Method for dynamic detection of a vehicle

Abstract

The application discloses a kind of dynamic detection methods of vehicle, the present application relates to the technical field of dynamic detection, based on the compensation detection of this visual occlusion area to determine occlusion dynamic target;Based on the occlusion dynamic target and dynamic driving stereogram determines enhanced panoramic view;Mark the multiple key geometric features of unmanned vehicle, according to multiple key geometric features, the current driving data of unmanned vehicle and enhanced panoramic view constructs dynamic detection system, guarantees the dynamic detection of unmanned vehicle in driving process.According to the area content of each dynamic detection area, corresponding dynamic early warning level and the driving state of unmanned vehicle determine the early warning control logic of unmanned vehicle;Based on the relative swing angle and the current driving data of unmanned vehicle determine corresponding variable characteristics, according to the variable characteristics and the multistage reasoning of early warning control logic of unmanned vehicle determine corresponding multi-modal emergency measures, improve the accuracy of multi-modal emergency measures.

Description

Technical Field

[0001] This invention relates to the technical field of dynamic detection, and more particularly to a method for dynamic detection of vehicles. Background Technology

[0002] With the development of technology, unmanned vehicles are a type of vehicle that drive autonomously without human intervention. Autonomous driving technology is gradually being applied to heavy trucks. Current technologies largely rely on the direct detection of visible targets, typically generating static stereoscopic images from road surface images captured by cameras. However, this method often fails to effectively handle visually obstructed areas when faced with large objects, inclement weather (such as rain and fog), or complex road conditions. This affects the accuracy of the dynamic detection system of unmanned vehicles, resulting in lower accuracy of their multimodal emergency response measures. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method for dynamic detection of vehicles.

[0004] This invention provides a method for dynamic detection of a vehicle, comprising:

[0005] When an unmanned vehicle is driving on the road, multiple road images and point cloud data detected by LiDAR are combined and spatially aligned, and the driving data of the unmanned vehicle is combined to determine the dynamic driving stereo map of the unmanned vehicle.

[0006] Multimodal learning is performed on the dynamic driving stereo image and the driving text information of the unmanned vehicle, and the visual occlusion area of ​​the unmanned vehicle is marked. The occluded dynamic target is determined based on the compensation detection of the visual occlusion area.

[0007] An enhanced panoramic image is determined based on the occluded dynamic target and the dynamic driving stereo image; multiple key geometric features of the unmanned vehicle are marked, and a dynamic detection system is constructed based on the multiple key geometric features, the current driving data of the unmanned vehicle, and the enhanced panoramic image;

[0008] The dynamic detection system outputs multiple dynamic detection areas, and determines the warning and control logic of the unmanned vehicle based on the area content of each dynamic detection area, the corresponding dynamic warning level, and the driving status of the unmanned vehicle.

[0009] The relative sway angle between the tractor and trailer in the unmanned vehicle is marked. Based on the relative sway angle and the current driving data of the unmanned vehicle, the corresponding variable features are determined. Based on the variable features and the multi-level reasoning of the early warning and control logic of the unmanned vehicle, the corresponding multimodal emergency measures are determined.

[0010] This invention provides a vehicle dynamic detection system, which is applied to the above-described vehicle dynamic detection method.

[0011] Compared with the prior art, the beneficial effects of the present invention are:

[0012] (1) When the unmanned vehicle is driving on the road, multiple road images and point cloud data detected by LiDAR are combined and spatially aligned, and the dynamic driving stereo map of the unmanned vehicle is determined by combining the driving data of the unmanned vehicle; multimodal learning is performed on the dynamic driving stereo map and the driving text information of the unmanned vehicle, and the visual occlusion area of ​​the unmanned vehicle is marked. The occlusion dynamic target is determined based on the compensation detection of the visual occlusion area; the enhanced panoramic map is determined based on the occlusion dynamic target and the dynamic driving stereo map; multiple key geometric features of the unmanned vehicle are marked, and a dynamic detection system is constructed based on multiple key geometric features, the current driving data of the unmanned vehicle and the enhanced panoramic map. The visual occlusion area is introduced and the enhanced panoramic map is further controlled, which improves the accuracy of the dynamic detection system and ensures the dynamic detection of the unmanned vehicle during the driving process.

[0013] (2) The dynamic detection system outputs multiple dynamic detection areas. Based on the area content of each dynamic detection area, the corresponding dynamic warning level and the driving status of the unmanned vehicle, the warning control logic of the unmanned vehicle is determined. The relative swing angle between the tractor and the trailer in the unmanned vehicle is marked. Based on the relative swing angle and the current driving data of the unmanned vehicle, the corresponding variable features are determined. Based on the variable features and the multi-level reasoning of the warning control logic of the unmanned vehicle, the corresponding multimodal emergency measures are determined. The variable features are further controlled, and the multi-level reasoning of the variable features and the warning control logic of the unmanned vehicle is realized, which improves the accuracy of the multimodal emergency measures. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the dynamic detection method for vehicles according to an embodiment of the present invention.

[0015] Figure 2 This is a flowchart illustrating step S11 in the vehicle dynamic detection method of this embodiment of the invention.

[0016] Figure 3 This is a flowchart illustrating step S12 in the vehicle dynamic detection method of this embodiment of the invention.

[0017] Figure 4 This is a flowchart illustrating step S13 in the vehicle dynamic detection method of this embodiment of the invention.

[0018] Figure 5 This is a flowchart illustrating step S14 in the vehicle dynamic detection method of this embodiment of the invention.

[0019] Figure 6 This is a flowchart illustrating step S15 of the vehicle dynamic detection method in this embodiment of the invention.

[0020] Figure 7 This is a schematic diagram of the structural composition of the vehicle dynamic detection system in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] Please see Figures 1 to 7 A dynamic vehicle detection method, applied to dynamic detection scenarios; the dynamic vehicle detection method includes:

[0023] Step S11: The unmanned vehicle drives on the road. Multiple road images and point cloud data detected by LiDAR are combined and spatially aligned. The dynamic driving stereo map of the unmanned vehicle is determined by combining the driving data of the unmanned vehicle.

[0024] Step S12: Perform multimodal learning on the dynamic driving stereo image and the driving text information of the unmanned vehicle, and mark the visual occlusion area of ​​the unmanned vehicle. Based on the compensation detection of the visual occlusion area, determine the occluded dynamic target.

