Automatic driving braking test method and device based on bionic vision and touch fusion

Abstract

The invention relates to an automatic driving braking test method and device based on bionic vision and tactile fusion, and the method comprises the steps: obtaining environment information in a current test scene based on visual equipment; simulating a collision process of a current vehicle and an obstacle, and acquiring tactile information corresponding to the collision process based on force feedback equipment; based on the original visual features corresponding to the environment information and the original tactile features corresponding to the tactile information, performing multi-modal feature fusion processing to obtain visual and tactile combined features; predicting vehicle body response information of the current vehicle corresponding to the candidate braking strategy based on the visual and tactile joint features, and obtaining an optimal braking strategy corresponding to the candidate braking strategy based on the vehicle body response information; and generating a braking control signal based on the optimal braking strategy. According to the invention, comprehensive and accurate evaluation of the braking performance of the automatic driving system under the emergency condition is realized.

Description

Technical Field

[0001] The present application relates to the technical field of autonomous driving, and particularly to an autonomous driving braking test method and device based on the fusion of bionic vision and touch. Background Art

[0002] With the rapid development of autonomous driving technology, the reliability and safety of vehicle emergency braking systems have become the focus of the industry. Traditional autonomous driving test methods mainly rely on a single visual perception system, obtaining environmental information through sensors such as cameras and lidar, and generating braking strategies based on predefined rules or simple models.

[0003] However, the existing technologies have revealed the following significant defects in practical applications: First, the perception accuracy of a single visual system drops sharply in complex environments such as rain, fog, backlight, and night, reducing the reliability of perception; Second, traditional test methods mostly use pure numerical simulation or finite element analysis, and cannot truly restore the mechanical feedback (such as impact force distribution, vibration frequency characteristics) when the vehicle collides with an obstacle; Third, traditional test methods rely on manual annotation of scenarios, with a low coverage rate of scenarios, and it is difficult to generate diverse dangerous scenarios to comprehensively evaluate the performance of autonomous driving systems; Finally, the existing braking strategies have insufficient robustness under uncertain conditions such as sensor noise and model parameter drift, and traditional test methods lack effective handling of uncertainties.

[0004] Therefore, there is an urgent need for an emergency braking test method that integrates multi-modal perception, physical model-driven, and intelligent optimization to solve core problems such as insufficient perception robustness in complex environments, limited coverage of dangerous scenarios, and low reliability of control strategies, promote a substantial breakthrough in autonomous driving safety technology, and achieve a comprehensive and accurate evaluation of the braking performance of autonomous driving systems in emergency situations. Summary of the Invention

[0005] To solve the core problems such as insufficient perception robustness in complex environments, limited coverage of dangerous scenarios, and low reliability of control strategies, and to achieve a comprehensive and accurate evaluation of the braking performance of autonomous driving systems in emergency situations, the present application provides an autonomous driving braking test method and device based on the fusion of bionic vision and touch.

[0006] In a first aspect, the present application provides an autonomous driving braking test method based on the fusion of bionic vision and touch, including:

[0007] S1. Based on a vision device, obtain the environmental information in the current test scenario. The environmental information includes weather, light, road surface information, obstacle information, and vehicle information of the host vehicle. The road surface information includes road surface friction coefficient, road surface flatness, road surface material, road surface slope, road surface curvature, road boundary, and lane line information. The obstacle information includes obstacle type and obstacle attribute information. The vehicle information of the host vehicle includes the current vehicle speed, vehicle body attitude, and vehicle body state;

[0008] S2. Simulate the collision process between the current vehicle and the obstacle, and based on a force feedback device, obtain the tactile information corresponding to the collision process;

[0009] S3. Based on the original visual features corresponding to the environmental information and the original tactile features corresponding to the tactile information, perform multi-modal feature fusion processing to obtain visual-tactile joint features;

[0010] S4. Based on the visual-tactile joint features, predict the vehicle body response information of the current vehicle corresponding to the candidate braking strategy, and based on the vehicle body response information, obtain the optimal braking strategy corresponding to the candidate braking strategy. The braking strategy is a parameter characterizing the braking control of the current vehicle in an emergency braking scenario. The vehicle body response information includes the longitudinal dynamic characteristics, lateral dynamic characteristics, vertical dynamic characteristics, and stability index of the vehicle;

[0011] S5. Generate a braking control signal based on the optimal braking strategy.

[0012] The beneficial effects of this application are as follows: By integrating the vision device and the force feedback device, comprehensive acquisition of environmental information and tactile information is realized, and the fusion perception of visual information and tactile information is achieved. By integrating visual-tactile multi-modal features, visual-tactile joint features are obtained, improving the perception accuracy and robustness. Based on the visual-tactile joint features, vehicle behavior can be accurately predicted, enhancing the test reliability.

[0013] Further, the S5 includes:

[0014] Based on the vehicle body dynamics model and the vehicle body response information corresponding to the optimal braking strategy, obtain the longitudinal control signal and the lateral control signal for implementing the optimal braking strategy. The vehicle body response information includes the first uncertainty quantification data of the control signal, and the first uncertainty quantification data is a parameter for evaluating the confidence level of the reliability of the control signal;

[0015] Based on the first uncertainty quantification data, perform uncertainty weighted fusion on the longitudinal control signal and the lateral control signal to obtain the fused longitudinal control signal and lateral control signal, and based on the collision probability between the current vehicle and the obstacle, calculate the comprehensive collision risk value corresponding to the optimal braking strategy;

[0016] If the comprehensive collision risk value exceeds a preset risk threshold, a braking control signal is generated based on the fused longitudinal control signal and lateral control signal.

[0017] The beneficial effects of adopting the above further solution are as follows: By considering both longitudinal and lateral movements simultaneously, more comprehensive braking control is achieved, improving the comprehensiveness of the braking strategy. Through the robustness management module, uncertainties can be handled, improving the reliability of the strategy in complex environments and enhancing the robustness of the braking strategy. By separately planning longitudinal and lateral movements, more refined braking control can be achieved, improving the braking effect and optimizing the braking effect. The strategy design based on collision risk assessment can better handle potential collision risks and enhance safety.

[0018] Further, calculating the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle includes:

[0019] Based on a preset comprehensive evaluation algorithm, comprehensively evaluate the collision probability and time-to-collision data to obtain the comprehensive collision risk value;

[0020] The comprehensive evaluation algorithm is expressed as:

[0021]

[0022] Among them, α1 represents a weight coefficient, and the weight coefficient is used to determine the priority of the time-to-collision data and the collision probability; β1 represents an attenuation coefficient, and the attenuation coefficient is used to control the attenuation rate of the time-to-collision data; P collision represents the collision probability; TTC represents the time-to-collision data; Risk represents the comprehensive collision risk value.

[0023] The beneficial effects of adopting the above further solution are as follows: The comprehensive collision risk assessment algorithm realizes multi-dimensional risk assessment of the comprehensive collision risk value by quantifying the collision probability and time urgency, avoiding over-braking or delayed response.

[0024] Further, before calculating the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle, it further includes:

[0025] Based on the visual-tactile joint features and a preset adaptive dynamic scene generation algorithm, randomly search for the candidate motion trajectories of the obstacle in different scenarios to obtain a set of candidate trajectory branches;

[0026] Based on a preset reinforcement learning algorithm, obtain the optimized trajectory parameters corresponding to the set of candidate trajectory branches, and the trajectory parameters include control point coordinates and speed curves;

[0027] Based on a preset Bessel curve formula, obtain a first target trajectory of the obstacle corresponding to the trajectory parameter;

[0028] Predict a second target trajectory of the current vehicle, and based on the first target trajectory and the second target trajectory, obtain the collision probability.

[0029] The beneficial effects of adopting the above further solution are as follows: In autonomous driving tests, by randomizing and optimizing control points, it is possible to more accurately perceive the surrounding situation in complex environments, obtain the collision probability, and efficiently generate diverse test dangerous scenarios, improving the test coverage and authenticity, thereby achieving a comprehensive and accurate assessment of the emergency braking performance of autonomous driving.

[0030] Further, the S2 includes:

[0031] Based on the road surface information and collision information, simulate the collision process, where the collision information includes collision type, relative collision speed, collision angle, collision contact area, obstacle material hardness characteristics, and collision force distribution and transmission path;

[0032] Based on the force feedback device, obtain the impact tactile signal during the collision process, and obtain the basic tactile signal during the collision process corresponding to the road surface information, the motion state tactile feedback signal during the collision process corresponding to the vehicle information of the self - vehicle, the vibration tactile signal during the collision process, and the pressure tactile signal during the collision process. The vehicle information of the self - vehicle further includes vehicle body attitude, suspension state, tire state, power system state, and braking system state;

[0033] Fuse the impact tactile signal with the basic tactile signal, the motion state tactile feedback signal, the vibration tactile signal, and the pressure tactile signal to obtain a comprehensive tactile signal stream;

[0034] Based on a preset mechanical and neural signal conversion model, obtain the tactile information corresponding to the comprehensive tactile signal stream;

[0035] The mechanical and neural signal conversion model is expressed as:

[0036]

[0037] where S(t1) represents the tactile information corresponding to the neural signal, F(t) represents the comprehensive tactile signal stream corresponding to the physical force signal, α2 represents the weight coefficient simulating the human sensitivity to the magnitude of the force, β2 represents the weight coefficient simulating the human sensitivity to the change rate of the force, and γ1 represents the weight coefficient simulating the human sensitivity to the duration of the force.

