoid a denominator of 0.

[0104] Step S102: Input the multi-source sensing data into the temporal feature pyramid branch for feature extraction to obtain full temporal behavior fusion features.

[0105] In this embodiment, a three-level temporal pyramid structure can be constructed to address the temporal continuity of post-collision behavior, forming a temporal feature pyramid branch. A BiLSTM network model can be used as the backbone network of this branch to capture behavioral trend features at different time scales, thus addressing the limitation of existing technologies that can only identify instantaneous motion states.

[0106] In practical applications, the aforementioned temporal feature pyramid branch can be, but is not limited to, convolutional layers, pooling layers, attention layers, and encoders. Specifically, the vehicle-mounted chip can input image frame data from multi-source perception data into the temporal feature pyramid branch for feature extraction, obtaining full-temporal behavior fusion features.

[0107] The following is about Figure 3 The implementation process of feature extraction from the aforementioned temporal feature pyramid branch is described below:

[0108] (1) Short time sequence layer: A 1-layer BiLSTM network model is adopted. The input is continuous image frame data within the first time period in the multi-source perception data to capture instantaneous change features such as the vehicle's brake lights illuminating, the steering wheel turning sharply, and the acceleration changing suddenly.

[0109] (2) Mid-time layer: A 1-layer BiLSTM network model is adopted. The input is continuous image frame data within the second time period in the multi-source perception data to capture the continuous behavioral features of the vehicle, such as the entire process of lane changing, rapid acceleration / deceleration trend, and frequent line crossing.

[0110] (3) Long time series layer: Two layers of BiLSTM are used. The input is continuous image frame data within the second time period in the multi-source perception data to capture long-term abnormal features such as serpentine driving of vehicles, continuous close following, and driver fatigue and distraction.

[0111] (4) Cross-level temporal fusion: The output features of the three levels of short temporal layer, medium temporal layer and long temporal layer are spliced ​​and fused, and the full temporal behavior fusion features are output through the fully connected layer to achieve full temporal feature coverage of "instantaneous action-continuous behavior-long-term anomaly".

[0112] Step S103: The multi-scale spatial fusion features and the full-time behavior fusion features are fused to obtain the multi-scale risk characterization features of the vehicle.

[0113] The vehicle-mounted chip can employ a feature fusion method to fuse multi-scale spatial fusion features and full-time behavioral fusion features to obtain multi-scale risk characterization features of the vehicle. Optionally, the aforementioned feature fusion method can be feature splicing, weighted fusion, and / or principal component analysis dimensionality reduction fusion, etc. In this embodiment, the aforementioned feature fusion method is a pyramid multi-scale fusion method.

[0114] In this embodiment, the vehicle chip can stitch together and fuse multi-scale spatial fusion features and full-temporal behavioral fusion features through the fully connected layer in the spatiotemporal feature fusion model, and then output the multi-scale risk characterization features of the vehicle.

[0115] It should be noted that, to meet the computing power limitations and real-time requirements of automotive chips, a lightweight approach can be used to optimize the network model across the entire link in this embodiment. This significantly reduces the number of model parameters and computational load without sacrificing feature extraction and inference accuracy. Optionally, the aforementioned lightweight approach can be INT8 symmetric quantization. The lightweight optimization can involve structured pruning of the convolutional channels at each level of each pyramid network model, retaining core causal feature channels and eliminating redundant channels, with a pruning ratio of 40%. Simultaneously, tensor decomposition can be used to optimize the cross-level fusion operators in each pyramid network model, decomposing large convolutional kernels into multiple smaller ones, further reducing computational latency.

[0116] In this embodiment, after lightweighting, the overall parameter size of each network model is less than 8M, the end-to-end inference latency of the automotive-grade chip (i.e., the automotive chip) can be less than or equal to 50ms, and the computing power consumption is less than or equal to 8 trillion integer operations per second. This lightweight optimization can fully meet the requirements of automotive real-time performance and mass production.

[0117] The technical solution in this application embodiment inputs multi-source sensing data into a spatial feature pyramid branch for feature extraction to obtain multi-scale spatial fusion features. Then, it inputs multi-source sensing data into a temporal feature pyramid branch for feature extraction to obtain full-temporal behavior fusion features. Finally, it fuses the multi-scale spatial fusion features and the full-temporal behavior fusion features to obtain multi-scale risk characterization features for the vehicle. This method can extract multi-scale spatial features through a spatial pyramid branch, taking into account both near-range details (such as vehicle distance and lateral offset) and far-range global semantics (such as traffic flow topology and road environment), thus solving the problem of losing far-range features. The problem of incomplete features in missing or occluded scenes is addressed by extracting features that accurately cover the spatial dimension. Simultaneously, by using a temporal pyramid branch, it dynamically captures instantaneous mutations, short-term persistence, and long-term anomalies across the entire temporal dimension, accurately identifying risk precursor signals such as rapid acceleration, rapid deceleration, and frequent changes. This compensates for the deficiency of traditional static features in reflecting dynamic risks. Furthermore, it integrates spatiotemporal features, complementing spatial details and temporal trends to form high-dimensional, highly robust, and comprehensive risk representation features. This provides high-quality input for subsequent intent recognition and risk quantification, fundamentally improving the accuracy and reliability of risk quantification.

[0118] The following describes the process of classifying and determining the behavioral intentions of targets behind a vehicle based on a behavioral intention prediction model and multi-source perception data. In one embodiment, as... Figure 4 As shown, the process in step S200 above can be implemented in the following way:

[0119] Step S201: Multi-dimensional feature extraction is performed based on multi-source perception data to obtain multiple rear collision risk characterization features of the vehicle.

[0120] In practical applications, automotive chips can employ feature extraction algorithms to extract multi-dimensional features from multi-source perception data, obtaining multiple rear-collision risk characterization features of the vehicle. Optionally, the aforementioned feature extraction algorithms can be Kalman filtering, least squares trajectory fitting, principal component analysis, etc.

[0121] Furthermore, the onboard chip can extract multi-dimensional features from multi-source perception data according to preset feature extraction rules, thereby obtaining multiple rear-collision risk characterization features of the vehicle. It should be noted that the onboard chip can calculate each rear-collision risk characterization feature of the vehicle based on at least a portion of the multi-source perception data.

[0122] In practical applications, to reduce the complexity of causal reasoning and improve the accuracy of risk identification, core pre-collision causal features such as vehicle brake light status, wheel slip rate, and abnormal driver actions can be extracted (for accurate collision risk prediction), along with fine-grained features from long distances, features under adverse weather and obstructed scenarios, and multi-scale temporal features (to reduce the false alarm rate). In this embodiment, the aforementioned rear collision risk characterization features may include the basic kinematic features and behavioral causal features of targets behind the vehicle, the global traffic flow features of the vehicle, and the driver's human factors. Tables 1 and 2 show the list of rear collision risk characterization features.

[0123] Table 1

[0124] Serial Number Feature categories Feature Name unit Physical interpretation 1 Basic kinematic characteristics of rear targets relative distance m Longitudinal relative distance between the rear target and the vehicle 2 Basic kinematic characteristics of rear targets relative speed m / s longitudinal relative speed between the target behind and the vehicle 3 Basic kinematic characteristics of rear targets Relative acceleration m / longitudinal relative acceleration between the target behind and the vehicle 4 Basic kinematic characteristics of rear targets Lateral offset m Lateral offset distance of the target behind relative to the lane centerline 5 Causal characteristics of the rear target's behavior Acceleration Impact m / The intensity of a target's rapid acceleration or deceleration is a key indicator of aggressive driving or abnormal maneuvers. 6 Causal characteristics of the rear target's behavior lateral acceleration m / The intensity of lateral maneuvers such as lane changes and serpentine driving by targets behind. 7 Causal characteristics of the rear target's behavior Wheel slip ratio % The degree of risk of the target's wheels locking up, skidding, or losing control. 8 Causal characteristics of the rear target's behavior Following distance (THW) s The time interval between the target behind and the vehicle in front represents the following risk. 9 Causal characteristics of the rear target's behavior Percentage of time spent crossing lane lines % The percentage of time a rear target's wheels are on the lane line per unit of time indicates the intention to change lanes. 10 Causal characteristics of the rear target's behavior Lighting status characteristics — The illumination status of turn signals, brake lights, and hazard lights of vehicles behind is represented by categorized codes.

[0125] Table 2

[0126] Serial Number Feature categories Feature Name unit Physical interpretation 11 Global traffic flow characteristics Target lane traffic density vehicles / km The density of traffic flow in the lanes surrounding a vehicle indicates the feasibility of changing lanes to avoid danger. 12 Global traffic flow characteristics Upstream target collision risk value — The probability of a collision between a rear-facing upstream target and the vehicle in front represents the risk of chain collisions. 13 Driver's human characteristics Mean reaction time s The driver's braking / steering response time to historical warnings characterizes the driver's responsiveness. 14 Driver's human characteristics Driving Style Coefficient — Aggressive / conservative tendencies quantified based on the driver's historical acceleration, deceleration, and lane-changing behavior. 15 Driver's human characteristics Rearview mirror observation frequency times / min The frequency with which a driver checks the interior / exterior rearview mirrors reflects their level of awareness of rear risks.