[0025] Step S13: Determine the enhanced panoramic image based on the occluded dynamic target and the dynamic driving stereo image; mark multiple key geometric features of the unmanned vehicle, and construct a dynamic detection system based on the multiple key geometric features, the current driving data of the unmanned vehicle, and the enhanced panoramic image;

[0026] Step S14: The dynamic detection system outputs multiple dynamic detection areas. Based on the area content of each dynamic detection area, the corresponding dynamic warning level, and the driving status of the unmanned vehicle, the warning and control logic of the unmanned vehicle is determined.

[0027] Step S15: Mark the relative sway angle between the tractor and trailer in the unmanned vehicle, determine the corresponding variable features based on the relative sway angle and the current driving data of the unmanned vehicle, and determine the corresponding multimodal emergency measures based on the variable features and the multi-level reasoning of the early warning and control logic of the unmanned vehicle.

[0028] refer to Figure 2 In step S11, the specific steps are as follows:

[0029] S111: Real-time monitoring of the unmanned vehicle's road driving process, and based on multiple cameras of the unmanned vehicle, multiple road images at different positions of the unmanned vehicle are collected. The pixel coordinate system of multiple road images is transformed to a unified LiDAR vehicle coordinate system. In this LiDAR vehicle coordinate system, multiple road images and point cloud data detected by LiDAR are combined and aligned in the same spatial dimension, and an RGB-D fusion map is generated. This RGB-D fusion map contains color texture and depth information.

[0030] S112: Real-time access to the chassis CAN bus data of the unmanned vehicle, extract multiple driving data of the unmanned vehicle, use multiple driving data to perform frame-by-frame motion distortion correction on the RGB-D fusion image, and output the corrected primary driving stereo image. Align the corrected primary driving stereo image with the vehicle status information with timestamps in the time dimension to construct the dynamic driving stereo image of the unmanned vehicle.

[0031] In the embodiments of this application, the road driving process of the unmanned vehicle is monitored in real time, and multiple road images at different positions of the unmanned vehicle are collected based on multiple cameras of the unmanned vehicle. The pixel coordinate system of the multiple road images is transformed to a unified LiDAR vehicle coordinate system. In the LiDAR vehicle coordinate system, the multiple road images and the point cloud data detected by the LiDAR are aligned in the same spatial dimension and an RGB-D fusion map is generated. The RGB-D fusion map contains color texture and depth information.

[0032] At this point, the unmanned vehicle is equipped with a multi-camera network covering 360 degrees, including forward-looking telephoto, forward-looking wide-angle, side-view, and rear-view cameras. The system uses GMSL (Gigabit Multimedia Serial Link) or Ethernet to collect "multiple road surface images" from different locations in real time at a high frame rate (such as 30FPS). At the same time, the system performs real-time synchronization processing on the image stream to ensure that within a microsecond time error, each camera captures photon information at the same moment, laying the timing foundation for subsequent fusion.

[0033] The camera outputs a 2D pixel coordinate system (unit: pixels) based on optical imaging principles, while the lidar outputs a 3D Cartesian coordinate system (unit: meters) based on physical ranging. The system utilizes pre-obtained camera intrinsic parameters (focal length, principal point) and extrinsic parameters (rotation matrix R, translation vector t) acquired through Zhang's calibration method to construct a rigid body transformation model. Through perspective projection equations, each pixel (u, v) in all camera images is inversely projected or mapped to a "lidar vehicle coordinate system" with the lidar center as the origin, achieving a geometric transition from the "image plane" to "vehicle space." Optionally, the lidar is a hybrid solid-state lidar, with multiple laser emitters arranged vertically, performing a 360° scan horizontally through the rotation of prisms or mirrors.

[0034] In a unified LiDAR vehicle coordinate system, the LiDAR point cloud is sparse (discrete 3D points), while the image data is dense (continuous pixel blocks). Algorithms typically employ voxel-based or point-based feature alignment strategies to fill the spatial structure of the point cloud with the semantic information of the image. The system corrects for parallax caused by different sensor installation positions, ensuring that the edges of objects identified in the image and the outlines of objects depicted in the point cloud are highly coincident in spatial position, completing the "alignment" operation, eliminating misalignment caused by perspective errors, and finally generating an information-rich "RGB-D fusion map". Here, RGB represents color texture information, which comes from the high-resolution camera; D represents depth information, which comes from the high-precision LiDAR. The system directly maps the RGB texture attributes of the image to the corresponding 3D point cloud data, so that each spatial point not only has geometric coordinates (x, y, z) but also color features (R, G, B). This fusion map retains both the accuracy of LiDAR in measuring distant objects and the semantic recognition capabilities of the camera for object materials, colors, and text (such as traffic lights and lane lines).

[0035] Specifically, the 360-degree multi-camera network (forward telephoto, forward wide-angle, side-view, and rear-view) of the unmanned vehicle is fully activated. As an unmanned heavy truck, the unmanned vehicle is traveling at 80 km / h in the rainy night, with time running out. The forward telephoto camera captures the glare area caused by oncoming headlights within a 200-meter range ahead, while the side-view camera is scanning the blind spot in the right lane. All cameras transmit raw image data to the central computing unit at a rate of 30 FPS via the GMSL link. The system uses hardware-level trigger signals to lock the scanning time of the LiDAR and the exposure time of the cameras within a microsecond error range. This means that every image acquired by the system at time T corresponds strictly to the spatial slice scanned by the LiDAR at time T, eliminating the risk of "ghosting" caused by image lag or advance.

[0036] The forward-looking wide-angle camera captured two bright red spots (the brake lights of the vehicle in front) in the upper left region of the image plane (pixel coordinates u=1200, v=300); however, these are only two-dimensional pixels, and the system does not know how far these two red spots are from the autonomous vehicle. The system calls the pre-stored Zhang calibration parameters to construct a rigid body transformation model. Through the inverse operation of the perspective projection equation, the system begins to search for the corresponding point in the "LiDAR vehicle coordinate system". Combined with the data from the forward-looking telephoto camera, the system successfully mapped the red spots on the image plane to their physical location in the vehicle coordinate system (X=+55m, Y=-0.5m, Z=0.8m), transforming the abstract pixels into a concrete physical space description.