[0038] The beneficial effects of adopting the above further solution are as follows: After fusing multiple tactile signals, a comprehensive tactile signal stream is obtained. By fusing tactile signals from multiple aspects, more comprehensive tactile feedback can be provided. Through the mechanical and neural signal conversion model, physical forces, pressures, vibrations, etc. can be converted into signals that can be perceived by the human nervous system, simulating the response characteristics of human mechanoreceptors (such as Pacinian corpuscles, Meissner corpuscles, etc.).

[0039] Further, S3 includes:

[0040] Based on a preset first deep feature extraction network, extract the original visual features, where the original visual features include image semantic information, regions of interest, and the first spatio-temporal information distribution features;

[0041] Based on a preset second deep feature extraction network, extract the original tactile features, where the original tactile features include the second spatio-temporal information distribution features;

[0042] Based on the multi-head self-attention structure, align the original visual features and the original tactile features, and based on a preset bilinear pooling algorithm, perform an outer product operation on the aligned original visual features and original tactile features to obtain the fused visual-tactile joint features.

[0043] The beneficial effects of adopting the above further solution are as follows: Combining tactile information with information related to vehicle motion can establish a mapping relationship between road conditions, collision events, and vehicle dynamic responses, improving the ability to understand and predict complex scenarios. Solving spatio-temporal misalignment through attention alignment and realizing cross-modal interaction and fusion through bilinear pooling, and finally generating multi-dimensional visual-tactile joint features.

[0044] Further, after aligning the original visual features and the original tactile features based on the multi-head self-attention structure, it further includes:

[0045] Based on a preset contrast loss function, perform similarity learning on the aligned original visual features and original tactile features;

[0046] The contrast loss function is expressed as:

[0047]

[0048] where Y represents a binary label. When the aligned original visual features and original tactile features come from the same test scenario, Y is 0, otherwise Y is 1; D wDenote the feature distance, which is a parameter characterizing the difference between the aligned original visual feature and original tactile feature; m represents a set boundary value, and the feature distances of the original visual feature and original tactile feature from different test scenarios are at least greater than the boundary value.

[0049] The beneficial effects of adopting the above further scheme are as follows: After aligning the original visual feature and original tactile feature, the semantic consistency of cross-modal features can be further optimized through the contrast loss function. The original visual feature and original tactile feature in the same scenario should have a high degree of similarity, enhancing the intra-modal consistency; the cross-modal feature distances of different scenarios need to exceed the preset boundary value to avoid confusion and strengthen the inter-modal distinguishability.

[0050] Further, S4 includes:

[0051] S41, based on a preset magic formula and the visual-tactile joint feature, calculate the tire forces corresponding to each wheel when executing the candidate braking strategy, where the tire forces include longitudinal forces and lateral forces;

[0052] S42, based on each of the tire forces and vehicle external forces, calculate the overall force information of the current vehicle, where the vehicle external forces include air resistance and gravity;

[0053] S43, based on the overall force information, calculate the vehicle acceleration, where the vehicle acceleration includes longitudinal acceleration, lateral acceleration, vertical acceleration, yaw acceleration, pitch acceleration, and roll acceleration;

[0054] S44, based on the vehicle acceleration, the current vehicle speed, the current position of the current vehicle, and a preset numerical integration algorithm, obtain the predicted speed and predicted position of the current vehicle at the next moment; and based on the predicted speed and predicted position, obtain the body response information corresponding to the candidate braking strategy;

[0055] S45, based on the body response information, perform a non-linear time-varying analysis to obtain the non-linear time-varying characteristics existing between the tires of the current vehicle and the ground, where the non-linear time-varying characteristics include non-linear friction coefficient, non-linear side slip characteristics, and time-varying road adhesion;

[0056] S46, if the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, adjust the candidate braking strategy; if the non-linear time-varying characteristics do not exceed the corresponding current safety constraint conditions, take the candidate braking strategy as the optimal braking strategy;

[0057] S47, take the adjusted candidate braking strategy as the new candidate braking strategy, and repeat the steps of S41 to S46 until the optimal braking strategy is obtained.

[0058] The beneficial effects of adopting the above further solution are as follows: identifying the complex characteristics of tire-ground interaction and predicting risks in real time through a state-space model. Through the cyclic iteration of S41 - S47, the optimal strategy is gradually approximated, improving the braking effect.

[0059] Furthermore, before determining whether the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, it further includes:

[0060] Based on the second uncertainty quantification data, the original safety constraint conditions are corrected to obtain the corrected robust constraint conditions, and the robust constraint conditions are used as the current safety constraint conditions. The second uncertainty quantification data is a statistical parameter representing the unpredictable deviation in the safety constraint process.

[0061] The beneficial effects of adopting the above further solution are as follows: by quantifying the impact of uncertainty, the original safety constraints can be corrected to generate robust constraint conditions. Thus, it can cover the worst-case perturbations, reduce the probability of out-of-control, enhance safety, and ensure braking stability.

[0062] In a second aspect, the present application provides an autonomous driving braking test device based on the fusion of bionic vision and touch, including:

[0063] An environment acquisition module, configured to acquire environmental information in the current test scenario based on a vision device. The environmental information includes weather, light, road surface information, obstacle information, and self-vehicle information. The road surface information includes road surface friction coefficient, road surface flatness, road surface material, road surface slope, road surface curvature, road boundary, and lane line information. The obstacle information includes obstacle type and obstacle attribute information. The self-vehicle information includes the current vehicle speed, vehicle body attitude, and vehicle body state;

[0064] A touch acquisition module, configured to simulate the collision process between the current vehicle and the obstacle and acquire touch information corresponding to the collision process based on a force feedback device;

[0065] A feature joint module, configured to perform multi-modal feature fusion processing on the original visual features corresponding to the environmental information and the original touch features corresponding to the touch information to obtain visual-touch joint features;

[0066] A prediction response module, configured to predict the body response information of the current vehicle corresponding to a candidate braking strategy based on the visual-touch joint features, and obtain the optimal braking strategy corresponding to the candidate braking strategy based on the body response information. The braking strategy is a parameter characterizing the braking control of the current vehicle in an emergency braking scenario. The body response information includes the longitudinal dynamic characteristics, lateral dynamic characteristics, vertical dynamic characteristics, and stability index of the vehicle;

[0067] A generation control signal module for generating a braking control signal based on the optimal braking strategy. Description of the Drawings

[0068] Figure 1 It is a schematic flowchart of the automatic driving braking test method based on the fusion of bionic vision and touch in the embodiment of the present application;

[0069] Figure 2 It is a schematic flowchart of obtaining the optimal braking strategy in the embodiment of the present application;

[0070] Figure 3 It is a structural block diagram of the automatic driving braking test device based on the fusion of bionic vision and touch in the embodiment of the present application. Detailed Description of the Embodiment

[0071] The following further describes the present application in detail with reference to the drawings.

[0072] The embodiment of the present application provides an automatic driving braking test method based on the fusion of bionic vision and touch. This method can be executed by a device, which can be a server or a terminal device. The server can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. The terminal device can be a smart phone, a tablet computer, a desktop computer, etc., but is not limited thereto.

[0073] As Figure 1 shown, an automatic driving braking test method based on the fusion of bionic vision and touch, with an electronic device as the execution subject, the main process of the method is described as follows (Steps S1 - S5): <##

[0074] Step S1: Based on the vision device, obtain the environmental information in the current test scene. The environmental information includes weather, light, road surface information, obstacle information, and vehicle information of the self - vehicle. The road surface information includes road surface friction coefficient, road surface flatness, road surface material, road surface slope, road surface curvature, road boundary, and lane line information. The obstacle information includes obstacle type and obstacle attribute information. The vehicle information of the self - vehicle includes the current vehicle speed, vehicle body attitude, and vehicle body state.

[0075] In this embodiment, the vision device can be a vision sensor. Preferably, the vision sensor can include a camera acquisition module and a lidar acquisition module. The camera acquisition module is used to acquire image information, and the lidar acquisition module is used to determine obstacle information.

[0076] The camera acquisition module can be a binocular stereo vision system composed of two zoom monochromatic cameras. The two zoom monochromatic cameras are respectively placed on the rearview mirrors on both sides of the vehicle to be measured. The horizontal field of view angle of the two zoom monochromatic cameras can be 110°, the vertical field of view angle can be 60°, the resolution can be 1920×1080 pixels, and the frame rate can be 30fps, which can meet the real-time requirements of autonomous driving. The two zoom monochromatic cameras can form a baseline distance of about 20-30 cm to obtain accurate visual information.

[0077] The lidar acquisition module can use a 16-line lidar. The detection range can be 100 meters, the angular resolution can be 0.1°, and the point cloud density can reach 300,000 points per second, which can provide high-precision spatial information.