[0127] In this embodiment, the upstream target can be understood as a vehicle located further behind all the rear targets and upstream of the traffic flow risk transmission chain. For example, [this vehicle] ← [rear target (the first vehicle directly behind this vehicle)] ← [upstream target 1] ← [upstream target 2] ← ... ← [upstream target n].

[0128] The above-mentioned acceleration impact ( The magnitude of acceleration impact indicates the stability and risk level of the driving behavior of the target behind the vehicle. A larger acceleration impact indicates that the target behind the vehicle has obvious abnormal behaviors such as rapid acceleration, rapid deceleration, and frequent lane changes. It is the core precursor signal of the rear collision risk of the vehicle and can be expressed by the following formula (2):

[0129] (2)

[0130] in, This represents the longitudinal acceleration of the target behind, in units of... .

[0131] Meanwhile, the aforementioned wheel slip ratio ( The wheel slip ratio represents the risk of loss of control of the target behind; if the vehicle's wheel slip ratio exceeds 20%, the vehicle has a high risk of brake lock-up and sideslip loss of control. In the embodiments of this application, the wheel slip ratio can be expressed by the following formula (3):

[0132] (3)

[0133] in, This indicates the vehicle's speed, measured in m / s. This indicates the angular velocity of the vehicle's wheels, measured in rad / s. This indicates the rolling radius of the vehicle's wheels, measured in meters (m).

[0134] In this application embodiment, each rear collision risk characterization feature can be divided into continuous numerical features (including relative distance, relative speed, relative acceleration, lateral offset, lateral acceleration, wheel slip ratio, following distance, target lane traffic density, upstream target collision risk value, average reaction time, driving style coefficient, and rearview mirror observation frequency), categorized coding features (including light status features), and temporal trajectory features (including acceleration impact and lane line crossing duration percentage).

[0135] Step S202: Perform multi-scale feature fusion based on the characteristics of each rear collision risk to determine the multi-scale feature matrix of the vehicle.

[0136] Specifically, the vehicle chip can use a feature fusion method to perform multi-scale feature fusion on the various rear collision risk characterization features to obtain the vehicle's multi-scale feature matrix.

[0137] Optionally, the aforementioned feature fusion method can be feature splicing, weighted fusion, and / or principal component analysis dimensionality reduction fusion, etc.

[0138] Step S203: Based on the behavioral intention prediction model and the multi-scale feature matrix, perform causal reasoning on the behavioral intention of the target behind to determine the behavioral intention of the target behind.

[0139] In practical applications, the vehicle-mounted chip can input a multi-scale feature matrix into the behavior intention prediction model, perform causal reasoning on the behavior intention of the target behind, and then output the behavior intention of the target behind.

[0140] The technical solution in this application embodiment extracts multi-dimensional features from multi-source perception data to obtain multiple rear-collision risk characterization features of the vehicle. Multi-scale feature fusion is then performed based on these features to determine the vehicle's multi-scale feature matrix. Finally, causal inference is conducted on the behavioral intent prediction model and the multi-scale feature matrix to determine the behavioral intent of the rear target. This method comprehensively covers rear-collision risk characterization features of various risk information such as distance, speed, acceleration, attitude, and lighting through multi-dimensional feature extraction, avoiding the one-sidedness caused by single features and ensuring the integrity of feature information. Furthermore, multi-scale feature fusion integrates features of different spatiotemporal granularities, taking into account both local details and global trends, enhancing the distinguishability and robustness of features, reducing noise interference, and performing causal inference based on the multi-scale feature matrix rather than simple data fitting. This enables accurate identification of the rear target's intent from the perspective of risk causes, distinguishing between normal and high-risk behaviors, and improving the accuracy and interpretability of behavioral intent recognition. In addition, the method combines multi-dimensional features with causal inference, providing a reliable basis for subsequent risk quantification, reducing the false alarm rate, and improving the overall accuracy of early warnings.

[0141] The process of extracting multiple rear-collision risk characteristics of a vehicle based on multi-source perception data is described below. In one embodiment, as... Figure 5 As shown, the process in step S201 above may include:

[0142] Step S211: Preprocess the multi-source sensing data to obtain preprocessed data; the preprocessing includes invalid removal, spatiotemporal alignment, trajectory tracking and smoothing filtering.

[0143] In practical applications, the vehicle-mounted chip can preprocess multi-source sensing data to obtain preprocessed data. Optionally, the above preprocessing may include data cleaning, standardization, normalization, noise reduction, interpolation, etc. However, in the embodiments of this application, the above preprocessing may include invalid data removal, spatiotemporal alignment, trajectory tracking, and smoothing filtering.

[0144] Optionally, the aforementioned invalid data removal can include pre-removal and post-removal. Pre-removal can be understood as removing glitches and abrupt data (abnormal and discrete data) caused by momentary malfunctions, electromagnetic interference, or occlusions in multi-source perception data. The aforementioned spatiotemporal alignment can be understood as aligning the timestamps of data collected by different types of sensors based on the pre-removed data. The aforementioned trajectory tracking can be understood as processing the target based on the temporal continuity of the target trajectory, including target association, effective target trajectory filtering (e.g., acquiring data with continuous vehicle driving time exceeding 3 seconds), and unified ID labeling. Furthermore, post-removal can be performed on the trajectory-tracked data, specifically removing invalid data such as continuous lane changes across multiple lanes and momentary occlusions, and then labeling it to form various types of operating condition data, including effective rear-collision related operating condition data, real accident operating condition data, high-risk operating condition data, and normal driving operating condition data.

[0145] Since the data collected by vehicle-mounted radar and cameras inevitably contains Gaussian noise and random noise, to ensure the reliability and stability of the data, the data obtained after further filtering can be smoothed and filtered. This smoothing and filtering process can employ methods such as moving averages and Kalman filtering to remove fluctuations caused by sensor noise and electromagnetic interference. In this embodiment, the smoothing and filtering process can be a symmetric exponential moving average filter. The symmetric exponential moving average filter is used to... Data obtained after removing time Smoothing filtering can be performed using the following formula (4):

[0146] (4)

[0147] in, express The smoothed value after filtering at each time step. express Forward filter value at time 10:00 express The back filter value at time 1. This represents the smoothing coefficient. In this embodiment, data smoothness and feature response speed can be balanced by taking any value between 0 and 1. It can be equal to .

[0148] Step S221: Feature extraction is performed on the preprocessed data to obtain multiple rear collision risk characterization features of the vehicle.

[0149] Furthermore, the onboard chip can employ feature extraction algorithms to extract features from the preprocessed data, thereby obtaining multiple rear-collision risk characterization features of the vehicle.

[0150] In one embodiment, such as Figure 6 As shown, the process of determining the vehicle's multi-scale feature matrix by performing multi-scale feature fusion based on the characteristics of each rear collision risk in step S202 above may include:

[0151] Step S202: Perform multi-scale pyramid branch mapping based on the characteristics of each post-collision risk to obtain the embedding features of each scale pyramid branch.

[0152] Among them, the vehicle chip can also perform multi-scale pyramid branch mapping according to the preset feature mapping strategy and the characteristics of each rear collision risk to obtain the embedding features of each scale pyramid branch.

[0153] In practical applications, the vehicle-mounted chip can generate a risk characterization feature matrix based on the risk characterization features of each rear collision, and then perform standardization processing on the risk characterization feature matrix to obtain a standardized feature matrix to eliminate differences in dimensions and scales. This standardization processing can be achieved through the following formula (5), namely:

[0154] (5)

[0155] It should be noted that each type of operating condition data is processed in parallel and independently, and each type of operating condition data can include n samples based on continuous time-series features. This represents the standardized value of the j-th feature of the n-th sample; This represents the rear collision risk characterization feature corresponding to the i-th row and j-th column in the risk characterization feature matrix.

[0156] Optionally, separate pyramid embedding branches can be constructed for small-scale fine-grained features, medium-scale behavioral features, and large-scale global features, respectively. The post-collision risk representation features of different scales are mapped to a unified high-dimensional semantic space through a fully connected layer. Each pyramid embedding branch can be implemented using the following formula (6):

[0157] (6)

[0158] in, This represents the embedding feature of the k-th scale pyramid embedding branch. , Let represent the weight matrix and bias term of the k-th scale pyramid embedding branch, respectively. Represents the ReLU activation function. This represents the normalized feature matrix for the k-th scale, where k represents large, small, or medium scale. Optionally, each normalized feature matrix can be constructed from features selected from the normalized feature matrix based on screening criteria for features at different scales.