[0037] The point cloud data from the LiDAR is sparse in this area. It may only scan a few discrete points on the bumper and roof of the vehicle cutting in (the vehicle in front), while missing the middle part of the vehicle body. At the same time, since the camera is installed on the top of the driver's cab and the LiDAR is installed below the bumper, there is a significant vertical parallax between the two. As a result, if not corrected, the outline of the vehicle body in the image will "float" above the point cloud. A voxel-based feature alignment strategy is adopted. The system divides the LiDAR point cloud into a three-dimensional voxel grid and fills the pixel features in the image into the corresponding voxels.

[0038] The system uses calibrated extrinsic parameters (rotation matrix R and translation vector t) to perform parallax correction, forcibly aligning the "car outline" identified in the image with the sparse "car top cloud" in the point cloud in spatial position; the algorithm eliminates false point clouds caused by raindrops (these point clouds cannot correspond to any entity objects in the image), and finally eliminates perspective error, ensuring that the geometric outline of the vehicle in the virtual 3D model perfectly matches the visual edge.

[0039] The system needs to comprehensively assess the status of the vehicle cutting in front, including not only the distance but also its type and behavioral intent. Geometric information (D): Using millimeter-level precision depth data provided by LiDAR, the relative distance between the cutting vehicle and the autonomous vehicle is accurately calculated to be 45 meters, and is rapidly decreasing. Semantic information (RGB): The system directly "attaches" the high-resolution RGB texture captured by the camera onto the 3D skeleton constructed from the point cloud. This instantly gives the originally blurry target, which was just a collection of point clouds, the red body color, the details of the illuminated brake lights, and the unique texture of a sedan, ultimately outputting an "RGB-D fusion map". In the RGB-D fusion map, the autonomous vehicle clearly "sees" a sedan ahead (RGB information), its brake lights are on (RGB semantics), and it is cutting into the adjacent lane at a distance of 45 meters from the vehicle (D depth information). This fusion map retains the ranging accuracy of LiDAR, which is not affected by light, and also uses the camera to identify the key semantic behavioral feature of "brake lights on".

[0040] Furthermore, the chassis CAN bus data of the unmanned vehicle is accessed in real time, and multiple driving data of the unmanned vehicle are extracted. The motion distortion of the RGB-D fusion image is corrected frame by frame using multiple driving data, and the corrected primary driving stereo image is output. The corrected primary driving stereo image is aligned with the vehicle status information with timestamps in the time dimension to construct the dynamic driving stereo image of the unmanned vehicle.

[0041] At this time, the system accesses the chassis CAN bus in real time through the vehicle gateway (usually using CANFD or vehicle Ethernet protocol) and parses the high-frequency signal according to the preset DBC (DatabaseCAN) file; it focuses on extracting "multiple driving data", including: wheel speed pulses of the four wheels, steering wheel angle, master cylinder pressure, longitudinal / lateral acceleration, yaw rate and displacement sensor data of the suspension system. These data reflect the real physical state of the vehicle under inertia and external forces, and provide kinematic parameter input for subsequent distortion correction.

[0042] Within the time it takes for a lidar and camera to scan a circle (e.g., 100ms), an unmanned vehicle may have moved several meters and its posture may have changed. If point clouds are directly stitched together, straight guardrails will become curved lines, and stationary objects will produce motion blur.

[0043] The system utilizes the extracted driving data and combines it with the vehicle kinematics model (such as a uniform speed or uniform acceleration model) to calculate the pose transformation matrix (rotation matrix and translation vector) of each frame of image or each LiDAR point at the time of emission relative to the start of the scan. Through inverse motion compensation, all data points are "back-projected" or "distorted" to the same reference coordinate system (usually the coordinate system at the center of the scan), eliminating the distortion of environmental data caused by the vehicle's own motion and outputting a "primary driving 3D map" with accurate geometry.

[0044] Due to sensor processing delays and transmission jitter, the time point at which the RGB-D fusion map (environmental data) is generated is often inconsistent with the time point at which the CAN data (vehicle status data) arrives. Therefore, the system introduces a globally unified time reference (such as PTP protocol or GPS timing) to strictly align the corrected primary driving stereo map with vehicle status information with precise timestamps (such as "the yaw rate of the unmanned vehicle at time T is 5deg / s"). The resulting "dynamic driving stereo map" not only contains high-precision environmental geometry information after distortion correction, but also perfectly matches the vehicle's own dynamic state at the moment of environmental perception, forming a four-dimensional (3D space + time) data field that includes the coupling relationship between the environment and the vehicle.

[0045] Specifically, when the unmanned vehicle is driving on a slippery road surface, the coefficient of friction between the tires and the road surface is at a critical state, and the chassis sensors are continuously outputting data reflecting the vehicle's dynamics. The system reads key signals defined in the DBC file at high frequency through the CANFD protocol. Wheel speed pulse: The system reads the wheel speed of the four wheels and calculates the current reference speed of the unmanned vehicle as 80km / h (22.2m / s). Steering wheel angle: The system detects that the steering wheel is in a state of slight correction (angle +2 degrees), and is not driving in a straight line. Inertial data: The IMU (Inertial Measurement Unit) shows that the longitudinal acceleration is 0, but the yaw rate fluctuates slightly (2deg / s) due to the uneven road surface. Suspension displacement: Due to the full load of 49 tons, the suspension displacement sensor shows that the vehicle height is lowered by about 80mm compared to the unloaded state, and the center of gravity position has changed. These data reflect the true kinematic state of the unmanned vehicle at time T, providing a crucial "motion reference" for subsequent distortion correction.

[0046] The system utilizes the acquired wheel speed and yaw rate, inputting them into the vehicle's kinematics model. It calculates that between the scan start time Tstart and the scan end time Tend, the unmanned vehicle undergoes a translation of 2.22 meters along the X-axis (longitudinal direction) and a slight rotation of 0.1 degrees around the Z-axis (vertical axis). For each point cloud data point acquired by the LiDAR, the system adds a precise timestamp. Based on this timestamp, the algorithm calculates the unmanned vehicle's pose offset relative to the scan center at that moment. Through matrix transformation, the system "reverse-translates" the point cloud data at the end of the scan by 2.22 meters and "corrects" the offset caused by the vehicle's rotation. The road guardrails that were originally curved and deformed in the point cloud are "straightened," and the outline of the car cutting in front is restored to its true physical proportions, eliminating any ghosting. The system outputs a geometrically accurate "primary driving 3D map."