[0078] Through parallax calculation, the electronic device can generate high-precision three-dimensional point cloud data, and the accuracy of the three-dimensional point cloud data can reach the centimeter level. In practical applications, the vision sensor can detect obstacles within a range of 100 meters and provide key information such as the relative position, size, and moving speed of the obstacles.

[0079] Environmental information refers to the set of dynamic three-dimensional space data outside the vehicle obtained through the vision sensor. Its perception range can cover a three-dimensional spherical space centered on the vehicle itself with a radius of 100 meters. The data update frequency can be not less than 30Hz, constituting the basic information layer for the autonomous driving system to perform environmental perception and decision-making.

[0080] Weather can include meteorological physical parameters and environmental interference degrees. Specifically, the meteorological physical parameters can include fog concentration (e.g., 0-500mg / m 3 )), wind speed (e.g., 0-30m / s), precipitation type (e.g., rain / snow / hail), and the environmental interference degree can include the coverage area ratio of raindrops on the lens surface (e.g., 0-100%), and the lens turbidity index (e.g., 0-1). Light can include optical characteristic parameters. Specifically, the optical characteristic parameters can include light intensity (e.g., 0-120kLux), visibility (e.g., 0-1000m), and rainfall level (e.g., 0-10mm / h).

[0081] The road surface flatness is the quantization value of unevenness such as potholes and bumps. The road surface material can be asphalt, cement, gravel, etc. The range of the friction coefficient μ can be expressed as 0.1-0.9. The friction coefficient μ corresponding to a dry asphalt road can be 0.8, the friction coefficient μ corresponding to a wet asphalt road can be 0.4, and the friction coefficient μ corresponding to an asphalt road covered with ice and snow can be 0.1-0.2.

[0082] The road surface slope can range from -15° to +15°. The road surface curvature refers to the degree of bending of the driving road, and the road surface curvature can affect the steering control of the vehicle under test. The road boundary can include a rigid boundary, a flexible boundary, and a virtual boundary. The rigid boundary can include a curb (e.g., height 15 ± 3 cm), a concrete guardrail (e.g., reflection intensity > 80 cd / lx), the flexible boundary can include a green belt (e.g., vegetation density eigenvalue 0.3 - 0.7), a rumble strip (e.g., tactile feedback frequency 50 Hz), and the virtual boundary can include a high-precision map digital fence (e.g., positioning accuracy 10 cm). The lane line information can include the type (e.g., solid line or dashed line), color, and position of the lane markings, etc.

[0083] The obstacle types can include static obstacles, dynamic obstacles, and special obstacles. Specifically, the static obstacles can include roadblocks, parked vehicles, buildings, curbs, etc., the dynamic obstacles can include pedestrians, cyclists, vehicles, animals, etc., and the special obstacles can include falling rocks, scattered objects, suddenly emerging obstacles, etc.

[0084] The obstacle attribute information includes the geometric characteristics, dynamic characteristics, and physical characteristics of the obstacles. Specifically, the geometric characteristics can include the size, shape, and position of the obstacles, the dynamic characteristics can include the speed, acceleration, and movement direction of the obstacles, and the physical characteristics can include the mass estimation and material hardness of the obstacles. The ego-vehicle information can include the vehicle motion state, body attitude, suspension state, tire state, powertrain state, braking system state, stability control system state, and vehicle mass distribution. Specifically, the vehicle motion state can include the current speed, acceleration, and yaw rate, the body attitude can include the pitch angle, roll angle, and yaw angle, the suspension state can include the compression amount of each wheel suspension, the tire state can include the load, slip ratio, and side slip angle of each tire, the powertrain state can include the engine speed and output torque, the braking system state can include the master cylinder pressure and the pressure distribution of each wheel, the stability control system state can include the ESP intervention degree, and the vehicle mass distribution can include the load condition and the centroid position.

[0085] In this embodiment, the main source of obtaining the ego-vehicle information can be the in-vehicle bus data. The original data of the vehicle ECU can be directly read through the CAN / FlexRay bus, and it can also have a refresh rate at the 10 ms level. Then, the ego-vehicle information is assisted by a vision device. Exemplarily, a vehicle-environment relative motion model can be established through a binocular stereo vision sensor, and the vehicle motion state can be deduced using the SLAM algorithm; the pixel displacement between the wheel arch and the wheel can be tracked to monitor the suspension state; the load can be deduced by the change of the body pitch angle to obtain an estimate of the vehicle mass distribution.

[0086] In this embodiment, more accurate depth information can be obtained through a vision device, which helps to more precisely locate obstacles and identify the environment, improving the accuracy of environmental perception. In addition to basic road surface information and obstacle information, environmental information also includes more extensive environmental factors such as weather, light, and vehicle information, enabling a more comprehensive assessment of the current driving environment and enhancing the comprehensiveness of environmental perception. By obtaining more comprehensive environmental information, various complex driving scenarios, such as bad weather or complex road conditions, can be better handled, improving adaptability in complex environments.

[0087] Step S2: Simulate the collision process between the current vehicle and the obstacle, and based on the force feedback device, obtain the tactile information corresponding to the collision process.

[0088] In this embodiment, a method combining finite element analysis (FEA) and rigid body dynamics models can be used. The environmental information can be used as input to simulate the collision process between the vehicle and different obstacles, and the impact force during the collision is output. The calculation resolution can reach 1 ms, and the calculation accuracy of the force can be ±0.1 N. The force feedback device obtains the impact tactile signal based on the impact force during the collision process, and fuses the impact tactile signal with other tactile signals from different sources and of different natures to obtain the tactile information. By performing internal fusion of multi-source tactile signals, the tactile information can comprehensively simulate the tactile feedback that a driver may feel in a real driving environment.

[0089] In this embodiment, the force feedback device may include a tactile sensor. The sensitivity of the tactile sensor can reach 0.05 N, and the sampling frequency is as high as 1000 Hz, capable of capturing the collision transient process at the millisecond level.

[0090] By setting specific environmental information, a collision process closer to the real situation can be simulated, improving the authenticity of the collision simulation. Through tactile simulation during the collision process, physical information can be converted into tactile signals that can be perceived by the human body, realizing the reproduction of human touch and improving the intuitiveness and interpretability of the test.

[0091] Step S3: Based on the original visual features corresponding to the environmental information and the original tactile features corresponding to the tactile information, perform multi-modal feature fusion processing to obtain visual-tactile joint features.

[0092] The electronic device includes a vision sub-network and a tactile sub-network. The vision sub-network is used to extract the original visual features corresponding to the environmental information, and the tactile sub-network is used to extract the original tactile features corresponding to the tactile information. The original visual features and the original tactile features are combined through a network fusion mechanism to obtain visual-tactile joint features.

[0093] In this embodiment, the vision sub-network can use ResNet-50 as the backbone network. The dimension of the last layer of feature map is 7×7×2048, and a 2048-dimensional vision feature vector is obtained through global average pooling. The tactile sub-network can be composed of three fully connected layers, and the dimensions of the hidden layers are 256, 512, and 256 respectively. The activation function can use ReLU, and finally a 256-dimensional tactile feature vector is output.

[0094] Step S4: Based on the visual-tactile joint feature, predict the body response information of the current vehicle corresponding to the candidate braking strategy, and based on the body response information, obtain the optimal braking strategy corresponding to the candidate braking strategy. The braking strategy is a parameter characterizing the braking control of the current vehicle in an emergency braking scenario. The body response information includes the longitudinal dynamic characteristics, lateral dynamic characteristics, vertical dynamic characteristics, and stability index of the vehicle.

[0095] In this embodiment, the body response information can also include tire dynamic characteristics and safety margin. The longitudinal dynamic characteristics can include speed change, acceleration / deceleration performance, braking distance. The lateral dynamic characteristics can include yaw rate, lateral acceleration, steering response. The vertical dynamic characteristics can include suspension displacement, vehicle body attitude (such as pitch, roll angle). The tire dynamic characteristics can include slip ratio, adhesion utilization rate. The vehicle stability index can include yaw rate, roll angle, understeer / oversteer judgment, roll risk assessment. The safety margin refers to the margin from the limit state, and the safety margin can include slip ratio margin and roll margin, etc.

[0096] The visual-tactile joint feature (1024-dimensional vector) can represent the spatio-semantic fusion information of the driving scenario, and can include features such as obstacle position / speed, road surface friction coefficient (μ), weather, current vehicle speed, yaw angle, centroid position, suspension compression, collision impact force waveform, high-frequency vibration energy distribution, etc.

[0097] Through the policy space, a combination of braking parameters can be sampled and generated. The parameter combination can include longitudinal braking force distribution, and the library in the longitudinal braking force distribution includes the front axle ratio and the rear axle ratio. When an emergency situation requiring braking is detected, the preset body response prediction model can predict the body response information such as braking distance and vehicle body stability corresponding to the candidate braking strategy according to conditions such as the road surface friction coefficient and the current vehicle speed in the current visual-tactile joint feature.

[0098] Step S5: Generate a braking control signal based on the optimal braking strategy.