[0159] In the embodiments of this application, there are significant differences in the dimensions, scales, and data types of different post-collision risk representation features. At the same time, some post-collision risk representation features are coupled and correlated. Directly inputting them into the network model for processing will lead to increased computational complexity and the problem that small-scale fine-grained features are submerged by large-scale features. Therefore, the embodiments of this application can use a multi-scale pyramid embedding network to reduce the dimensionality and unify the representation of each standardized feature matrix, so as to retain the causal correlation characteristics of features at different scales while reducing the dimensionality.

[0160] Step S203: Perform multi-scale feature fusion on each embedded feature to obtain the multi-scale feature matrix of the vehicle.

[0161] Furthermore, a feature fusion method can be used to fuse the embedded features at multiple scales to obtain the vehicle's multi-scale feature matrix.

[0162] In this embodiment, the embedded features of each pyramid branch at each scale can be weighted and fused using an attention mechanism to calculate the causal contribution rate of each feature dimension, retain the main feature dimension with a cumulative contribution rate of over 95%, and complete the dimensionality reduction process.

[0163] The aforementioned weighted fusion and contribution rate can be achieved through the following formula (7):

[0164] (7)

[0165] In the above formula, This represents the attention weight of the k-th scale pyramid branch; This represents the multi-scale feature matrix of the fused vehicle. This indicates the degree of causal influence of the j-th feature dimension on the result of "collision risk warning"; A label indicating the risk of a rear collision; This represents the function for calculating covariance.

[0166] It should be noted here that... The value range is [0, 1], and the sum of the causal influence levels corresponding to all feature dimensions is 1. In the embodiments of this application, all multi-scale feature matrices can be processed to achieve early warning, or some features can be selected from all multi-scale feature matrices according to the degree of causal influence for further processing to achieve early warning.

[0167] Meanwhile, the labeling of rear-collision risk This represents a discrete value that characterizes the true extent of the rear collision risk at any given time. The essence of causal influence is to quantify the degree to which a change in the j-th feature will cause the rear collision risk label Y to transition from low risk to high risk. Table 3 shows the labels. The classification criteria.

[0168] Table 3

[0169] Tag Y Level of rear collision risk definition Basis for labeling preprocessed data 0 Risk-free There is no possibility of a rear collision. Under normal driving conditions, with no abnormal behavior from the target behind, there is no risk of collision within the next 10 seconds. 1 Low risk There is a potential for a collision, but the risk is extremely low. The vehicle behind is following normally and has not engaged in any unusual maneuvers; the probability of a collision within the next 10 seconds is less than 10%. 2 Medium risk There is a clear possibility of a rear-end collision, and the driver should be aware. If a target behind exhibits initial abnormal behavior (such as frequently crossing the line or following too closely), the probability of a collision within the next 5 seconds is 10% to 30%. 3 High risk The risk of a rear-end collision is extremely high, and evasive action must be taken immediately. If a target behind exhibits clearly dangerous behavior (such as a forced lane change or rapid acceleration), the probability of a collision within the next 3 seconds is 30% to 90%.

[0170] In order to avoid the one-sidedness of information at a single scale, this embodiment can map collision risk representation features according to scale, embed and fuse the original risk features in layers, effectively improve the feature quality and expressive power, achieve full-dimensional and full-granular coverage of risk features, and provide reliable basic data for subsequent accurate identification and early warning.

[0171] The technical solution in this application embodiment preprocesses multi-source perception data to obtain preprocessed data, and extracts features from the preprocessed data to obtain multiple rear collision risk characterization features of the vehicle. The above method can first preprocess the original multi-source perception data and then extract features from the preprocessed data, which can improve the accuracy of the feature extraction results and provide reliable basic data for subsequent accurate identification and early warning.

[0172] The following describes the process of causal inference of the behavioral intention of the target based on the behavioral intention prediction model and the multi-scale feature matrix. In one embodiment, as... Figure 7 As shown, the process of step S203 above may include:

[0173] Step S213: Input the multi-scale feature matrix into the behavior intention prediction model and output the posterior probability of the target's upcoming behavior intentions.

[0174] It should be noted that after completing the feature dimensionality reduction process, in order to accurately identify the differences in collision risk corresponding to different behaviors of the target behind, the behavioral intentions of the target behind can be classified to construct a causal link between features, intentions, and risks.

[0175] Specifically, the vehicle-mounted chip can input a multi-scale feature matrix into the behavior intention prediction model, and output the posterior probability of various behavior intentions that the target behind will soon perform. In the embodiments of this application, in order to output the posterior probabilities of different behavior intentions and fully adapt to the subsequent causal inference architecture, the above-mentioned behavior intention prediction model can be a Bayesian network model.

[0176] Optionally, the aforementioned behavioral intent may include, but is not limited to, normal following, aggressive following, intention to forcibly change lanes, intention to brake suddenly, loss of vehicle control, and abnormal driver operation.

[0177] In this embodiment of the application, the behavioral intent prediction model can use formula (8) to process the multi-scale feature matrix to obtain the posterior probability of the target's upcoming behavioral intents, that is:

[0178] (8)

[0179] in, This represents the posterior probability that the target behind is about to perform the i-th type of action. Let represent the class-conditional probability that the target behind is about to perform the i-th type of behavior. This represents the prior probability that the target behind exhibits the intention to perform the i-th type of behavior.

[0180] Step S223: Determine the behavioral intention corresponding to the largest posterior probability among all posterior probabilities as the behavioral intention of the target.

[0181] In practical applications, the vehicle-mounted chip can compare or take the maximum value of each posterior probability, and determine the behavioral intent corresponding to the maximum posterior probability among all posterior probabilities as the behavioral intent of the target behind.

[0182] In the technical solution of this application embodiment, a multi-scale feature matrix is ​​input into the behavior intention prediction model, and the posterior probability of each behavior intention that the target is about to perform is output. The behavior intention corresponding to the maximum posterior probability among the posterior probabilities is determined as the behavior intention of the target. The above method can use a high-quality and comprehensive multi-scale feature matrix as input for behavior intention determination, providing high-quality basic data for behavior intention determination and reducing misjudgment of behavior intention due to feature bias or noise. At the same time, the above method can first obtain the posterior probability of different behavior intentions that the target is about to perform, so as to quantify the possibility of each behavior intention, transforming fuzzy behavior judgment into a probabilistic expression, avoiding absolute judgment, and improving the scientificity and interpretability of behavior intention determination. In addition, using the maximum posterior probability to determine the final behavior intention can simplify the decision-making logic while ensuring accuracy.

[0183] The process of identifying vehicle collision risks based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data is described below. In one embodiment, as... Figure 8 As shown, the process in step S300 above can be implemented in the following way:

[0184] Step S310: Construct a global traffic flow topology map of vehicles based on multi-scale risk characterization features and multi-source perception data.

[0185] In practical applications, the vehicle chip can also call an online topology mapping tool, inputting multi-scale risk characterization features and multi-source perception data into the online topology mapping tool, which outputs a global traffic flow topology map of the vehicle.

[0186] In addition, the onboard chip can also acquire a pre-trained topology graph construction model, and then input multi-scale risk representation features and multi-source perception data into the topology graph construction model. The topology graph construction model outputs a constructed global traffic flow topology graph of the vehicle.

[0187] Optionally, the aforementioned topology graph construction model may include at least one implementation of convolutional neural network model, fully connected neural network model, residual neural network model, recurrent recurrent neural network model, long short-term memory neural network model, etc.

[0188] Step S320: Input the global traffic flow topology map into the graph pyramid network model to output the multi-scale fusion features of vehicles.

[0189] It should be noted that existing rear collision warning technologies can only identify direct collision risks of single targets within the vehicle's sensor line of sight, completely failing to cover chain collision risks beyond visual range and across lanes, which account for more than 60% of highway rear collision accidents. Based on this, this application embodiment can construct a Graph Pyramid Network (GPN) model to achieve multi-scale risk transmission in global traffic flow and prediction and quantification of beyond-visual-range chain collision risks, improving the overall accident coverage and filling gaps in scenario coverage. Simultaneously, traditional Graph Neural Network (GNN) models can only handle single-scale graph structures and suffer from low computational efficiency in complex traffic flow scenarios. This application embodiment can construct a three-level Graph Pyramid Network model to achieve the aggregation of multi-scale risk features and accurate calculation of transmission coefficients.

[0190] Specifically, the onboard chip can input the global traffic flow topology map into the graph pyramid network model and output the multi-scale fusion features of vehicles.