[0047] The system incorporates a GPS / PTP timing system to align the timestamps of all data sources to a unified global clock. The algorithm identifies a 50ms delay in the environmental data. To match this instant, the system uses a kinematic model to extrapolate the yaw rate (e.g., 5deg / s) and acceleration collected at time T to calculate the predicted state 50ms later. The system deeply binds the distortion-corrected RGB-D environmental data with the extrapolated vehicle state data, generating a "dynamic driving 3D map." In the dynamic driving 3D map, not only is the position and texture of the car cutting in front clearly displayed (environmental information), but the actual state of the unmanned vehicle at the instant it perceives this scene is also accurately marked: "fully loaded, low center of gravity, yaw rate is rapidly increasing, and it is trending towards a left turn" (vehicle state). This four-dimensional data field perfectly reproduces the spatiotemporal coupling relationship between the unmanned vehicle and the environment at time T+50ms.

[0048] refer to Figure 3 In step S12, the specific steps are as follows:

[0049] S121: Collect driving text information of unmanned vehicles, introduce a multimodal learning mechanism into the dynamic driving stereo map and driving text information of unmanned vehicles, and analyze the road scene of unmanned vehicles through cross-modal semantic association to mark the visual occlusion areas that cannot be directly observed by visual sensors.

[0050] S122: For visually obstructed areas, the laser radar beam penetration characteristics are used to extract point cloud echoes behind the obstruction. Combined with the temporal information of the dynamic driving stereo map, Kalman filtering is used to track the motion trend around the obstructed area to determine multiple dynamic features. Based on the characteristic shape, relative position and corresponding point cloud echo of the multiple dynamic features, the obstructed dynamic target is determined, and the specific position and motion state of the obstructed dynamic target in the obstructed state are marked.

[0051] In the embodiments of this application, driving text information of unmanned vehicles is collected, a multimodal learning mechanism is introduced into the dynamic driving stereo map and the driving text information of unmanned vehicles, and the road scene of unmanned vehicles is analyzed through cross-modal semantic association to mark the visual occlusion areas that cannot be directly observed by the visual sensors.

[0052] At this point, the system not only collects visual images, but also extracts road sign text (such as "speed limit 80", "exit merging", "construction ahead") through the OCR (Optical Character Recognition) module. At the same time, it extracts the semantic attributes of the current road segment (such as "sharp bend", "tunnel", "toll station") from high-precision maps and navigation systems. This text information is encoded into natural language vectors or structured labels, which serve as the "text modality" input for multimodal learning, providing prior knowledge for subsequent logical reasoning.

[0053] The system employs a Transformer-based multimodal encoder to extract features from the "dynamic driving stereo image" (visual modality) from S11 and the acquired "driving text information" (text modality). Through a projection layer, the visual feature vectors and text feature vectors are mapped to the same high-dimensional latent feature space. In this shared space, the visual features representing "red light" and the text features representing "stop signal" are geometrically close, thus achieving cross-modal semantic alignment.

[0054] The system utilizes a cross-modal attention mechanism to calculate the correlation weight between visual regions and text semantics. For example, when the text features show "merging area" and the visual features detect "lane line interruption", a strong correlation is formed between the two. Through this correlation, the system deeply analyzes the semantic logic of the road scene, infers the implicit rules or potential risk points of the current scene, and identifies those areas that, although not visually presented, should exist semantically.

[0055] Based on the above reasoning results, the system performs reverse marking in the dynamic driving stereo map; for areas that cannot be observed by visual sensors (cameras) due to physical obstruction, but are determined to be high-risk or key interactions through semantic association, the system marks them as "visual occlusion areas". This marking is not based on pixels, but on a mask based on probability and logic. It clearly indicates to the subsequent perception modules that there is a very high probability that there are unobserved dynamic targets in this area, and targeted compensation detection must be performed (such as calling radar penetration or performing trajectory prediction).

[0056] Specifically, because the autonomous vehicle is in a nighttime rainstorm environment, the visual sensors are physically limited (such as being obscured by water mist or insufficient light), and cannot fully understand complex traffic rules and potential risks based solely on geometric images; at this time, the autonomous vehicle is driving on the highway, the forward-facing camera captures the roadside signs, and the onboard navigation system locates the current road segment attributes.

[0057] The system's OCR module identifies the text "Construction Ahead" and "Change Lane to Left" on the roadside reflective signs in the enhanced image, as well as "Speed ​​Limit 60" (original speed limit 100, speed reduction due to construction) on the auxiliary signs; the high-precision map interface returns the attribute labels of the current road segment: "Construction Section", "Number of Lanes Reduced (4 to 2)", and "Slippery Road Surface"; this text information is converted into natural language vectors (such as "Construction Ahead, Change Lane to Left"), which serve as the system's text modal input, indicating that the road structure ahead of the unmanned vehicle will undergo a sudden change;

[0058] The multimodal encoder extracts visual features (such as the conical geometric features of the cones and the yellow reflective texture) and textual features ("construction" and "lane change") from the "dynamic driving stereo image". In the shared feature space, the system finds that the visual "cone array" and the textual semantic "construction area" are highly overlapping in geometric distance. At the same time, the visual "blurred / worn left lane line" is strongly associated with the textual semantic "change lane to the left". This step allows the autonomous vehicle not only to know that there are obstacles, but also to understand the semantic logic of "road closure for construction", which means that the traffic rule of "changing lane to the left" must be followed.

[0059] The system calculates the correlation between visual regions and text semantics; the text prompts "construction ahead" and "lane narrowing" usually mean that traffic will merge and become congested; when the visual system detects a faint, highly reflective target (the merging car) in the blurred area on the right, the text semantic "construction merging" significantly enhances the attention weight of that area; the system infers that: due to construction ahead, vehicles in the right lane (such as this car) are highly likely to urgently change lanes to the left (the lane where the unmanned vehicle is located) to avoid the road closure. Although this behavior has just appeared visually, it is a high-probability event that is bound to happen in semantic logic.

[0060] In the dynamic driving 3D map, the system performs reverse marking on the blind spot of the A-pillar on the right side of the unmanned vehicle and the "rear area" obscured by the car. The system marks these areas as "high-risk visual occlusion areas," which is a probability-based mask. Although they are not visible now (because they are obscured by the car in front or the A-pillar), according to the semantic logic of "construction merging" and the vehicle's movement trend, there is a high probability that a following car or a smaller obstacle (such as a motorcycle) is cutting in from the back in this area. The system outputs a scene analysis map rich in semantic understanding. The scene analysis map clearly indicates the "construction attribute" in front and the "high-risk existence" of the right blind spot. This forces the dynamic detection system in S13 to focus on trajectory prediction and anti-shake processing for these areas that are not directly observed visually, and cannot simply treat them as empty ground.