[0099] In this embodiment, by integrating a vision device and a force feedback device, comprehensive acquisition of environmental information and tactile information is achieved, and the fusion perception of visual information and tactile information is realized. By integrating visual-tactile multimodal features, visual-tactile joint features are obtained, improving the perception accuracy and robustness. Based on the visual-tactile joint features, vehicle behavior can be accurately predicted, enhancing the test reliability. This application performs excellently in complex environments, providing strong guarantee for the driving safety of autonomous vehicles, promoting the industry to develop to a higher level, and providing important support for the safety assessment of autonomous driving.

[0100] In this embodiment, S2 may specifically include:

[0101] Based on the road surface information and collision information, simulate the collision process, where the collision information includes collision type, relative collision speed, collision angle, collision contact area, hardness characteristics of obstacle material, and collision force distribution and transmission path;

[0102] Based on the force feedback device, obtain the impact tactile signal during the collision process, and obtain the basic tactile signal during the collision process corresponding to the road surface information, the motion state tactile feedback signal during the collision process corresponding to the vehicle information, the vibration tactile signal during the collision process, and the pressure tactile signal during the collision process. The vehicle information further includes body attitude, suspension state, tire state, power system state, and braking system state;

[0103] Fuse the impact tactile signal with the basic tactile signal, the motion state tactile feedback signal, the vibration tactile signal, and the pressure tactile signal to obtain a comprehensive tactile signal stream;

[0104] Based on a preset mechanical and neural signal conversion model, obtain the tactile information corresponding to the comprehensive tactile signal stream;

[0105] The mechanical and neural signal conversion model is expressed as:

[0106]

[0107] where S(t1) represents the tactile information corresponding to the neural signal, F(t1) represents the comprehensive tactile signal stream corresponding to the physical force signal, α2 represents the weight coefficient simulating the human sensitivity to the magnitude of force, β2 represents the weight coefficient simulating the human sensitivity to the rate of change of force, and γ1 represents the weight coefficient simulating the human sensitivity to the duration of force. Exemplarily, α2 can be set to 1.0, β2 can be set to 0.5, and γ1 can be set to 0.2.

[0108] Collision types can include frontal, side, and rear-end collisions, which can be determined through 3D point cloud data analysis. The relative collision speed can be ∈[0, 120] km / h, which can be calculated through lidar multi-frame tracking. The collision angle can be ∈[0°, 360°], which can be the result of visual SLAM positioning. The collision contact area can be ∈[0, 2] m 2 , and finite element deformation simulation calculation. The hardness characteristics of the obstacle material can be ∈[0, 10], and it is matched with a preset parameter library. For example, the hardness characteristic of metal = 8 and the hardness characteristic of plastic = 3. The collision force distribution and transmission path can be collected through a pressure sensor array.

[0109] Based on road surface information and collision information, collision dynamics modeling is carried out to simulate the collision process. Based on a force feedback device, impact tactile signals, basic tactile signals, motion state tactile feedback signals, vibration tactile signals, and pressure tactile signals during the collision process are synchronously collected, and time synchronization marks can be made for the above multiple tactile signals. After fusing the above multiple tactile signals, a comprehensive tactile signal stream is obtained. By fusing tactile signals from multiple aspects, more comprehensive tactile feedback can be provided.

[0110] The fusion process of multiple tactile signals can specifically include: extracting the key features corresponding to various original tactile signals; performing feature fusion and enhancement on each key feature to complete the above fusion process.

[0111] Exemplarily, the impact tactile signal can be obtained by extracting the maximum gradient change rate of the force field matrix, the basic tactile signal can be obtained by calculating the product of the friction coefficient and the contact area, the motion tactile signal can be obtained by fusing the roll angle change rate and the suspension vibration frequency, the vibration tactile signal can be obtained by extracting the energy of 5 main frequency bands through Fourier transform, and the pressure tactile signal can be obtained by calculating the average pressure and gradient distribution in the contact area.

[0112] In this embodiment, the main functions of the mechanics and nerve signal conversion model include:

[0113] 1. Mapping from physical quantities to perceptual quantities: converting physical forces, pressures, vibrations, etc. into signals that can be perceived by the human nervous system, and simulating the response characteristics of human mechanoreceptors (such as Pacinian corpuscles, Meissner corpuscles, etc.);

[0114] 2. Simulation of perceptual characteristics: considering the characteristic that the human body is more sensitive to the rate of change of force than the absolute force value; simulating the non-linear characteristics and saturation effect of force perception; considering the time integration effect of force perception;

[0115] 3. Generation of multi-dimensional tactile experience: combining multi-dimensional tactile characteristics such as force, texture, temperature, etc.; simulating the cooperative response of different types of receptors; generating a comprehensive tactile experience.

[0116] Through the mechanical and neural signal conversion model, the physical force signal F(t) can be converted into a neural tactile signal S(t), which can accurately simulate the different sensitivities of the human body to the magnitude, change rate, and duration of force.

[0117] In this embodiment, the S3 may specifically include:

[0118] Based on a preset first deep feature extraction network, extract the original visual features, where the original visual features include image semantic information, regions of interest, and a first spatio-temporal information distribution feature;

[0119] Based on a preset second deep feature extraction network, extract the original tactile features, where the original tactile features include a second spatio-temporal information distribution feature;

[0120] Based on the multi-head self-attention structure, align the original visual features and the original tactile features, and based on a preset bilinear pooling algorithm, perform an outer product operation on the aligned original visual features and original tactile features to obtain the fused visual-tactile joint features.

[0121] The visual sub-network is the first deep feature extraction network, and the tactile sub-network is the second deep feature extraction network. The original visual data and environmental information can be used as the input of the first deep feature extraction network, and the first deep feature extraction network can output the original visual features. Among them, the image semantic information can be the semantic segmentation result, the region of interest can be the obstacle bounding box, and the first spatio-temporal information distribution feature can be a 2048-dimensional vector time-series - spatial encoding vector. The original visual data can be an RGB image sequence collected by a visual device.

[0122] The dimension of the feature map of the last layer corresponding to the first deep feature extraction network can be 7×7×2048. The 7×7×2048 feature map is compressed into a 2048-dimensional vector through global average pooling, retaining the global semantic information.

[0123] In this embodiment, data such as the original tactile data, the current vehicle speed, and the vehicle body state can be used as the input of the second deep feature extraction network. The second deep feature extraction network can output the original tactile features. The second spatio-temporal information distribution feature can be a 256-dimensional time-series - mechanical encoding vector. The original tactile data can be a multi-dimensional force signal collected by a tactile sensor. The second deep feature extraction network can be composed of three fully connected layers, and the hidden layer dimensions are 256, 512, and 256 respectively. The activation function can use ReLU, and finally, a feature vector corresponding to the 256-dimensional original tactile features can be output.

[0124] In this embodiment, the temporal sequence features of the input of the first depth feature extraction network and the input of the second depth feature extraction network can be extracted respectively to obtain the spatio-temporal information distribution features describing the corresponding entire data stream, so as to obtain the original tactile features with translational invariance and the original tactile features.

[0125] As an alternative implementation of this embodiment, the original visual data and environmental information can be processed by a convolutional neural network to extract feature maps of 4 different scales, and the feature dimensions are 64×64×32, 32×32×64, 16×16×128, 8×8×256 respectively, so as to capture the spatial information of different scales; the original tactile data, the current vehicle speed, the vehicle body state and other data can be processed by a long short-term memory network (LSTM) to extract a 512-dimensional temporal feature vector, and the time window is set to 250 ms, which is sufficient to capture typical collision events.

[0126] The multi-head self-attention structure processes the original visual features and the original tactile features simultaneously to align the features of the two and obtain the feature alignment result. The structural parameters of the multi-head self-attention alignment structure can include: the number of heads is 8, the hidden layer dimension is 512, and the positional encoding generates spatio-temporal position markers by sine / cosine functions.

[0127] The fusion of the original visual features and the original tactile features can adopt the Bilinear Pooling method, and the bilinear pooling algorithm can be expressed as:

[0128]

[0129] where V represents the original visual features, T represents the original tactile features, represents the outer product operation, and W represents the learnable projection matrix. The feature dimension of the finally fused visual-tactile joint features is 1024, which is sufficient to characterize the characteristics of complex driving scenarios.

[0130] The first depth feature extraction network can extract high-dimensional semantic features (2048 dimensions) through ResNet-50 and enhance the key regions by combining spatio-temporal attention. The second depth feature extraction network can fuse information such as touch and vehicle movement through a three-layer fully connected structure (256-512-256) and output 256-dimensional dynamic features. By combining the tactile information with the information related to vehicle movement, the mapping relationship between the road surface condition, the collision event and the vehicle dynamic response can be established, and the understanding and prediction ability of complex scenarios can be improved. By solving the spatio-temporal misalignment through attention alignment and realizing cross-modal interaction fusion through bilinear pooling, 1024-dimensional joint features are finally generated.

[0131] In this embodiment, after aligning the original visual features and the original tactile features based on the multi-head self-attention structure, it further includes:

[0132] Based on a preset contrast loss function, perform similarity learning on the aligned original visual features and original tactile features;

[0133] The contrast loss function is expressed as:

[0134]

[0135] where Y represents a binary label. When the aligned original visual features and original tactile features come from the same test scenario, Y is 0; otherwise, Y is 1; D w represents the feature distance, which is a parameter characterizing the difference between the aligned original visual features and original tactile features; m represents a set boundary value, and the feature distance between the original visual features and original tactile features from different test scenarios is at least greater than the boundary value.