[0191] Specifically, a directed weighted traffic flow topology graph, i.e., a global traffic flow topology graph, can be constructed based on radar data, image frame data, and global traffic flow data of the vehicle's environment. , Represents a set of nodes. Represents the edge set (including directed edges) ), Represents the weight set (including edge weights) Global traffic flow topology maps can characterize the risk relationships between vehicles, overcoming the line-of-sight limitations of vehicle sensors.

[0192] Each node in the global traffic flow topology graph corresponds to a single vehicle in the traffic flow. Node attributes include vehicle position, speed, acceleration, heading angle, and collision risk probability. The directed edges in the global traffic flow topology graph... Represents a node With nodes There are risk associations among them, including following relationships between vehicles, parallel relationships in adjacent lanes, and lane change conflicts; edge weights Represents a node Risk to nodes The intensity of risk transmission is dynamically calculated based on the relative distance, relative speed, lane matching degree, and feasibility of lane change and avoidance by the vehicle in front between the two nodes. Optionally, the lane matching degree can be a topological association parameter calculated based on the lateral position of the target lane and the semantic information of the lane lines. The lane matching degree is 1.0 when the target lane is the same lane, 0.6 when the target lane is an adjacent lane, and 0 when the target lane is a non-adjacent cross-lane lane.

[0193] Among them, edge weight This can be achieved using the following formula (9):

[0194] (9)

[0195] In the formula, The Sigmoid activation function is used to normalize edge weights to the [0, 1] interval; The weights are learnable coefficients (any value between 0 and 1), obtained by training on a public accident dataset, with the optimal value being... , , , ; This is the normalized relative distance; Normalized relative velocity; Lane matching degree; Feasibility of changing lanes to avoid the vehicle in front.

[0196] The processing procedure for the graph pyramid network model is described here:

[0197] (1) Hierarchical pooling and feature aggregation

[0198] The bottom fine-grained layer uses individual vehicles as nodes, inputting the multi-scale risk representation features of a single vehicle from the global traffic flow topology map. It extracts direct risk association features between nodes (such as the direct impact of a vehicle braking suddenly on a following vehicle) through graph convolutional layers (GCN), outputting updated node features (each node carrying direct risk information with adjacent vehicles), which serve as input to the middle mesoscale layer. The middle mesoscale layer uses graph clustering pooling to aggregate associated vehicles in the same lane or adjacent lanes into vehicle cluster nodes, extracting risk transmission features between vehicle cluster nodes, which serve as input to the top global scale layer. The top global scale layer uses global pooling to aggregate vehicle clusters across the entire road segment into topological nodes across the entire road segment, extracting global risk diffusion trend features, i.e., global risk diffusion features.

[0199] (2) Upsampling and feature correction: The global risk diffusion features are transmitted to the middle mesoscale layer and the bottom fine-grained layer through upsampling, and the features of the middle mesoscale layer and the bottom fine-grained layer are corrected to realize the constraint of global risk on local risk.

[0200] Step S330: Based on the multi-scale fusion features, the normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and the lane matching degree between the vehicle and the upstream target, determine the risk transmission coefficient from the upstream target to the vehicle.

[0201] Specifically, the vehicle-mounted chip can perform arithmetic operations on multi-scale fusion features, normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and lane matching degree between the vehicle and the upstream target to obtain the risk transmission coefficient of the risk from the upstream target to the vehicle.

[0202] Optionally, the above arithmetic operations may include at least one of addition, subtraction, multiplication, logarithmic, exponential, and division operations.

[0203] In addition, the vehicle chip can also acquire a pre-trained algorithm model, and then input the multi-scale fusion features, the normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and the lane matching degree between the vehicle and the upstream target into the algorithm model to obtain the risk transmission coefficient of the risk from the upstream target to the vehicle.

[0204] In this embodiment, the risk transmission coefficient from upstream targets to vehicles can be calculated based on the multi-scale fusion features output by the bottom fine-grained layer. The value ranges from 0 to 1.0, and the calculation formula is as follows:

[0205] (10)

[0206] in, Use the Sigmoid activation function; This represents the normalized relative distance between the upstream target and the vehicle. Normalized relative velocity; For lane matching accuracy, use 1.0 for the same lane, 0.6 for adjacent lanes, and 0 for cross-lane lanes; The feasibility of the vehicle changing lanes to avoid the vehicle in front is assessed, with a value ranging from 0 to 1.0. These are learnable weights.

[0207] Step S340: Based on the causal reasoning model and risk transmission coefficient, the rear collision risk of the vehicle is quantified to obtain the quantified value of the rear collision risk.

[0208] In practical applications, the above causal reasoning model can be a chain structure, a branching structure, a convergence structure, a structural causal model, a dynamic causal structure, etc., but in the embodiments of this application, the above causal reasoning model can be a Bayesian causal network model.

[0209] It should be noted that the collision risk quantification value will replace the traditional fixed collision time threshold triggering logic, which can provide interpretable and highly robust dependency data for early warning decisions.

[0210] The technical solution in this application constructs a global traffic flow topology map of vehicles based on multi-scale risk characterization features and multi-source perception data. This global traffic flow topology map is then input into a graph pyramid network model to output the multi-scale fusion features of the vehicles. Based on the multi-scale fusion features, the normalized relative distance and normalized relative speed between the vehicle and its corresponding upstream target, and the lane matching degree between the vehicle and the upstream target, a risk transmission coefficient is determined to transmit the risk from the upstream target to the vehicle. Based on a causal inference model and the risk transmission coefficient, the rear-collision risk of the vehicle is quantified to obtain a quantified value. This method, by constructing a global traffic flow topology map of vehicles, can comprehensively depict the direct and indirect, line-of-sight and beyond-line-of-sight risk relationships between vehicles, overcoming the limitations of single-target risk perception and effectively covering multi-vehicle chain collisions. The scenario involves inputting the topology graph into a graph pyramid network model to extract and fuse features at multiple scales, including single vehicle, vehicle group, and global levels. This approach considers both local risk details and global traffic flow trends, accurately capturing cross-lane and multi-hop risk transmission paths. Based on this, the risk transmission coefficient from upstream targets to vehicles can be determined, quantifying the intensity of upstream risk transmission and making risk transmission judgments more accurate and objective. Furthermore, the above method can rely on a causal inference model to fuse risk transmission coefficients for risk quantification, moving away from the traditional fixed threshold mode for prediction. It achieves accurate differentiation between direct and chain collision risks from the perspective of risk causes and transmission patterns, effectively reducing the false alarm and missed alarm rates of chain collision risks. This enables a global, dynamic, and interpretable post-collision risk assessment, significantly improving the accuracy, robustness, and reliability of early warnings in complex scenarios.

[0211] The following describes the process of quantifying the rear-collision risk of a vehicle based on the causal reasoning model and risk transmission coefficient. In one embodiment, as... Figure 9 As shown, the process in step S340 above can be implemented in the following way:

[0212] Step S341: Based on the risk transmission coefficient and the collision causal probability of the upstream target to its preceding vehicle, determine the indirect chain collision causal probability of the upstream target.

[0213] In practical applications, the vehicle chip can process the risk transmission coefficient and the causal probability of the collision between the upstream target and the vehicle in front according to the preset processing strategy, so as to obtain the indirect chain collision causal probability of the upstream target.

[0214] Simultaneously, the onboard chip can perform arithmetic operations on the risk transmission coefficient and the causal probability of a collision between an upstream target and its preceding vehicle to obtain the indirect chain collision causal probability of the upstream target. Optionally, the preceding vehicle of the upstream target can be understood as the vehicle directly in front of the upstream target, that is, the downstream vehicle immediately adjacent to the upstream target in the risk transmission chain.

[0215] In this embodiment of the application, the vehicle-mounted chip can process the risk transmission coefficient and the collision causal probability of the upstream target to its preceding vehicle according to the following formula (11) to obtain the indirect chain collision causal probability of the upstream target. ,like:

[0216] (11)

[0217] In the above formula, This represents the causal probability of an upstream target colliding with the vehicle in front of it. This represents the risk transmission coefficient, indicating the transmission of risk from the upstream target to the vehicle. For example... Figure 10 The diagram shows the structure of the graph pyramid network model and the process architecture for determining the risk transmission coefficient.

[0218] Step S342: Input the indirect chain collision causal probability, the causal effect weight of the rear target's behavioral intention on the collision risk and the prior probability into the causal inference model to perform causal association inference and determine the direct collision causal probability of the rear target on the vehicle.

[0219] Specifically, the vehicle-mounted chip can input the causal probability of indirect chain collisions, the causal effect weight of the behavioral intention of the rear target on the collision risk, and the prior probability into the causal inference model to perform causal association inference, and then output the direct collision causal probability of the rear target on the vehicle.