[0061] Furthermore, for visually obstructed areas, the line penetration characteristics of LiDAR are utilized to extract point cloud echoes behind the obstruction. Combined with the temporal information of the dynamic driving stereo map, Kalman filtering is used to track the motion trend around the obstructed area to determine multiple dynamic features. Based on the characteristic shape, relative position, and corresponding point cloud echo of multiple dynamic features, the obstructed dynamic target is determined, and the specific position and motion state of the obstructed dynamic target under the obstruction state are marked. This comprehensive consideration of the characteristic shape, relative position, and corresponding point cloud echo of multiple dynamic features ensures the accuracy of obstructing dynamic targets.

[0062] At this point, the laser beam emitted by the lidar has a certain penetrating ability, especially for glass, sparse vegetation, and gaps under the vehicle in front. For the "visual occlusion area" marked by S121, the system no longer only focuses on the first echo signal (i.e., the first reflected signal, usually the surface of the occlusion object), but focuses on analyzing the second or last echo signal. Through signal processing algorithms, it filters out the strong reflection point cloud belonging to the vehicle in front (occlusion object), and extracts and retains those weaker "point cloud echoes" from behind the occlusion object. Although these point clouds are sparse, they carry the geometric information of the occluded area.

[0063] The system incorporates the temporal information of the dynamic driving 3D map and uses extended Kalman filter (EKF) or unscented Kalman filter (UKF) to track potential targets within the occluded area. The algorithm establishes a state vector (including position, velocity, acceleration, etc.) and uses a motion model (such as a uniform velocity or uniform acceleration model) to predict the possible position of the target at the current moment. The extracted sparse point cloud is correlated with the predicted trajectory through the observation equation to continuously correct the prediction error. This process can smooth noise and accurately lock the "motion trend" of targets around the occluded area.

[0064] Based on the established trajectory, the system further analyzes the essential attributes of the target. The system extracts "multiple dynamic features" from three dimensions: characteristic morphology: analyzes the geometric dimensions and aspect ratio after point cloud clustering to determine whether it is a low-profile shape similar to a car or a small shape similar to a pedestrian; relative position: calculates the relative distance and angle relationship between the target and obstructions (such as the vehicle in front) to determine whether it is overtaking or being followed; point cloud echo characteristics: analyzes the reflection intensity of the echo to distinguish between metallic objects, non-metallic objects, or road clutter. These features constitute a high-dimensional feature vector used to describe the target's "identity information".

[0065] Based on the extracted dynamic feature vectors, the system uses a classifier (such as a support vector machine (SVM) or a neural network) to match the target category and formally determine the "occluded dynamic target" (such as "vehicle cutting in front" or "pedestrian appearing out of nowhere"). The system maps the optimal state value estimated by Kalman filtering to the global coordinate system, accurately marking the "specific location" (such as 3.5 meters behind the left of the vehicle in front) and "motion state" (such as longitudinal speed of 15 m / s and lateral speed to the left of 1.5 m / s) of the target in the occluded state. This completes the digital reconstruction of an "invisible" target.

[0066] Specifically, the sedan had cut in front of the unmanned vehicle on its right front; visually (S121), the sedan's large body blocked the lane that was originally on its right, creating a blind spot; in addition, the rain curtain caused by the heavy rain also created partial obstruction; when the laser beam of the lidar shone on the metal surface of the sedan, it generated an extremely strong first echo (sedan outline); however, some of the laser beam passed through the gaps under the sedan or the front and rear windshields (non-metallic parts), illuminating objects further behind.

[0067] S122 targets the occluded area marked by S121, suppressing the first echo signal and focusing on analyzing the second and final echoes. The system filters out strong reflective point clouds belonging to the car body and extracts a batch of weaker but present point cloud data. To the right rear of the car (within the visual blind spot), the system sparsely "emerges" some point clouds, which belong to a small truck that is being followed by the car and is being occluded by the car. Without this step, the autonomous vehicle would have no idea that there is a car behind the car.

[0068] The system uses an extended Kalman filter (EKF) to establish a state vector. Based on historical frame data, the algorithm predicts the location where the occluded truck should appear in the next moment. The system matches the extracted sparse penetrating point cloud with the predicted trajectory. When the car moves to the left and reveals a gap, several high-confidence point clouds are captured instantly. By continuously correcting the prediction error, the system finds that although the point cloud is sparse, its trajectory shows that the target is accelerating forward and does not follow the car to change lanes to the left, but stays in the right lane and goes straight. This accurately locks the target's "movement trend".

[0069] The system clusters sparse point clouds and analyzes their geometric dimensions; with an aspect ratio of approximately 1.5 and a height of approximately 1.8 meters, it is determined to be a van-like shape similar to a small truck. The relationship between the target and the obstruction (car) is calculated; the target is located 3 meters to the right rear of the car and is being followed. The point cloud intensity is analyzed, and the echo intensity is found to be high and has metallic reflection characteristics, ruling out the possibility of rain clutter or plastic traffic cones. A high-dimensional feature vector is generated, describing the target as "a moving object with a certain metallic volume located in the right lane, closely following the car in front".

[0070] Specifically, the system utilizes the prediction update mechanism of the Kalman filter to solve the problem of inaccurate localization of sparse point clouds. Based on the target's historical data from the previous few frames, a state vector containing position, velocity, and acceleration is established, and a motion model (such as a uniform velocity or uniform acceleration model) is used to predict the target's theoretically expected spatial position at the current moment. When new sparse point cloud data arrives, the system does not simply treat the point cloud coordinates as the target position, but rather treats them as observations. These observations are then weighted and matched with the predicted values ​​in the filter's "observation equation." Since the predicted trajectory has already locked the approximate search range of the target, the system only needs to verify the existence of the sparse point cloud within a very small neighborhood (confidence interval) near the predicted position. This "prediction-guided observation" approach greatly reduces the search space, allowing even the few reflection points that fall within the error ellipsoid of the predicted trajectory to be associated with the target with high confidence. By continuously correcting the prediction residuals, the system can solve for the most accurate three-dimensional coordinates of the target at the current moment, achieving convergence from "fuzzy observation" to "precise position."