[0136] In this embodiment, the Euclidean distance can be used to measure the cross-modal feature difference. When the aligned original visual features and original tactile features come from the same test scenario (i.e., a positive sample pair), D w should approach 0; when from different test scenarios (negative sample pairs), D w needs to be greater than the boundary value m.

[0137] After aligning the original visual features and original tactile features, in this embodiment, the contrast loss function (Contrastive Loss) can further optimize the semantic consistency of the cross-modal features. Specifically, the original visual features and original tactile features in the same scenario should have a high degree of similarity (the feature distance approaches 0), enhancing the intra-modal consistency; the cross-modal feature distances in different scenarios need to exceed the preset boundary value to avoid confusion and strengthen the inter-modal distinguishability.

[0138] As Figure 2 shown, in this embodiment, S4 can specifically include:

[0139] S41. Based on a preset magic formula and the visual-tactile joint features, calculate the tire forces corresponding to each wheel when executing the candidate braking strategy. The tire forces include longitudinal forces and lateral forces;

[0140] S42. Based on each of the tire forces and the vehicle external forces, calculate the overall force information of the current vehicle. The vehicle external forces include air resistance and gravity;

[0141] S43. Based on the overall force information, calculate the vehicle acceleration. The vehicle acceleration includes longitudinal acceleration, lateral acceleration, vertical acceleration, yaw acceleration, pitch acceleration, and roll acceleration;

[0142] S44. Based on the vehicle acceleration, the current vehicle speed, the current position of the current vehicle, and a preset numerical integration algorithm, obtain the predicted speed and predicted position of the current vehicle at the next moment; and based on the predicted speed and predicted position, obtain the body response information corresponding to the candidate braking strategy;

[0143] S45. Based on the body response information, perform a non-linear time-varying analysis to obtain the non-linear time-varying characteristics existing between the tires of the current vehicle and the ground. The non-linear time-varying characteristics include non-linear friction coefficient, non-linear side slip characteristics, and time-varying road adhesion;

[0144] S46. If the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, adjust the candidate braking strategy; if the non-linear time-varying characteristics do not exceed the corresponding current safety constraint conditions, use the candidate braking strategy as the optimal braking strategy;

[0145] S47. Use the adjusted candidate braking strategy as the new candidate braking strategy, and repeat the steps of S41 to S46 until the optimal braking strategy is obtained.

[0146] In this embodiment, the wheel force analysis uses the classical Magic Formula to calculate the tire force. The Magic Formula can be expressed as:

[0147] F = Dsin(Carctan(Bx - E(Bx - arctan(Bx))))

[0148] where F represents the tire force; x represents the slip ratio or the side slip angle. When x is the slip ratio, the tire force is the longitudinal force, and when x is the side slip angle, the tire force is the lateral force; D represents the peak factor, which determines the magnitude of the maximum tire force; C represents the shape factor, which affects the curve shape; B represents the stiffness factor, which is related to the initial slope of the curve; E represents the curvature factor, which determines the peak position and the characteristics of the descending section; B, C, D, and E are empirical parameters, and the typical values can be 10, 1.9, 1, and 0.97 respectively.

[0149] The calculation method of the slip ratio can be expressed as:

[0150] Longitudinal slip ratio = (ω7r - v) / max(ω7r, v)

[0151] where ω7 represents the tire angular velocity, r represents the effective radius of the tire, and v represents the wheel center line speed. For example, the braking slip ratio is positive (0 - 1), the driving slip ratio is negative (-1 - 0), the slip ratio is 1 when completely locked, and the slip ratio is 0 when in pure rolling.

[0152] The calculation method of the side slip angle can be expressed as:

[0153] Slip angle = arctan(v y / v x )

[0154] where v y represents the lateral velocity of the tire, and v x represents the longitudinal velocity of the tire.

[0155] Applying Newton's laws of motion and solving for the vehicle acceleration corresponding to the current vehicle based on the overall force information, a six-dimensional acceleration vector can be obtained, namely longitudinal acceleration, lateral acceleration, vertical acceleration, yaw acceleration, pitch acceleration, and roll acceleration.

[0156] In this embodiment, the numerical integration algorithm can be the Runge-Kutta integration method, and the time step can be 1 - 10 ms.

[0157] The inputs for the analysis of non-linear time-varying characteristics can include: the combined visual and tactile features, i.e., a 1024-dimensional fusion vector, which characterizes road surface adhesion, collision risk, etc.; candidate braking strategies, including parameters such as braking force distribution ratio, braking pressure gradient, etc.

[0158] Based on different tire motion states, the non-linear time-varying characteristics existing between the tire and the ground are identified. Specifically, the tire motion states can include:

[0159] Pure rolling state: The tire has no slip or side slip, such as driving in a straight line at a constant speed;

[0160] Longitudinal slip state: Slip occurs between the tire and the ground during hard acceleration or hard braking;

[0161] Lateral slip state: The tire generates a slip angle during turning;

[0162] Combined slip state: Both longitudinal slip and lateral slip exist simultaneously, such as braking during turning;

[0163] Free rolling state: The tire rolls without being subjected to driving or braking forces;

[0164] Locked state: The tire is completely locked and no longer rotates, such as when the emergency braking ABS fails.

[0165] The non-linear time-varying characteristics can include:

[0166] Nonlinear friction coefficient: The phenomenon that the friction coefficient first increases and then decreases as the slip ratio increases;

[0167] Nonlinear side slip characteristics: The lateral force no longer has a linear relationship with the side slip angle at large side slip angles;

[0168] Tire elastic hysteresis: The hysteresis phenomenon of the tire force and deformation during the loading and unloading processes;

[0169] Time-varying road surface adhesion: For example, the friction coefficient changes suddenly when driving through a puddle;

[0170] Temperature dependence: The change in tire temperature leads to the change of friction characteristics;

[0171] Vertical load sensitivity: The non-linear phenomenon that the change of vertical force leads to the change of friction coefficient.

[0172] The processing of non-linear time-varying characteristic analysis can include: identifying the time-varying road surface adhesion, extracting the change rate of the road surface friction coefficient from the joint visual and tactile features. For example, when it is detected that the water film thickness increases, the road surface friction coefficient can be decreased by 30%; analyzing the non-linear cornering characteristics. When the cornering angle > 8°, the lateral force saturation correction can be triggered. For example, the lateral force attenuation coefficient = 0.7.

[0173] The current safety constraints can include the slip rate threshold, the roll angle threshold, and the predicted time to collision TTC (i.e., the time-to-collision data). For example, the slip rate threshold can be 0.2, the roll angle threshold can be 15°, and emergency braking is triggered when the predicted time to collision TTC is less than 1.5 s. If the slip rate corresponding to the predicted current candidate braking strategy is greater than 0.2 according to the joint visual and tactile features, it is determined that the safety constraint is exceeded.

[0174] In the electronic device, there is a mapping relationship between the comparison of the real-time non-linear time-varying characteristics and the corresponding current safety constraint conditions, and the correction of the current candidate braking strategy. Exemplarily, when the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, the braking force distribution gradient can be decreased (such as from 500 Bar / ms to 300 Bar / ms), and the ESP intervention intensity can be increased.

[0175] In this embodiment, the non-linear time-varying analysis can adopt the stochastic differential equation modeling. Considering the random change of the road surface friction coefficient, a state space model is established. The non-linear time-varying analysis model can be expressed as:

[0176]

[0177] Wherein, X represents the state vector, characterizing the tire motion state and mechanical characteristics, and is directly related to the real-time motion state and non-linear mechanical response of the tire; u represents the control input, such as the instruction affecting the tire motion state corresponding to the candidate braking strategy; w(t2) represents the noise term, describing the uncertainty in the tire-ground interaction; f and g represent non-linear functions. The f function can accurately describe the non-linear time-varying relationship of the tire force, and the g function can quantify the amplification or suppression effect of the random perturbation on the tire state.

[0178] In this embodiment, the Deep Deterministic Policy Gradient (DDPG) can be used to optimize the candidate braking strategy to obtain the optimal braking strategy. The DDPG algorithm can include an Actor network and a Critic network. The Actor network can directly output continuous action values, and the Critic network can evaluate the action value. By inputting the vehicle body response information into the Actor network and the Critic network, the vehicle body response can be bound to the action, and the value of the action in a specific state can be evaluated. Through the coupling of the vehicle body response information and the Q-value gradient, the policy gradient optimization can be directly targeted at the vehicle dynamic characteristics (such as reducing side slip and suppressing pitching).

[0179] The optimization objective is the loss function minimized by the Critic network, and the minimized loss function can be expressed as:

[0180]

[0181] where y i = r i + γ2Q′(s i+1 , μ′(s i+1 |θ μ′ )|θ Q′ ) is the target value, γ2 is the discount factor used to balance the weight of the current reward and future benefits. The larger the value, the more importance is attached to the long-term return. The discount factor can be set to 0.99. θ Q and θ μ are the parameters of the Critic network and the Actor network respectively; θ Q′ and θ μ′ are the corresponding target network parameters. The size of the experience replay buffer can be set to 10,000, which ensures the independence of training samples and reduces the possibility of overfitting. N represents the batch size, and the batch size can be 64, which can balance the gradient estimation accuracy and the computational efficiency. The learning rate can be 0.001, which can control the parameter update step size and ensure the convergence stability.