[0220] In this embodiment of the application, the causal probability of a direct collision between a rear target and a vehicle is determined based on a causal inference model. The implementation process can be expressed by the following formula (12):

[0221] (12)

[0222] in, This represents the scene correction factor, with a value range of 0.5 to 2.0. It can be dynamically adjusted based on the vehicle's speed, the type of road it is on, and the road surface adhesion coefficient. Indicates the first The causal effect weights of behavioral intentions on collision risk are obtained through training using a causal discovery algorithm.

[0223] Step S343: The maximum value between the indirect chain collision causal probability and the direct collision causal probability is determined as the post-collision risk quantification value.

[0224] The onboard chip can compare or maximize the causal probabilities of indirect chain collisions and direct collisions, and determine the maximum value of the two as the quantified value of the vehicle's rear collision risk. ,Right now .

[0225] The technical solution in this application embodiment determines the indirect chain collision causal probability of the upstream target based on the risk transmission coefficient and the collision causal probability of the upstream target to its preceding vehicle. The indirect chain collision causal probability, the causal effect weight of the rear target's behavioral intent on the collision risk, and the prior probability are input into a causal inference model for causal association inference to determine the direct collision causal probability of the rear target to the vehicle. The maximum value between the indirect chain collision causal probability and the direct collision causal probability is determined as the quantified value of the rear collision risk. This method utilizes the risk transmission coefficient of the upstream target's risk to the vehicle and the collision causal probability of the upstream target to its preceding vehicle. The method accurately calculates the causal probability of indirect chain collisions with upstream targets, effectively identifying chain risks beyond visual range and across lanes. Simultaneously, it combines the causal effect weights of the behavioral intentions of rear targets with prior probabilities to achieve causal association reasoning, scientifically determining the causal probability of direct collisions between vehicles. This differentiates the risk contribution corresponding to different behavioral intentions, making risk assessment more interpretable. Furthermore, the method comprehensively integrates both direct and chain collision risks, avoiding missed warnings due to a single risk type. It achieves comprehensive and accurate quantification of rear collision risks, significantly improving the completeness and reliability of risk assessment in highly complex traffic flow environments.

[0226] The process of providing rear-collision warning for vehicles based on the quantified rear-collision risk value is described below. In one embodiment, as... Figure 11 As shown, the process in step S400 above may include:

[0227] Step S410: Determine the rear collision risk level of the vehicle based on the rear collision risk quantification value.

[0228] In practical applications, the onboard chip can call a classification tool, input the vehicle's rear collision risk quantification value into the classification tool, and the classification tool will return the vehicle's rear collision risk level.

[0229] Meanwhile, the onboard chip can obtain the correlation between different collision risk quantification values ​​and different collision risk levels, and compare the vehicle's rear collision risk quantification value with each collision risk quantification value in the correlation, and determine the collision risk level corresponding to the collision risk quantification value that is equal to the rear collision risk quantification value as the vehicle's rear collision risk level.

[0230] In this embodiment of the application, the rear collision risk of a vehicle can be divided into 5 levels (as shown in Table 4) based on the rear collision risk quantification value, corresponding to different warning and control logics, so as to realize the closed-loop logic of the entire chain from "early warning reminder - pre-safety preparation - active intervention - collision protection" of the vehicle.

[0231] Table 4

[0232] Collision risk level Rear Collision Risk Quantification Early warning strategy Risk-free Without any warning actions, the vehicle maintains its normal driving state. Low risk Level 1 Warning: The vehicle's infotainment display shows a rear-end target risk warning, illuminates the vehicle's brake lights, and provides a two-way alert to the driver and the target behind. Medium risk Level 2 Warning: A red warning light on the instrument panel and an audible and visual warning sound, hazard warning lights (double flashers) illuminate, seatbelt pretensioning and braking system pre-pressurization are activated, and collision preparation is completed. High risk Level 3 intervention: Enhanced audible and visual warnings with seat vibration and continuous buzzing, assisting the driver in making minor lane changes or accelerating to avoid danger within safe boundaries; automatically applying the electronic parking brake when the vehicle is stationary to prevent rolling after a rear-end collision. Extremely high risk Level 4 protection: Full and stable braking force is maintained, doors are locked, seat headrests and seating position are adjusted, windows are closed, seat belts are continuously tightened, airbag deployment logic is optimized, and collision risks are broadcast to surrounding vehicles via vehicle-to-vehicle communication.

[0233] Step S420: Obtain the corresponding warning strategy based on the rear collision risk level, and issue a rear collision warning to the vehicle according to the warning strategy.

[0234] In this embodiment of the application, the vehicle chip can acquire a pre-trained algorithm model, then input the rear collision risk level of the vehicle into the algorithm model and output the corresponding warning strategy, and provide a rear collision warning to the vehicle according to the warning strategy.

[0235] In addition, the onboard chip can also obtain an information mapping table between different collision risk levels and different warning strategies, and obtain the warning strategy corresponding to the collision risk level that matches the rear collision risk level from the information mapping table, as the warning strategy corresponding to the vehicle.

[0236] In practical applications, when the vehicle's collision risk probability, risk transmission coefficient, relative distance and / or speed exceed the corresponding warning threshold, the on-board chip can trigger a rear collision warning.

[0237] The technical solution in this application determines the rear-collision risk level of a vehicle based on the rear-collision risk quantification value, obtains a corresponding warning strategy based on the rear-collision risk level, and issues a rear-collision warning to the vehicle according to the warning strategy. This method matches the vehicle's rear-collision risk quantification value with its rear-collision risk level, transforming the abstract risk quantification result into a clear and distinguishable level classification. This solves the problem of high false alarm rates in traditional warning systems that use fixed collision time thresholds to trigger warnings and cannot adapt to different scenario requirements. Furthermore, this method can match corresponding warning strategies based on risk levels to achieve tiered warnings and avoid... This method addresses the drawbacks of excessive warnings that interfere with drivers in low-risk scenarios and insufficient warnings in high-risk scenarios, balancing the effectiveness and comfort of warnings. Furthermore, the method can implement rear-collision warnings based on an appropriate warning strategy, accurately delivering warning information according to risk levels. This provides drivers with clear guidance on hazard avoidance and can trigger high-intensity warnings in high-risk scenarios, allowing sufficient time for hazard avoidance maneuvers. Simultaneously, it reduces warning interference in low-risk scenarios, improving the driving experience and further enhancing the practicality, accuracy, and user-friendliness of rear-collision warnings. It ensures warning effectiveness while also prioritizing driving safety and comfort.

[0238] The process of providing rear collision warning for vehicles based on the aforementioned warning strategy is described below. In one embodiment, as... Figure 12 As shown, the process of providing a rear collision warning to the vehicle according to the warning strategy in step S420 above may include:

[0239] Step S421: Determine the vehicle's driving scenario data based on multi-source perception data, and determine the vehicle's driving style data based on the vehicle's historical driving information.

[0240] In practical applications, automotive chips can perform identification, matching, mapping, and analysis on at least some of the data from multi-source sensing data to obtain vehicle driving scenario data; and can perform identification, matching, mapping, and analysis on the vehicle's historical driving information to obtain vehicle driving style data.

[0241] In addition, the onboard chip can input multi-source perception data and the vehicle's historical driving information into a pre-trained algorithm model, which outputs the vehicle's driving scene data and driving style data.

[0242] Optionally, the vehicle's driving scenario data may include scenarios such as highways, urban roads, and low-speed parking where the vehicle is located, as well as the interaction information in these scenarios. This interaction information may include traffic flow, vulnerable road users, intersections, traffic lights, oncoming traffic, turning, and other traffic interaction conditions.

[0243] Meanwhile, the vehicle's historical driving information can include the driver's driving behavior and response data to historical warnings over a period of time prior to the current moment. Specifically, it includes multi-dimensional information such as the driver's acceleration and deceleration habits, lane change frequency, response time to warnings, line of sight status, and frequency of rearview mirror observation.

[0244] Once a driver completes a total of 15 valid driving cycles, the onboard chip can automatically identify the driver's human characteristics and driving style based on the vehicle's historical driving information, and classify them into three types of driving styles: aggressive, conventional, and conservative. Correspondingly, the vehicle's driving style data can include data on the types of aggressive, conventional, and conservative driving styles.

[0245] Step S422: Adjust the warning strategy based on the vehicle's driving scenario data and driving style data, and obtain the adjusted warning strategy.

[0246] In this embodiment, to improve the adaptability of the warning strategy across all scenarios, for the same rear collision risk level, corresponding control logic can be formulated based on the vehicle's driving scenario and driving style data. Specifically, the warning strategy can be jointly adjusted based on the vehicle's driving scenario data and driving style data to obtain the adjusted warning strategy.

[0247] Among them, the onboard chip can adjust the warning strategy based on the vehicle's driving scenario data.