[0071] For sparse point clouds, although their spatial distribution is discrete, their motion trajectory over time is continuous and conforms to physical laws. The system analyzes the displacement changes of sparse point clouds in consecutive frames, calculates optical flow or displacement vectors, and thus solves for the target's velocity vector and heading angle. For example, even if the number of point clouds is insufficient to piece together a complete vehicle body, as long as its echo center shows a stable longitudinal displacement and almost zero lateral displacement in consecutive frames, the system can determine that the target is in a "straight-ahead" state rather than a "lane-changing" state. At the same time, the system combines the echo intensity characteristics of the lidar to distinguish between real targets and environmental noise. The echo intensity of real occluded targets is significantly higher than that of the surrounding background (such as rainwater or road surface) and has specific material reflection characteristics. By double-checking this intensity characteristic with the motion trend, the system can effectively eliminate false clutter points and ensure that the data source used to calculate the motion state is real and reliable. The system uses Support Vector Machine (SVM) to classify the feature vectors, officially identifying the target as "an occluded dynamic target - a small truck". The system maps the optimal state value estimated by EKF to the global coordinate system. The system accurately marks the small truck 45 meters to the right front of the unmanned vehicle, with a longitudinal speed of 20 m / s (slightly faster than the unmanned vehicle) and a lateral speed of 0 (not changing lanes). Although the unmanned vehicle cannot see the small truck visually (in the camera view) (it is blocked by the car), the small truck has been digitally reconstructed in the output of S122. This is crucial for subsequent risk avoidance decisions - the unmanned vehicle knows: "Although I can avoid the car in front by changing lanes to the left, I must pay attention to the space on the left because there is a fast car following closely behind the cutting vehicle on the right. I cannot blindly turn the steering wheel to the right."

[0072] At this moment, the sedan cut in from the right of the driverless vehicle and drove in the lane in front of the driverless vehicle on the right; the obscured "minivan" was right behind the red sedan; although the lidar was installed under the bumper and had the advantage of a low viewpoint, the car in front (the sedan) acted as an obstruction and was located right between the lidar and the minivan; for the lidar installed at the front of the car, the large body of the red sedan constituted a huge "visual barrier", completely blocking its line of sight to detect the minivan behind.

[0073] Although the small truck is objectively located in the right lane of the road, it is completely blocked by the vehicle in front and is "invisible" in the sensor field of view of the unmanned vehicle. At the same time, for a fully loaded unmanned heavy truck like the unmanned vehicle, when the front of the truck makes a sharp left turn to avoid the vehicle in front, due to the huge inertia, the trailer (rear) will not immediately follow the front of the truck to the left. Instead, it will be violently swung to the right due to centrifugal force (i.e., "turning over" or "Jackknife" tendency).

[0074] This rightward swing of the trailer causes the vehicle's physical footprint to expand instantaneously to the right. At this moment, the right rear corner of the trailer sweeps into the space that originally belonged to the right lane. Since the minivan is located directly behind the vehicle in front (i.e., in front of the right lane) and is traveling at a relatively high speed (20m / s), if the unmanned vehicle blindly changes lanes sharply to the left, causing the trailer to swing violently to the right, it is highly likely that the right rear side of the trailer will scrape against or even collide with the minivan traveling straight in the right lane. Therefore, the purpose of S122 reconstructing this minivan is not to turn the steering wheel to the right to avoid it, but to limit the magnitude and speed of the left lane change. The system knows that there is a fast vehicle on the right and must control the amount of the trailer's tail swing to prevent its tail from swinging onto the minivan when avoiding the vehicle in front.

[0075] The sedan, positioned to the right front, obstructed the view of the small truck, creating a blind spot. The S122 system used lidar to penetrate the underside or gaps of the sedan, detecting and locking onto the small truck within the blind spot. When swerving to the left to avoid the sedan, the system, based on the reconstructed position of the small truck, identified a risk of a "trailer skidding" collision on the right. Therefore, instead of blindly counter-steering to the right, the system implemented "constrained avoidance"—that is, while swerving to the left, it used differential braking and other means to suppress the trailer's rightward sway, ensuring that the front of the sedan avoided the sedan while the rear did not scrape the small truck on the right. By fully considering multi-sensor fusion, occlusion penetration detection, and the special dynamic constraints of large vehicles, the system perfectly solved the safety avoidance problem in complex scenarios.

[0076] refer to Figure 4 In step S13, the specific steps are as follows:

[0077] S131: Based on the detection of the occluded dynamic target, the position coordinates and attribute information of the occluded dynamic target are determined, and the position coordinates and attribute information of the occluded dynamic target are loaded into the dynamic driving stereoscopic image. At this time, through semantic alignment and edge smoothing processing, the information gaps in the dynamic driving stereoscopic image are filled, and a corresponding enhanced panoramic image is generated. The enhanced panoramic image includes the visible target and the occluded dynamic target.

[0078] S132: Collect the overall shape of the unmanned vehicle, and determine multiple key geometric features based on the overall shape of the unmanned vehicle, the corresponding driving data and attitude data. The multiple key geometric features cover the steering wheel difference trajectory, turning radius and outer contour sway value. Based on the multiple key geometric features, construct the physical constraint layer of the unmanned vehicle during the driving process.

[0079] S133: The physical constraint layer is combined with the current driving data of the unmanned vehicle and the enhanced panoramic image. During the combination process, multiple dynamic detection items are constructed. Based on the multi-level training of each dynamic detection item, the dynamic detection system of the unmanned vehicle is determined. This dynamic detection system constructs a real-time digital twin model and dynamically presents the collision risk between the vehicle body and obstacles in the enhanced panoramic image.

[0080] In the embodiments of this application, the position coordinates and attribute information of the occluded dynamic target are determined based on the detection of the occluded dynamic target, and the position coordinates and attribute information of the occluded dynamic target are loaded into the dynamic driving stereoscopic image. At this time, the information gaps in the dynamic driving stereoscopic image are filled through semantic alignment and edge smoothing processing, and a corresponding enhanced panoramic image is generated. The enhanced panoramic image includes both visible targets and occluded dynamic targets.

[0081] At this point, the system locks onto the "occluded dynamic target" determined in S12 and extracts its state estimates: precise 3D position coordinates (based on the vehicle coordinate system), geometric dimensions (length, width, and height), category labels (such as cars, pedestrians, and non-motorized vehicles), and motion vectors (velocity and heading angle). The system packages these structured attribute information and, through a spatiotemporal synchronization mechanism, "rigidly writes" them as virtual entities into the corresponding grids or voxels of the "dynamic driving 3D map" generated in S11, giving them the same status as directly observed entities at the data level.