[0182] r i represents the immediate reward, reflecting the immediate benefit of the current state-action pair. The immediate reward function can be expressed as:

[0183]

[0184] where S i represents the braking distance, θ i represents the pitching angle, S margin represents the slip rate margin, ω1 represents the braking efficiency weight, ω2 represents the comfort weight, and ω3 represents the stability weight.

[0185] Constructing an immediate reward function based on vehicle body response information can guide the Critic network to learn a braking strategy that balances safety, comfort, stability, and efficiency. It is easy to understand that an immediate reward function can also be constructed by combining the vehicle body response information with the comprehensive collision risk value between the vehicle and the obstacle.

[0186] In this embodiment, before determining whether the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, it further includes:

[0187] Based on the second uncertainty quantification data, the original safety constraint conditions are corrected to obtain the corrected robust constraint conditions, and the robust constraint conditions are used as the current safety constraint conditions. The second uncertainty quantification data is a statistical parameter representing the unpredictable deviation in the safety constraint process.

[0188] The second uncertainty quantification data represents the random deviation that cannot be accurately modeled in the safety constraint analysis, and may include sensor noise statistics, model prediction errors, and environmental disturbance statistics. The mathematical representation of the second uncertainty quantification data can be a Gaussian distribution or a non-Gaussian distribution.

[0189] The correction objective of the original safety constraint conditions can be: embedding a safety margin in the original constraint conditions to cover the impact of uncertainty. The correction method of the original safety constraint conditions can include threshold relaxation or probabilistic constraints. Threshold relaxation is to expand the tolerance range of the safety boundary according to the uncertainty statistical parameters, and probabilistic constraints are to convert the deterministic constraints into a probabilistic guarantee form.

[0190] By quantifying the impact of uncertainty, the original safety constraints can be corrected to generate robust constraint conditions. Thus, it can cover the worst-case perturbations, reduce the out-of-control probability, improve safety, and ensure braking stability.

[0191] In this embodiment, the S5 includes:

[0192] Based on the vehicle body dynamics model and the vehicle body response information corresponding to the optimal braking strategy, obtain the longitudinal control signal and the lateral control signal for implementing the optimal braking strategy. The vehicle body response information includes the first uncertainty quantification data of the control signal, and the first uncertainty quantification data is a parameter of the confidence level for evaluating the reliability of the control signal;

[0193] Based on the first uncertainty quantification data, perform uncertainty weighted fusion on the longitudinal control signal and the lateral control signal to obtain the fused longitudinal control signal and lateral control signal, and calculate the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle;

[0194] If the comprehensive collision risk value exceeds a preset risk threshold, a braking control signal is generated based on the fused longitudinal control signal and lateral control signal.

[0195] The vehicle body dynamics model is a vehicle body dynamics model of a vehicle body response prediction model based on physical information. The vehicle body dynamics model can output a longitudinal control signal and a lateral control signal. The vehicle body response information includes first uncertainty quantification data, and the first uncertainty quantification data can include longitudinal signal confidence, lateral signal confidence, and an uncertainty covariance matrix.

[0196] By simultaneously considering longitudinal and lateral motions, more comprehensive braking control is achieved, improving the comprehensiveness of the braking strategy. Through the robustness management module, uncertainties can be handled, improving the reliability of the strategy in complex environments and enhancing the robustness of the braking strategy. By separately planning longitudinal and lateral motions, more refined braking control can be achieved, improving the braking effect and optimizing the braking effect. The strategy design based on collision risk assessment can better handle potential collision risks and enhance safety.

[0197] In this embodiment, calculating the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle specifically may include:

[0198] Based on a preset comprehensive evaluation algorithm, comprehensively evaluate the collision probability and time-to-collision data to obtain the comprehensive collision risk value;

[0199] The comprehensive evaluation algorithm is expressed as:

[0200]

[0201] where α1 represents a weight coefficient, and the weight coefficient is used to determine the priority of the time-to-collision data and the collision probability; β1 represents a decay coefficient, and the decay coefficient is used to control the decay rate of the time-to-collision data; P collision represents the collision probability; TTC represents the time-to-collision data; and Risk represents the comprehensive collision risk value.

[0202] The comprehensive collision risk assessment algorithm realizes multi-dimensional risk assessment of the comprehensive collision risk value by quantifying the collision probability and time urgency, avoiding over-braking or delayed response. The collision probability can be updated in real time. When the Risk value exceeds the risk threshold, emergency braking is triggered, and the risk threshold can be set to 0.8.

[0203] In this embodiment, a smoothing function can be used for longitudinal deceleration planning, and the smoothing function can be expressed as:

[0204]

[0205] Among them, a max represents the maximum deceleration, which can be set to 9.8 m / s 2 , γ3 represents the smoothing factor, which can take a value of 0.3 to ensure the balance between passenger comfort and vehicle stability.

[0206] By adopting a smoothing function for longitudinal deceleration planning, the following beneficial effects are achieved:

[0207] 1. Smooth braking force transition: Avoid the impact caused by sudden changes in braking force, reduce the pitching fluctuation of the vehicle body during braking, and improve passenger comfort;

[0208] 2. Optimize braking efficiency: According to the tire slip ratio characteristics, avoid frequent intervention of ABS; Maximize the braking force within the allowable range of tire-road adhesion; Gradually increase the braking force to avoid instantaneous locking of the tires;

[0209] 3. Consider the vehicle's dynamic response: Adapt to the suspension response time; Reduce the instability caused by load transfer; Keep the vehicle body attitude stable;

[0210] 4. Ensure the realizability of control signals: Consider the dynamic characteristics of the braking system actuator; Prevent the braking system from saturating; Adapt to the response characteristics of different braking actuators.

[0211] For example, at the beginning of emergency braking, the electronic device will gradually increase the braking force to the maximum value according to the smoothing function instead of immediately applying the maximum braking force to obtain better braking effect and stability.

[0212] In this embodiment, before calculating the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle, the following steps are further included:

[0213] Based on the visual-tactile joint features and a preset adaptive dynamic scene generation algorithm, randomly search for the candidate motion trajectories of the obstacle in different scenarios to obtain a set of candidate trajectory branches;

[0214] Based on a preset reinforcement learning algorithm, obtain the optimized trajectory parameters corresponding to the set of candidate trajectory branches, where the trajectory parameters include control point coordinates and speed curves;

[0215] Based on a preset Bessel curve formula, obtain the first target trajectory of the obstacle corresponding to the trajectory parameters;

[0216] Predict the second target trajectory of the current vehicle, and based on the first target trajectory and the second target trajectory, obtain the collision probability.

[0217] In this embodiment, before calculating the comprehensive collision risk value, the obstacle motion trajectory is generated and optimized through an adaptive dynamic scenario generation algorithm and reinforcement learning, and the collision probability is calculated based on the trajectory prediction of the host vehicle and the obstacle.

[0218] The adaptive dynamic scenario generation algorithm can adopt a combination of Monte Carlo Tree Search (MCTS) and evolutionary algorithms. The reward function of Monte Carlo Tree Search can be expressed as:

[0219] R = ω3·Risk + ω4·Diversity - ω5·Similarity

[0220] Where Risk represents the scene risk degree, that is, the comprehensive collision risk value between the obstacle trajectory and the current vehicle; Diversity represents the diversity with the tested scenes, that is, the difference degree from the generated trajectories; Similarity represents the similarity with the real scene, that is, the matching degree with the real accident scene; ω3, ω4, and ω5 respectively represent weight coefficients, and ω3, ω4, and ω5 can be set to 0.5, 0.3, and 0.2 respectively to balance the test coverage rate and authenticity.

[0221] The generation of the obstacle trajectory can be parameterized by the Bezier Curve. The Bezier Curve formula can be expressed as:

[0222]

[0223] Where P i is the control point, n is the curve order, the typical value is 3, and t4 is the parameter. By randomizing the positions of the control points, diverse motion trajectories can be generated to cover various possible dangerous scenarios. The control points are the key points used to define the shape of the Bezier curve, and the position of each control point determines the direction and curvature of the curve. The parameter t4 can be a variable that varies within the interval [0, 1] and is used to represent the position ratio of the curve from the starting point to the ending point.

[0224] In the autonomous driving test, by randomizing and optimizing the control points and combining the time mapping of t, it is possible to more accurately perceive the surrounding situation in a complex environment, obtain the collision probability, and efficiently generate diverse test dangerous scenarios, improving the test coverage rate and authenticity, thereby achieving a comprehensive and accurate evaluation of the emergency braking performance of autonomous driving.