[0248] Taking complex scenarios such as urban roads with low vehicle speeds, dense traffic flow, numerous vulnerable road users, intersections, and traffic lights as an example, the core risks are injuries from low-speed reversing, fatal accidents from parked car doors opening, collisions from non-motorized vehicles weaving through traffic, and collisions from forced lane changes at intersections. Therefore, adjusting the early warning strategy can focus on suppressing safety protection and false alarms within the strategy. Specific methods for adjusting the early warning strategy include:

[0249] (1) When the vehicle speed is below 30km / h, the causal effect weight of the characteristics of vulnerable traffic participants can be increased, and the risk warning for non-motorized vehicles and pedestrians can be strengthened.

[0250] (2) When it is detected that a person inside the vehicle is grasping the door handle or the door is unlocked, a door opening risk warning can be triggered in advance, and In such cases, the car doors can be temporarily locked to prevent accidents caused by opening the doors.

[0251] (3) When the vehicle is less than 150m from the stop line at the intersection and less than 150m from the traffic lights, unnecessary lane change and risk avoidance interventions can be suppressed, and only the sound and light warnings can be retained to avoid misoperation in intersection scenarios.

[0252] (4) When a target behind the vehicle is detected to have a tendency to leave the current lane, that is, when the following formula (13) is satisfied, the risk weight of the target behind the vehicle can be reduced to avoid false warnings:

[0253] (13)

[0254] in, The heading angle of the target behind, in rad; This indicates the angular velocity of the target behind, expressed in rad / s.

[0255] Taking complex scenarios such as high-speed traffic, high risk of chain collisions, and merging traffic from ramps and exits as examples, the core risks are high-speed rear-end collisions, multi-vehicle chain collisions, and collisions caused by forced lane changes. Therefore, adjusting the warning strategy can focus on identifying and actively avoiding beyond-visual-range risks within the strategy. Specific methods for adjusting the warning strategy include:

[0256] (1) It can increase the weight of chain collision risk and strengthen the beyond-line-of-sight risk identification of the graph pyramid network model. When a multi-vehicle chain collision risk is detected, the warning level can be increased 2 seconds in advance.

[0257] (2) When the vehicle is less than 2000m away from the highway ramp or service area entrance / exit, it can suppress lane change and risk avoidance intervention in lanes not recommended by navigation, so as to avoid missing the ramp;

[0258] (3) When there is traffic merging into the adjacent motor vehicle lane within 800m ahead of the vehicle, it shall not be used as the target lane for evacuation, so as to avoid secondary risks caused by the merging traffic.

[0259] (4) When a vehicle is in a high-risk scenario, the time to collision (TTC) of the vehicles in front and behind in the target lane can be verified. When the vehicle collides with the vehicle in front in the target lane, the time to collision is... Time of collision between the vehicle and the vehicle behind in the target lane Furthermore, when the longitudinal safety distance between the vehicle and the vehicle in the parallel lane after entering the target lane is greater than 8m, lane change avoidance intervention can be triggered. The collision time TTC can be obtained by the following formula (14):

[0260] (14)

[0261] In the formula, , These represent the speeds of the vehicle in front and the vehicle behind in the target lane, respectively, in m / s; To represent a very small constant, in order to avoid the denominator being 0; Indicates the current longitudinal speed of the vehicle ahead in the target lane (in m / s); This indicates the vehicle's current longitudinal speed (in m / s). This indicates the current longitudinal speed (in m / s) of the vehicle following in the target lane.

[0262] Taking low-speed parking or stationary scenarios as an example, the core risks are reversing collisions, being rear-ended while stationary, and door-opening collisions. Therefore, adjusting the warning strategy can focus on protecting vulnerable road users and providing a safety net. Specific methods for adjusting the warning strategy include:

[0263] (1) When the vehicle is in reverse gear, the feature extraction of pedestrians, non-motorized vehicles, and low obstacles behind can be enhanced. When this happens, emergency braking can be triggered;

[0264] (2) When the vehicle is parked and stationary, lowering the warning threshold can trigger the hazard lights and seat belt pretensioning in advance, and automatically apply the parking brake to prevent the vehicle from rolling away after being rear-ended.

[0265] Meanwhile, the onboard chip can dynamically adjust the rear collision warning threshold and the driver's intervention level (i.e., the degree of intervention) in the warning strategy based on the vehicle's driving style data, achieving personalized and adaptive strategy. This significantly reduces false alarms while ensuring safety, improving user acceptance and human-like performance. In this embodiment, the rear collision warning threshold adjustment parameters for different driving styles are shown in Table 5.

[0266] Table 5

[0267] Driving style type Rear collision warning threshold offset under low risk level Rear collision warning threshold shift under medium risk level Rear collision warning threshold offset under high risk level Intervention intensity Early warning coefficient radical -5% -5% -5% 1.2 1.3 Standard 0 0 0 1.0 1.0 conservative +5% +5% +5% 0.8 0.8

[0268] Among them, a negative rear collision warning threshold offset indicates a reduction in the rear collision warning threshold triggered by the warning, thus advancing the warning timing; an intervention intensity greater than 1 indicates increased intervention intensity, while an intervention intensity less than 1 indicates decreased intervention intensity.

[0269] Step S423: Provide a rear collision warning for the vehicle according to the adjusted warning strategy.

[0270] Furthermore, the corresponding rear collision warning information can be directly output according to the adjusted warning strategy to provide rear collision warning for the vehicle.

[0271] In one embodiment, after performing the process in step S400 above, as Figure 13 As shown, the above method also includes:

[0272] Step S500: Based on the vehicle's rear collision risk level and multi-source perception data, obtain the closed-loop control strategy for the vehicle's environment.

[0273] It should be noted that before issuing a rear collision warning, the vehicle can perform an efficient lane change to ensure its safety. Specifically, to avoid a vehicle slowing down ahead, the vehicle can switch to an adjacent lane. The process for determining the efficiency of a lane change can be as follows: First, select an adjacent lane (the lane the vehicle can enter). If it is determined that the adjacent lane does not meet the vehicle's traffic efficiency requirements, then no lane change request will be initiated. Otherwise, determine whether the vehicle's current environment meets the constraints of the intersection and / or traffic lights. If not, no lane change request will be initiated. Otherwise, determine whether the vehicle meets the protection constraints for door opening / vulnerable party participants. If not, no lane change request will be initiated. Otherwise, a lane change request can be initiated.

[0274] Step S600: Based on the closed-loop control strategy of the vehicle's environment, perform safety control on the vehicle.

[0275] This embodiment can perform safety control on the vehicle based on a closed-loop control strategy in the vehicle's environment to ensure vehicle safety. At the same time, it can design differentiated control strategies for different scenarios and drivers' driving styles to improve control accuracy.

[0276] The technical solution in this application embodiment determines the vehicle's driving scenario data based on multi-source perception data, and determines the vehicle's driving style data based on the vehicle's historical driving information. Based on the vehicle's driving scenario data and driving style data, the warning strategy is adjusted to obtain the adjusted warning strategy. The vehicle is then given a rear collision warning according to the adjusted warning strategy. The above method can design differentiated warning strategies for different scenarios and drivers' driving styles, which can improve the accuracy of the warning and avoid false warnings.

[0277] In one embodiment, this application also provides a vehicle rear collision warning method, which includes the following processes:

[0278] (1) Input the multi-source sensing data into the spatial feature pyramid branch for feature extraction to obtain multi-scale spatial fusion features;

[0279] (2) Input the multi-source sensing data into the temporal feature pyramid branch for feature extraction to obtain full temporal behavior fusion features;

[0280] (3) The multi-scale spatial fusion features and the full-time behavior fusion features are fused to obtain the multi-scale risk characterization features of the vehicle;

[0281] (4) Preprocess the multi-source sensing data to obtain preprocessed data; the preprocessing includes invalid removal, spatiotemporal alignment, trajectory tracking and smoothing filtering;

[0282] (5) Extract features from the preprocessed data to obtain multiple rear collision risk characterization features of the vehicle;

[0283] (6) Perform multi-scale pyramid branch mapping based on the characteristics of each post-collision risk to obtain the embedding features of each scale pyramid branch;

[0284] (7) Perform multi-scale feature fusion on each embedded feature to obtain the multi-scale feature matrix of the vehicle;

[0285] (8) Input the multi-scale feature matrix into the behavioral intention prediction model and output the posterior probability of the target's upcoming behavioral intentions;

[0286] (9) The behavioral intention corresponding to the largest posterior probability among all posterior probabilities is determined as the behavioral intention of the target behind.

[0287] (10) Construct a global traffic flow topology map of vehicles based on multi-scale risk characterization features and multi-source perception data;

[0288] (11) Input the global traffic flow topology map into the graph pyramid network model and output the multi-scale fusion features of vehicles;

[0289] (12) Based on the multi-scale fusion characteristics, the normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and the lane matching degree between the vehicle and the upstream target, determine the risk transmission coefficient from the upstream target to the vehicle;

[0290] (13) Determine the indirect chain collision causal probability of the upstream target based on the risk transmission coefficient and the collision causal probability of the upstream target to its preceding vehicle;

[0291] (14) Input the indirect chain collision causal probability, the causal effect weight of the rear target's behavioral intention on the collision risk and the prior probability into the causal inference model to perform causal association inference and determine the direct collision causal probability of the rear target on the vehicle.