[0082] By utilizing a contextual semantic network, the texture and geometric features of the environment surrounding the occluded target (such as road surface texture and lane line orientation) are analyzed. The pose and features of the occluded target are adjusted to conform to physical laws (e.g., ensuring that the wheels are close to the ground rather than suspended in the air). An image inpainting algorithm based on partial differential equations (PDE) or a generative filling network is used to feather the edges of the occluded target and fill in pixel holes caused by coordinate transformation or occlusion. This step eliminates compositing traces, allowing the newly added target to naturally "integrate" into the original image both visually and geometrically, thus eliminating data discontinuities.

[0083] After the above processing, the system generates an "enhanced panoramic image" containing all the information. This image is no longer a single-view observation, but a dense environmental map containing multiple information: visible targets: vehicles, pedestrians, guardrails, etc. directly observed by cameras and lidar; occluded dynamic targets: dynamic entities that were originally occluded, which have been reconstructed and verified by algorithms.

[0084] This image not only contains rich RGB-D information (color and depth), but also semantic labels and motion states, providing complete, continuous and error-free input data for building the physical constraint layer in S132.

[0085] Specifically, S122 has determined that there is a "small truck" to the right rear of the car, but in the original point cloud and image generated by S11, this location only shows the rear of the car, with the area behind it empty or cluttered. The system extracts the state estimate of the small truck output by S122: 3D position (45 meters to the right front of the unmanned vehicle, with a lateral offset of 3.5 meters), geometric dimensions (5.2 meters long, 2.0 meters wide, and 2.2 meters high), category label (Van / Small Truck), and motion vector (longitudinal velocity 20 m / s, heading angle 0 degrees).

[0086] Through a spatiotemporal synchronization mechanism, the system packages these attributes; in the 3D voxel grid of the dynamic driving stereo map, the system forcibly activates the grid at the corresponding position; although there is no data visually, the system gives these voxels the same confidence level as the real LiDAR point cloud at the data level; in the digital space, the area behind the car is no longer an "unknown area", but is filled with this small truck that has been "rigidly written".

[0087] The contextual semantic network analyzed the texture of the road surface in the area (inferred to be asphalt) and the direction of the lane lines (slightly curved); the system applied physical constraints to force the bottom geometric center of the small truck to be "attached" to the road surface, eliminating the "suspended" phenomenon caused by radar scattering or elevation angle error, and ensuring that the wheels are in close contact with the ground.

[0088] To address the data gaps at the boundary between the target and the original environment (such as the edge of a car), the system employs an image inpainting algorithm based on partial differential equations (PDEs). The edge pixels of the minivan are feathered, and filling pixels are generated based on the hue of the surrounding rainy night (dark and high contrast) to fill the point cloud holes caused by occlusion. The minivan looks very natural in terms of image and geometry, as if it were originally "seen" by the sensor.

[0089] The system retains the car directly seen by the camera, the road signs in the distance, and the reflection of water on the road as "visible targets"; it seamlessly integrates the reconstructed and verified small truck as "occluded dynamic targets"; each voxel in the map not only contains RGB (color, from visual restoration) and D (depth, from radar estimation), but also is bound to a semantic label ("small truck") and motion state ("high-speed straight-through"); the system generates an "enhanced panoramic map"; in the enhanced panoramic map, the unmanned vehicle can clearly perceive a complete traffic flow situation: a car cutting in front (hazard source), and a small truck following closely behind the car on its right rear without any intention to change lanes (potential collision risk, because if the unmanned vehicle brakes suddenly, the small truck may rear-end the car and affect the unmanned vehicle). This map without blind spots provides complete input for the physical constraint layer in step S132, enabling the system to perform high-precision trajectory simulation.

[0090] Furthermore, the overall shape of the unmanned vehicle is collected, and several key geometric features are determined based on the overall shape of the unmanned vehicle, the corresponding driving data and attitude data. These key geometric features cover the steering wheel difference trajectory, turning radius and outer contour sway value. Based on these key geometric features, a physical constraint layer for the unmanned vehicle during driving is constructed, which takes into account the overall shape of the unmanned vehicle, the corresponding driving data and attitude data, and ensures the accuracy of the key geometric features.

[0091] At this point, the system collects the "overall shape" parameters of the unmanned vehicle, including static factory parameters such as the length of the tractor, trailer length, wheelbase, track width, articulation point position, and the maximum external dimensions of the vehicle. Simultaneously, it accesses the vehicle's "driving data" (such as steering wheel angle, wheel speed, and throttle opening) and "attitude data" (such as yaw rate, pitch angle, roll angle, and the relative articulation angle between the tractor and trailer) in real time. Through coordinate transformation, the system maps these dynamic data onto the vehicle's static geometric model, forming a dynamic vehicle body model that deforms in real time with the motion state.

[0092] The key calculations focus on three core indicators that determine collision risk: Steering wheel difference trajectory (inner wheel difference): This calculates the deviation between the center of the front wheel's turning angle and the center of the rear wheel's (or trailer's) trajectory. Especially for heavy trucks, during low-speed, large-angle turns, the rear wheel trajectory may cut into the inner side of the front wheel's trajectory, forming a "death crescent" area. Turning radius: Based on the current steering wheel angle, vehicle speed, and road adhesion coefficient, the instantaneous radius of curvature of the vehicle's center of gravity is calculated, along with the turning centers of the tractor and trailer, to deduce the fan-shaped path swept by the vehicle. Outer profile sway value: For articulated trailers, this calculates the lateral offset (fishtail amplitude) of the trailer relative to the tractor during lane changes or turns, as well as the vertical outer profile expansion caused by body roll.

[0093] The calculated geometric features are transformed into a "occupancy field" in the spatiotemporal domain. The system uses the calculated features to construct a set of 3D polyhedra surrounding the unmanned vehicle, namely the "physical constraint layer". This constraint layer is not only the current static bounding box of the vehicle, but also includes the spatial volume that the vehicle may occupy in the next few seconds due to movement (such as tail-swing or inner wheel difference cutting). It defines the physical boundaries that the vehicle cannot cross, including both collision avoidance constraints on external obstacles and dynamic constraints on the vehicle's own stability (such as anti-rollover).