[0225] Based on the same technical concept, the present application also provides an autonomous driving braking test device based on the fusion of bionic vision and touch, as Figure 3 shown. The autonomous driving braking test device 200 based on the fusion of bionic vision and touch mainly includes:

[0226] Obtain an environment module 201, which is used to obtain environmental information in the current test scenario based on a vision device. The environmental information includes weather, light, road surface information, obstacle information, and vehicle information of the host vehicle. The road surface information includes road surface friction coefficient, road surface flatness, road surface material, road surface slope, road surface curvature, road boundary, and lane line information. The obstacle information includes obstacle type and obstacle attribute information. The vehicle information of the host vehicle includes the current vehicle speed, vehicle body attitude, and vehicle body state;

[0227] Obtain a tactile module 202, which is used to simulate the collision process between the current vehicle and the obstacle, and obtain tactile information corresponding to the collision process based on a force feedback device;

[0228] A feature combination module 203, which is used to perform multimodal feature fusion processing on the original visual features corresponding to the environmental information and the original tactile features corresponding to the tactile information to obtain visual-tactile combined features;

[0229] A prediction response module 204, which is used to predict the vehicle body response information of the current vehicle corresponding to a candidate braking strategy based on the visual-tactile combined features, and obtain the optimal braking strategy corresponding to the candidate braking strategy based on the vehicle body response information. The braking strategy is a parameter characterizing the braking control of the current vehicle in an emergency braking scenario. The vehicle body response information includes the longitudinal dynamic characteristics, lateral dynamic characteristics, vertical dynamic characteristics, and stability index of the vehicle;

[0230] A generated control signal module 205, which is used to generate a braking control signal based on the optimal braking strategy

[0231] Optionally, the generated control signal module 205 includes:

[0232] A first acquisition sub-module, which is used to obtain a longitudinal control signal and a lateral control signal for implementing the optimal braking strategy based on a vehicle body dynamics model and the vehicle body response information corresponding to the optimal braking strategy. The vehicle body response information includes first uncertainty quantification data of the control signal, and the first uncertainty quantification data is a parameter of the confidence level for evaluating the reliability of the control signal;

[0233] A weighted fusion sub-module, which is used to perform uncertainty weighted fusion on the longitudinal control signal and the lateral control signal based on the first uncertainty quantification data to obtain a fused longitudinal control signal and a lateral control signal, and calculate a comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle;

[0234] A signal generation sub-module, configured to generate the braking control signal based on the fused longitudinal control signal and lateral control signal when the comprehensive collision risk value exceeds a preset risk threshold.

[0235] Optionally, the weighted fusion sub-module includes:

[0236] A comprehensive evaluation sub-module, configured to comprehensively evaluate the collision probability and time-to-collision data based on a preset comprehensive evaluation algorithm to obtain the comprehensive collision risk value

[0237] Optionally, before the weighted fusion sub-module, it includes:

[0238] A random search sub-module, configured to randomly search for candidate motion trajectories of obstacles in different scenarios based on the visual-tactile joint features and a preset adaptive dynamic scene generation algorithm to obtain a set of candidate trajectory branches;

[0239] An obtaining trajectory parameter sub-module, configured to obtain optimized trajectory parameters corresponding to the set of candidate trajectory branches based on a preset reinforcement learning algorithm, where the trajectory parameters include control point coordinates and speed curves;

[0240] A second obtaining sub-module, configured to obtain a first target trajectory of the obstacle corresponding to the trajectory parameters based on a preset Bezier curve formula;

[0241] A third obtaining sub-module, configured to predict a second target trajectory of the current vehicle, and obtain the collision probability based on the first target trajectory and the second target trajectory.

[0242] Optionally, the obtaining tactile module 202 includes:

[0243] A simulated collision sub-module, configured to simulate the collision process based on the road surface information and collision information, where the collision information includes collision type, collision relative speed, collision angle, collision contact area, obstacle material hardness characteristics, and collision force distribution and transmission path;

[0244] A fourth obtaining sub-module, configured to obtain the impact tactile signal during the collision process based on the force feedback device, and obtain the basic tactile signal during the collision process corresponding to the road surface information, the motion state tactile feedback signal during the collision process corresponding to the vehicle information, the vibration tactile signal during the collision process, and the pressure tactile signal during the collision process, where the vehicle information further includes body attitude, suspension state, tire state, power system state, and braking system state;

[0245] A signal fusion sub-module, configured to fuse the impact haptic signal with the basic haptic signal, the motion state haptic feedback signal, the vibration haptic signal, and the pressure haptic signal to obtain a comprehensive haptic signal stream;

[0246] A fifth acquisition sub-module, configured to acquire the haptic information corresponding to the comprehensive haptic signal stream based on a preset mechanical and neural signal conversion model.

[0247] Optionally, the feature joint module 203 includes:

[0248] A first extraction sub-module, configured to extract the original visual features based on a preset first deep feature extraction network, where the original visual features include image semantic information, regions of interest, and first spatio-temporal information distribution features;

[0249] A second extraction sub-module, configured to extract the original haptic features based on a preset second deep feature extraction network, where the original haptic features include second spatio-temporal information distribution features;

[0250] An alignment and fusion sub-module, configured to align the original visual features and the original haptic features based on a multi-head self-attention structure, and perform an outer product operation on the aligned original visual features and original haptic features based on a preset bilinear pooling algorithm to obtain the fused visual-haptic joint features.

[0251] Optionally, after the alignment and fusion sub-module, it further includes:

[0252] A similarity learning sub-module, configured to perform similarity learning on the aligned original visual features and original haptic features based on a preset contrast loss function.

[0253] Optionally, the prediction response module 204 includes:

[0254] A first calculation sub-module, configured to calculate the tire forces corresponding to each wheel when executing the candidate braking strategy based on a preset magic formula and the visual-haptic joint features, where the tire forces include longitudinal forces and lateral forces;

[0255] A second calculation sub-module, configured to calculate the overall force information of the current vehicle based on each of the tire forces and vehicle external forces, where the vehicle external forces include air resistance and gravity;

[0256] A third calculation sub-module, configured to calculate the vehicle acceleration based on the overall force information, where the vehicle acceleration includes longitudinal acceleration, lateral acceleration, vertical acceleration, yaw acceleration, pitch acceleration, and roll acceleration;

[0257] A prediction response sub-module, configured to obtain the predicted speed and predicted position of the current vehicle at the next moment based on the vehicle acceleration, the current vehicle speed, the current position of the current vehicle, and a preset numerical integration algorithm; and obtain the body response information corresponding to the candidate braking strategy based on the predicted speed and predicted position.

[0258] A non-linear time-varying analysis sub-module, configured to perform non-linear time-varying analysis based on the body response information to obtain the non-linear time-varying characteristics existing between the tires of the current vehicle and the ground, where the non-linear time-varying characteristics include non-linear friction coefficient, non-linear side slip characteristics, and time-varying road adhesion.

[0259] An adjustment sub-module, configured to adjust the candidate braking strategy if the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions; and use the candidate braking strategy as the optimal braking strategy if the non-linear time-varying characteristics do not exceed the corresponding current safety constraint conditions.

[0260] A repetition sub-module, configured to use the adjusted candidate braking strategy as a new candidate braking strategy, and repeatedly execute the processing from the first calculation sub-module to the adjustment sub-module until the optimal braking strategy is obtained.

[0261] Optionally, before the adjustment sub-module, it includes:

[0262] A correction constraint sub-module, configured to correct the original safety constraint conditions based on the second uncertainty quantification data to obtain the corrected robust constraint conditions, and use the robust constraint conditions as the current safety constraint conditions, where the second uncertainty quantification data is a statistical parameter representing the unpredictable deviation in the safety constraint process.

[0263] In one example, the modules in any of the above devices may be one or more integrated circuits configured to implement the above methods. For example: one or more application specific integrated circuits (ASICs), or, one or more digital signal processors (DSPs), or, one or more field programmable gate arrays (FPGAs), or a combination of at least two of these integrated circuit forms.

[0264] For another example, when the modules in the device can be implemented in the form of a processing element scheduler, the processing element can be a general-purpose processor, such as a central processing unit (CPU) or other processors that can call programs. For another example, these modules can be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0265] In this application, names may be given to various objects such as various messages / information / devices / network elements / systems / devices / actions / operations / processes / concepts, etc. It can be understood that these specific names do not constitute a limitation on the relevant objects, and the given names may change with factors such as scenarios, contexts, or usage habits. The understanding of the technical meaning of the technical terms in this application should be mainly determined from the functions and technical effects reflected / executed in the technical solutions.

[0266] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and modules described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0267] Those of ordinary skill in the art can realize that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0268] The term "comprising", "including" or any other variant thereof is intended to cover a non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0269] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of this application, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0270] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0271] Although the embodiments of this application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting this application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. An automatic driving braking test method based on the fusion of bionic vision and touch, characterized in that, Including: S1. Based on a vision device, obtain environmental information in the current test scenario. The environmental information includes weather, light, road surface information, obstacle information, and self-vehicle information. The road surface information includes road surface friction coefficient, road surface flatness, road surface material, road surface slope, road surface curvature, road boundary, and lane line information. The obstacle information includes obstacle type and obstacle attribute information. The self-vehicle information includes the current vehicle speed, vehicle body attitude, and vehicle body state; S2. Simulate the collision process between the current vehicle and the obstacle, and based on a force feedback device, obtain tactile information corresponding to the collision process; S3. Based on the original visual features corresponding to the environmental information and the original tactile features corresponding to the tactile information, perform multi-modal feature fusion processing to obtain visual-tactile joint features; S4. Based on the visual-tactile joint features, predict the vehicle body response information of the current vehicle corresponding to the candidate braking strategy, and based on the vehicle body response information, obtain the optimal braking strategy corresponding to the candidate braking strategy. The braking strategy is a parameter characterizing the braking control of the current vehicle in an emergency braking scenario. The vehicle body response information includes the longitudinal dynamic characteristics, lateral dynamic characteristics, vertical dynamic characteristics, and stability index of the vehicle; S5. Based on the optimal braking strategy, generate a braking control signal.