[0292] (15) The maximum value between the indirect chain collision causal probability and the direct collision causal probability is determined as the post-collision risk quantification value;

[0293] (16) Determine the rear collision risk level of the vehicle based on the rear collision risk quantification value;

[0294] (17) Obtain the corresponding early warning strategy based on the rear collision risk level, and determine the vehicle's driving scenario data based on multi-source perception data, and determine the vehicle's driving style data based on the vehicle's historical driving information.

[0295] (18) Adjust the warning strategy based on the vehicle's driving scenario data and driving style data, and obtain the adjusted warning strategy;

[0296] (19) Provide rear collision warning for the vehicle in accordance with the adjusted warning strategy;

[0297] (20) Based on the vehicle's rear collision risk level and multi-source perception data, obtain the closed-loop control strategy in the vehicle's environment.

[0298] (21) Based on the closed-loop control strategy of the vehicle's environment, the vehicle is controlled for safety.

[0299] The specific execution process of (1) to (21) above can be found in the description of the above embodiments. The implementation principle and technical effect are similar, and will not be repeated here.

[0300] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0301] Based on the same inventive concept, this application also provides a vehicle rear collision warning device for implementing the aforementioned vehicle rear collision warning method. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations of one or more vehicle rear collision warning device embodiments provided below can be found in the limitations of the vehicle rear collision warning method described above, and will not be repeated here.

[0302] In one embodiment, Figure 14 This is a schematic diagram of the structure of a vehicle rear collision warning device in one embodiment of this application. The vehicle rear collision warning device provided in this embodiment can be applied to an on-board chip in a vehicle. Figure 14 As shown, the vehicle rear collision warning device of this application embodiment may include: a feature processing module 11, a classification module 12, a risk identification module 13, and a warning module 14, wherein:

[0303] Feature processing module 11 is used to perform multi-scale feature processing on the multi-source perception data of the vehicle using a spatiotemporal feature fusion model to obtain the multi-scale risk characterization features of the vehicle.

[0304] Classification module 12 is used to classify the behavioral intentions of the target behind the vehicle based on the behavioral intention prediction model and multi-source perception data, and to determine the behavioral intentions of the target behind the vehicle.

[0305] The risk identification module 13 is used to identify the collision risk of the vehicle based on the behavioral intent of the target behind, multi-scale risk characterization features and multi-source perception data, and obtain the quantitative value of the vehicle's rear collision risk.

[0306] The warning module 14 is used to provide rear collision warnings to vehicles based on the rear collision risk quantification value.

[0307] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0308] In one embodiment, the spatiotemporal feature fusion model includes a spatial feature pyramid branch and a temporal feature pyramid branch; the feature processing module 11 is specifically used for:

[0309] Multi-source sensing data is input into the spatial feature pyramid branch for feature extraction to obtain multi-scale spatial fusion features;

[0310] Multi-source sensing data is input into the temporal feature pyramid branch for feature extraction to obtain full temporal behavior fusion features;

[0311] Multi-scale spatial fusion features and full-time behavioral fusion features are fused to obtain multi-scale risk characterization features of vehicles.

[0312] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0313] In one embodiment, the classification module 12 includes: a feature extraction unit, a feature fusion unit, and a causal reasoning unit, wherein:

[0314] The feature extraction unit is used to extract multi-dimensional features based on multi-source perception data to obtain multiple rear collision risk characterization features of the vehicle.

[0315] The feature fusion unit is used to perform multi-scale feature fusion based on the characteristics of each rear collision risk to determine the multi-scale feature matrix of the vehicle.

[0316] The causal reasoning unit is used to perform causal reasoning on the behavioral intentions of the target based on the behavioral intention prediction model and the multi-scale feature matrix, and to determine the behavioral intentions of the target.

[0317] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0318] In one embodiment, the feature extraction unit is specifically used for:

[0319] The multi-source sensing data is preprocessed to obtain preprocessed data; the preprocessing includes invalid removal, spatiotemporal alignment, trajectory tracking and smoothing filtering.

[0320] Feature extraction is performed on the preprocessed data to obtain multiple rear collision risk characterization features of the vehicle.

[0321] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0322] In one embodiment, the feature fusion unit is specifically used for:

[0323] Multi-scale pyramid branch mapping is performed based on the characteristics of each post-collision risk to obtain the embedding features of each scale pyramid branch.

[0324] Multi-scale feature fusion is performed on each embedded feature to obtain the multi-scale feature matrix of the vehicle.

[0325] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0326] In one embodiment, the causal reasoning unit is specifically used for:

[0327] The multi-scale feature matrix is ​​input into the behavioral intent prediction model, and the posterior probability of the target's upcoming behavioral intent is output.

[0328] The behavioral intent corresponding to the largest posterior probability among all posterior probabilities is determined as the behavioral intent of the target.

[0329] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0330] In one embodiment, the risk identification module 13 includes: a topology graph construction unit, a first processing unit, a transmission coefficient determination unit, and a second processing unit, wherein:

[0331] The topology graph construction unit is used to construct a global traffic flow topology graph for vehicles based on multi-scale risk characterization features and multi-source perception data.

[0332] The first processing unit is used to input the global traffic flow topology map into the graph pyramid network model and output the multi-scale fusion features of vehicles.

[0333] The transmission coefficient determination unit is used to determine the risk transmission coefficient from the upstream target to the vehicle based on multi-scale fusion features, normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and lane matching degree between the vehicle and the upstream target.

[0334] The second processing unit is used to quantify the rear collision risk of a vehicle based on a causal reasoning model and a risk transmission coefficient, and obtain a quantified value of the rear collision risk.

[0335] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0336] In one embodiment, the second processing unit is specifically used for:

[0337] Based on the risk transmission coefficient and the causal probability of collision between the upstream target and the vehicle in front, the indirect chain collision causal probability of the upstream target is determined.

[0338] The causal probability of indirect chain collisions, the causal effect weight of the behavioral intention of the rear target on the collision risk, and the prior probability are input into the causal inference model to perform causal association inference and determine the direct collision causal probability of the rear target on the vehicle.

[0339] The maximum value between the causal probability of indirect chain collisions and the causal probability of direct collisions is determined as the quantified value of post-collision risk.

[0340] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0341] In one embodiment, the early warning module 14 includes: a level determination unit and an early warning unit, wherein:

[0342] The rating determination unit is used to determine the rear collision risk level of a vehicle based on the rear collision risk quantification value.

[0343] The warning unit is used to obtain the corresponding warning strategy based on the rear collision risk level, and to issue a rear collision warning to the vehicle according to the warning strategy.

[0344] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0345] In one embodiment, the early warning unit is specifically used for:

[0346] The vehicle's driving scenario data is determined based on multi-source perception data, and the vehicle's driving style data is determined based on the vehicle's historical driving information.

[0347] Based on vehicle driving scenario data and driving style data, the warning strategy is adjusted, and the adjusted warning strategy is obtained.

[0348] According to the adjusted warning strategy, a rear collision warning will be issued for the vehicle.

[0349] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0350] In one embodiment, the vehicle rear collision warning device further includes: a safety control module, wherein:

[0351] The safety control module is used to obtain the closed-loop control strategy of the vehicle's environment based on the vehicle's rear collision risk level and multi-source perception data, and to perform safety control on the vehicle based on the closed-loop control strategy of the vehicle's environment.

[0352] The vehicle rear collision warning device provided in this application embodiment can be used to execute the technical solution in the above-described vehicle rear collision warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0353] Specific limitations regarding the vehicle rear collision warning device can be found in the limitations of the vehicle rear collision warning method described above, and will not be repeated here. Each module in the aforementioned vehicle rear collision warning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the vehicle's onboard chip in hardware form or independently of it, or stored in the vehicle's memory in software form, so that the onboard chip can call and execute the corresponding operations of each module.

[0354] In one embodiment, a vehicle is provided, the internal structure diagram of which can be as follows: Figure 15As shown, the vehicle includes an onboard chip, memory, and network interface connected via a system bus. The onboard chip provides processing power. The vehicle's memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The vehicle's database stores multi-source perception data. The vehicle's network interface is used to communicate with external endpoints via a network connection. When the computer program is executed by the onboard chip, it implements a rear-collision warning method.

[0355] Those skilled in the art will understand that Figure 15 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the vehicle to which the present application is applied. A specific vehicle may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0356] In one embodiment, a cloud is also provided, including a memory and an onboard chip, the memory storing a computer program, the onboard chip executing the computer program to perform the following steps:

[0357] A spatiotemporal feature fusion model is used to perform multi-scale feature processing on the multi-source perception data of vehicles to obtain multi-scale risk characterization features of vehicles.