2. The method for testing the braking of an autonomous vehicle based on the fusion of bionic vision and touch according to claim 1, wherein The S5 includes: Based on the vehicle body dynamics model and the vehicle body response information corresponding to the optimal braking strategy, obtain the longitudinal control signal and the lateral control signal for implementing the optimal braking strategy. The vehicle body response information includes the first uncertainty quantification data of the control signal, and the first uncertainty quantification data is a parameter of the confidence level for evaluating the reliability of the control signal; Based on the first uncertainty quantification data, perform uncertainty weighted fusion on the longitudinal control signal and the lateral control signal to obtain the fused longitudinal control signal and lateral control signal, and based on the collision probability between the current vehicle and the obstacle, calculate the comprehensive collision risk value corresponding to the optimal braking strategy; If the comprehensive collision risk value exceeds a preset risk threshold, then based on the fused longitudinal control signal and lateral control signal, generate the braking control signal.

3. The method for testing the automatic driving braking based on the fusion of bionic vision and touch according to claim 2, characterized in that, The calculating the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle includes: Based on a preset comprehensive evaluation algorithm, comprehensively evaluate the collision probability and the time-to-collision data to obtain the comprehensive collision risk value; The comprehensive evaluation algorithm is expressed as: Among them, α1 represents a weight coefficient, and the weight coefficient is used to determine the priority of the time-to-collision data and the collision probability; β1 represents a decay coefficient, and the decay coefficient is used to control the decay rate of the time-to-collision data; P collision represents the collision probability; TTC represents the time-to-collision data; Risk represents the comprehensive collision risk value.

4. The method for testing the braking of an autonomous vehicle based on the fusion of bionic vision and touch according to claim 2, wherein, Before calculating the comprehensive collision risk value corresponding to the optimal braking strategy based on the collision probability between the current vehicle and the obstacle, it further includes: Based on the visual-tactile joint features and a preset adaptive dynamic scenario generation algorithm, randomly search for the candidate motion trajectories of the obstacle in different scenarios to obtain a set of candidate trajectory branches; Based on a preset reinforcement learning algorithm, obtain the optimized trajectory parameters corresponding to the set of candidate trajectory branches. The trajectory parameters include control point coordinates and speed curves; Based on a preset Bessel curve formula, obtain the first target trajectory of the obstacle corresponding to the trajectory parameters; Predict the second target trajectory of the current vehicle, and based on the first target trajectory and the second target trajectory, obtain the collision probability.

5. A method for testing the braking of an autonomous vehicle based on the fusion of bionic vision and touch according to any one of claims 1 to 4, characterized in that, The S2 includes: Based on the road surface information and collision information, simulate the collision process, where the collision information includes collision type, relative collision speed, collision angle, collision contact area, hardness characteristics of the obstacle material, and collision force distribution and transmission path; Based on the force feedback device, obtain the impact tactile signal during the collision process, and obtain the basic tactile signal during the collision process corresponding to the road surface information, the motion state tactile feedback signal during the collision process corresponding to the vehicle information of the own vehicle, the vibration tactile signal during the collision process, and the pressure tactile signal during the collision process. The vehicle information of the own vehicle also includes body attitude, suspension state, tire state, power system state, and braking system state; Fuse the impact tactile signal with the basic tactile signal, the motion state tactile feedback signal, the vibration tactile signal, and the pressure tactile signal to obtain a comprehensive tactile signal stream; Based on a preset mechanical and neural signal conversion model, obtain the tactile information corresponding to the comprehensive tactile signal stream; The mechanical and neural signal conversion model is expressed as: Wherein, S(t1) represents the tactile information corresponding to the neural signal, F(t1) represents the comprehensive tactile signal stream corresponding to the physical force signal, α2 represents the weight coefficient simulating the human sensitivity to the magnitude of the force, β2 represents the weight coefficient simulating the human sensitivity to the change rate of the force, and γ1 represents the weight coefficient simulating the human sensitivity to the duration of the force.

6. The method for testing the automatic driving braking according to claim 5, which is based on the fusion of bionic vision and touch, is characterized in that The S3 includes: Based on a preset first deep feature extraction network, extract the original visual features, where the original visual features include image semantic information, region of interest, and first spatio-temporal information distribution features; Based on a preset second deep feature extraction network, extract the original tactile features, where the original tactile features include second spatio-temporal information distribution features; Based on the multi-head self-attention structure, align the original visual features and the original tactile features, and based on a preset bilinear pooling algorithm, perform an outer product operation on the aligned original visual features and original tactile features to obtain the fused visual-tactile joint features.

7. The method for testing the braking of an autonomous vehicle based on the fusion of bionic vision and touch according to claim 6, wherein, After aligning the original visual features and the original tactile features based on the multi-head self-attention structure, it further includes: Based on a preset contrast loss function, perform similarity learning on the aligned original visual features and original tactile features; The contrast loss function is expressed as: Among them, Y represents a binary label. When the aligned original visual feature and original tactile feature come from the same test scenario, Y is 0; otherwise, Y is 1; D w represents the feature distance, which is a parameter characterizing the difference between the aligned original visual feature and original tactile feature; m represents a set boundary value, and the feature distance between the original visual feature and original tactile feature from different test scenarios is at least greater than the boundary value.

8. A method for testing the braking of an autonomous vehicle based on the fusion of bionic vision and touch according to claim 1, characterized in that, The S4 includes: S41, based on a preset magic formula and the visual-tactile joint features, calculate the tire forces corresponding to each wheel when executing the candidate braking strategy, where the tire forces include longitudinal forces and lateral forces; S42, based on each of the tire forces and the vehicle external forces, calculate the overall force information of the current vehicle, where the vehicle external forces include air resistance and gravity; S43. Calculate the vehicle acceleration based on the overall force information, where the vehicle acceleration includes longitudinal acceleration, lateral acceleration, vertical acceleration, yaw acceleration, pitch acceleration, and roll acceleration; S44. Based on the vehicle acceleration, the current vehicle speed, the current position of the current vehicle, and a preset numerical integration algorithm, obtain the predicted speed and predicted position of the current vehicle at the next moment; and based on the predicted speed and predicted position, obtain the body response information corresponding to the candidate braking strategy; S45. Based on the body response information, perform a non-linear time-varying analysis to obtain the non-linear time-varying characteristics between the tires of the current vehicle and the ground, where the non-linear time-varying characteristics include non-linear friction coefficient, non-linear side-slip characteristics, and time-varying road adhesion; S46. If the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, adjust the candidate braking strategy; if the non-linear time-varying characteristics do not exceed the corresponding current safety constraint conditions, use the candidate braking strategy as the optimal braking strategy; S47. Use the adjusted candidate braking strategy as the new candidate braking strategy, and repeat the steps of S41 to S46 until the optimal braking strategy is obtained.

9. The automatic driving braking test method based on the fusion of bionic vision and touch according to claim 8, wherein, Before determining whether the non-linear time-varying characteristics exceed the corresponding current safety constraint conditions, it further includes: Based on the second uncertainty quantification data, correct the original safety constraint conditions to obtain the corrected robust constraint conditions, and use the robust constraint conditions as the current safety constraint conditions, where the second uncertainty quantification data is a statistical parameter representing the unpredictable deviation in the safety constraint process.

10. An automatic driving braking test device based on the fusion of bionic vision and touch, characterized in that, It includes: An environment acquisition module, configured to acquire the environment information in the current test scenario based on a vision device, where the environment information includes weather, light, road surface information, obstacle information, and self-vehicle information, the road surface information includes road surface friction coefficient, road surface flatness, road surface material, road surface slope, road surface curvature, road boundary, and lane line information, the obstacle information includes obstacle type and obstacle attribute information, and the self-vehicle information includes the current vehicle speed, body attitude, and body state; A tactile acquisition module, configured to simulate the collision process between the current vehicle and the obstacle, and acquire the tactile information corresponding to the collision process based on a force feedback device; A feature fusion module, configured to perform multi-modal feature fusion processing on the original visual features corresponding to the environment information and the original tactile features corresponding to the tactile information to obtain visual-tactile combined features; A prediction response module, configured to predict the body response information of the current vehicle corresponding to the candidate braking strategy based on the visual-tactile combined features, and obtain the optimal braking strategy corresponding to the candidate braking strategy based on the body response information, where the braking strategy is a parameter representing the braking control of the current vehicle in an emergency braking scenario, and the body response information includes the longitudinal dynamic characteristics, lateral dynamic characteristics, vertical dynamic characteristics, and stability index of the vehicle; A control signal generation module, configured to generate a braking control signal based on the optimal braking strategy.

Images