[0358] Based on the behavioral intent prediction model and multi-source perception data, the behavioral intent of the target behind the vehicle is classified and the behavioral intent of the target behind the vehicle is determined.

[0359] Based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data, the collision risk of the vehicle is identified, and the quantitative value of the vehicle's rear collision risk is obtained.

[0360] Based on the quantified value of rear collision risk, a rear collision warning is issued for the vehicle.

[0361] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by an on-board chip, performs the following steps:

[0362] A spatiotemporal feature fusion model is used to perform multi-scale feature processing on the multi-source perception data of vehicles to obtain multi-scale risk characterization features of vehicles.

[0363] Based on the behavioral intent prediction model and multi-source perception data, the behavioral intent of the target behind the vehicle is classified and the behavioral intent of the target behind the vehicle is determined.

[0364] Based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data, the collision risk of the vehicle is identified, and the quantitative value of the vehicle's rear collision risk is obtained.

[0365] Based on the quantified value of rear collision risk, a rear collision warning is issued for the vehicle.

[0366] In one embodiment, a computer program product is provided, including a computer program that, when executed by an on-board chip, performs the following steps:

[0367] A spatiotemporal feature fusion model is used to perform multi-scale feature processing on the multi-source perception data of vehicles to obtain multi-scale risk characterization features of vehicles.

[0368] Based on the behavioral intent prediction model and multi-source perception data, the behavioral intent of the target behind the vehicle is classified and the behavioral intent of the target behind the vehicle is determined.

[0369] Based on the behavioral intent of the target behind, multi-scale risk characterization features, and multi-source perception data, the collision risk of the vehicle is identified, and the quantitative value of the vehicle's rear collision risk is obtained.

[0370] Based on the quantified value of rear collision risk, a rear collision warning is issued for the vehicle.

[0371] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0372] The technical features in the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0373] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for rear collision warning of a vehicle, characterized in that, The method includes: A spatiotemporal feature fusion model is used to perform multi-scale feature processing on the multi-source perception data of the vehicle to obtain the multi-scale risk characterization features of the vehicle. Based on the behavioral intent prediction model and the multi-source perception data, the behavioral intent of the target behind the vehicle is classified and the behavioral intent of the target behind the vehicle is determined. Based on the behavioral intent of the rear target, the multi-scale risk characterization features, and the multi-source perception data, the collision risk of the vehicle is identified, and the rear collision risk quantification value of the vehicle is obtained. Based on the aforementioned rear collision risk quantification value, a rear collision warning is issued for the vehicle.

2. The method according to claim 1, characterized in that, The spatiotemporal feature fusion model includes a spatial feature pyramid branch and a temporal feature pyramid branch; The method employs a spatiotemporal feature fusion model to perform multi-scale feature processing on the vehicle's multi-source perception data, obtaining the vehicle's multi-scale risk characterization features, including: The multi-source sensing data is input into the spatial feature pyramid branch for feature extraction to obtain multi-scale spatial fusion features. The multi-source sensing data is input into the temporal feature pyramid branch for feature extraction to obtain full temporal behavior fusion features; The multi-scale spatial fusion features and the full-time behavioral fusion features are fused to obtain the multi-scale risk characterization features of the vehicle.

3. The method according to claim 1, characterized in that, The process of classifying and determining the behavioral intentions of the targets behind the vehicle based on the behavioral intention prediction model and the multi-source perception data includes: Based on the multi-source perception data, multi-dimensional feature extraction is performed to obtain multiple rear collision risk characterization features of the vehicle; Based on the rear collision risk characterization features described above, multi-scale feature fusion is performed to determine the multi-scale feature matrix of the vehicle. Based on the behavioral intent prediction model and the multi-scale feature matrix, causal reasoning is performed on the behavioral intent of the rear target to determine the behavioral intent of the rear target.

4. The method according to claim 3, characterized in that, The step of extracting multi-dimensional features from the multi-source perception data to determine multiple rear-collision risk characterization features of the vehicle includes: The multi-source sensing data is preprocessed to obtain the preprocessed data; the preprocessing includes invalid removal, spatiotemporal alignment, trajectory tracking, and smoothing filtering. Feature extraction is performed on the preprocessed data to obtain multiple rear collision risk characterization features of the vehicle.

5. The method according to claim 3, characterized in that, The step of performing multi-scale feature fusion based on the respective rear-collision risk characterization features to determine the multi-scale feature matrix of the vehicle includes: Based on the post-collision risk characterization features described above, a multi-scale pyramid branch mapping is performed to obtain the embedding features of each scale pyramid branch. Multi-scale feature fusion is performed on each of the embedded features to obtain the multi-scale feature matrix of the vehicle.

6. The method according to claim 3, characterized in that, The step of determining the behavioral intention of the rear target by performing causal inference based on the behavioral intention prediction model and the multi-scale feature matrix includes: The multi-scale feature matrix is ​​input into the behavioral intent prediction model, and the posterior probability of the target's upcoming behavioral intent is output. The behavioral intent corresponding to the largest posterior probability among all the posterior probabilities is determined as the behavioral intent of the rear target.

7. The method according to any one of claims 1-6, characterized in that, The step of identifying the collision risk of the vehicle based on the behavioral intent of the rear target, the multi-scale risk characterization features, and the multi-source perception data, and obtaining a quantitative value of the rear collision risk of the vehicle, includes: Based on the multi-scale risk characterization features and the multi-source perception data, a global traffic flow topology map of the vehicle is constructed. The global traffic flow topology is input into the graph pyramid network model, which outputs the multi-scale fusion features of the vehicles. Based on the multi-scale fusion features, the normalized relative distance and normalized relative speed between the vehicle and the corresponding upstream target, and the lane matching degree between the vehicle and the upstream target, the risk transmission coefficient of the risk from the upstream target to the vehicle is determined; Based on the causal reasoning model and the risk transmission coefficient, the rear collision risk of the vehicle is quantified to obtain the quantified value of the rear collision risk.

8. The method according to claim 7, characterized in that, The process of quantifying the rear-collision risk of the vehicle based on the causal inference model and the risk transmission coefficient to obtain the quantified value of the rear-collision risk includes: Based on the risk transmission coefficient and the collision causal probability of the upstream target to its preceding vehicle, the indirect chain collision causal probability of the upstream target is determined. The indirect chain collision causal probability, the causal effect weight of the rear target's behavioral intention on the collision risk, and the prior probability are input into the causal inference model to perform causal association inference and determine the direct collision causal probability of the rear target on the vehicle. The maximum value between the indirect cascading collision causal probability and the direct collision causal probability is determined as the post-collision risk quantification value.

9. The method according to any one of claims 1-6, characterized in that, The method of providing a rear collision warning to the vehicle based on the rear collision risk quantification value includes: The rear collision risk level of the vehicle is determined based on the rear collision risk quantification value. Based on the rear collision risk level, a corresponding warning strategy is obtained, and a rear collision warning is issued to the vehicle according to the warning strategy.

10. The method according to claim 9, characterized in that, The step of providing a rear collision warning to the vehicle according to the warning strategy includes: The driving scenario data of the vehicle is determined based on the multi-source perception data, and the driving style data of the vehicle is determined based on the historical driving information of the vehicle. Based on the vehicle's driving scenario data and the driving style data, the warning strategy is adjusted, and the adjusted warning strategy is obtained. According to the adjusted warning strategy, a rear collision warning is issued to the vehicle.

11. The method according to any one of claims 1-6, characterized in that, The method further includes: Based on the rear collision risk level of the vehicle and the multi-source perception data, a closed-loop control strategy for the environment in which the vehicle is located is obtained. Based on the closed-loop control strategy of the vehicle's environment, the vehicle is subjected to safety control.

12. A vehicle rear collision warning device, characterized in that, The device includes: The feature processing module is used to perform multi-scale feature processing on the multi-source perception data of the vehicle using a spatiotemporal feature fusion model to obtain the multi-scale risk characterization features of the vehicle. The classification module is used to classify the behavioral intentions of the target behind the vehicle based on the behavioral intention prediction model and the multi-source perception data, and to determine the behavioral intentions of the target behind the vehicle. The risk identification module is used to identify the collision risk of the vehicle based on the behavioral intent of the rear target, the multi-scale risk characterization features, and the multi-source perception data, and obtain the rear collision risk quantification value of the vehicle. The warning module is used to provide a rear collision warning to the vehicle based on the rear collision risk quantification value.

13. A vehicle, comprising a memory and an on-board chip, wherein the memory stores a computer program, characterized in that, When the vehicle-mounted chip executes the computer program, it implements the steps of the method according to any one of claims 1-11.

Images