[0094] Specifically, the unmanned vehicle is performing an emergency avoidance maneuver. The driver (or the autonomous driving system) has already turned the steering wheel to the left, and the massive cab and trailer have begun to twist relative to each other. The system calls up the unmanned vehicle's factory configuration: the cab is 4 meters long, the trailer is 13.5 meters long, the wheelbase distribution, and the key articulation point position (traction seat position). The system accesses chassis data in real time and detects that the steering wheel angle has turned 120 degrees to the left. At the same time, the attitude data shows that the relative articulation angle between the cab and the trailer has reached 8 degrees (the trailer lags behind and shows a folding trend). The system uses coordinate transformation to superimpose these dynamic data onto the static geometric model. The generated model is no longer a straight train shape, but a dynamic vehicle body with a slight "V" shape, accurately reflecting the physical posture of the trailer swinging to the right and outward under inertia.

[0095] Although the unmanned vehicle was not making a low-speed, sharp turn, its rear track still contracted inwards during an emergency lane change. The system calculations revealed that due to the trailer's large mass (49 tons), its rear axle track experienced a 1.5-meter inner wheel difference contraction relative to the tractor's track. This meant that while the tractor's front avoided the obstacle on the right, the trailer's rear wheels might cut into it. Based on the current left-turn angle and lateral acceleration, the system calculated the instantaneous radius of curvature of the tractor's center of gravity to be approximately 200 meters. However, due to the trailer's tendency to sideslip, its center of rotation shifted lag behind the tractor. The system calculated the trailer's lateral sway value relative to the tractor. Due to the centrifugal force generated by the sharp turn, the rightmost end of the trailer (the rear of the trailer) swung outwards by 2.2 meters, resulting in the unmanned vehicle's actual width being far greater than its static width.

[0096] The system uses the calculated inner wheel difference trajectory and trailer tail-swing amplitude to construct a wraparound 3D polyhedron set around the unmanned vehicle. This set is not tightly attached to the vehicle body, but rather an "expanded envelope" with reserved dynamic deformation margin. Considering that the trailer is tail-swinging violently, the physical constraint layer extends 2.5 meters to the right rear in a fan-shaped space. This means that the system logically believes: "Although the front of the vehicle is here now, the rear of my trailer will definitely sweep across this area on the right within 0.5 seconds." This constraint layer defines an insurmountable boundary. If the car or small truck in S131 enters this "predicted fan-shaped space," the system determines that a physical collision is unavoidable and must trigger emergency braking or differential control in S152.

[0097] Through S132, the unmanned vehicle not only sees obstacles on the map, but also clearly "perceives" its own dangerous movement posture; the system clearly knows: "Although the front of the vehicle avoids the vehicle in front by turning left, the huge trailer is violently tail-swinging to the right, and there is a serious inner wheel difference cutting in;" This provides accurate boundary conditions of the vehicle itself for the subsequent area division in S141 and risk assessment in S142.

[0098] Therefore, the physical constraint layer is combined with the current driving data of the autonomous vehicle and the enhanced panoramic image. During the combination process, multiple dynamic detection items are constructed. Based on the multi-level training of each dynamic detection item, the dynamic detection system of the autonomous vehicle is determined. This dynamic detection system constructs a real-time digital twin model and dynamically presents the collision risk between the vehicle body and obstacles in the enhanced panoramic image. It is compatible with the overall consideration of multi-level training of each dynamic detection item, ensuring the accuracy of the dynamic detection system of the autonomous vehicle. At the same time, visual occlusion areas are introduced and the enhanced panoramic image is further controlled, which improves the accuracy of the dynamic detection system and ensures the dynamic detection of the autonomous vehicle during driving.

[0099] At this point, the system takes the "physical constraint layer" (describing the dynamic envelope of the unmanned vehicle) constructed in S132, the real-time collected "current driving data" (speed, acceleration), and the "enhanced panoramic view" (containing the complete environment of occluded targets) generated in S131 as input; it maps the vehicle's physical constraint layer to the world coordinate system of the enhanced panoramic view to ensure that the vehicle's future sweep path and the position of obstacles in the environment are compared under the same spatiotemporal reference; based on the mapping results, the system instantiates several specific "dynamic detection items". These items are not general detection boxes, but detection tasks for specific interaction logic, such as: "right-side crush detection based on inner wheel difference trajectory", "forward collision avoidance detection based on relative speed", "side scraping detection based on trailer sway", and "cutting conflict detection based on blind spot target".

[0100] The system performs hierarchical training on each constructed dynamic detection item, forming a layered detection architecture: Level 1 (Physical Rule Layer): Based on rigid body kinematics, it quickly determines whether geometric spaces overlap (e.g., whether the vehicle body directly presses on an obstacle); Level 2 (Statistical Learning Layer): Using a classifier trained on historical driving data (e.g., SVM, Random Forest), it determines the obstacle's movement intention (e.g., whether it illegally changes lanes); Level 3 (Deep Prediction Layer): Using LSTM or Transformer networks, it predicts long-term trajectories in complex interaction scenarios and determines the potential collision probability; The trained models at each level are integrated to determine the final "dynamic detection system," which can automatically schedule different levels of detection items according to the complexity of the current scene, ensuring both real-time performance in simple scenarios and accuracy in complex scenarios.

[0101] The dynamic detection system utilizes high-fidelity vehicle dynamics and environmental models to construct a "real-time digital twin model." This model is not only completely synchronized with the real world at the current moment but also extrapolates the situation 3-5 seconds into the future. The model includes the dynamic response characteristics of the unmanned vehicle (such as braking lag and trailer tail-swing) and the predicted trajectories of obstacles in the environment. In the digital twin space, the system calculates the spatiotemporal intersection between the physical constraint layer of the unmanned vehicle (including inner wheel difference and sway) and obstacles in the enhanced panoramic image. By calculating indicators such as TTC (Time to Collision) and Minimum Safe Distance (MTTC), it dynamically quantifies and presents the "collision risk," which is usually output in the form of a risk potential field map or heat map, intuitively showing the proximity and danger level of various parts of the vehicle body to surrounding obstacles.

[0102] Specifically, the unmanned vehicle is making a sharp left turn to avoid an obstacle, the trailer is swinging to the right, and there is a small truck that is obscured in the blind spot on the right. The system maps the physical constraint layer generated by S132 (including the envelope of the inner wheel difference trajectory and the 2.2-meter drift) to the enhanced panoramic world coordinate system of S131.

[0103] Based on the current interaction logic, the system instantiates four key "dynamic detection items": "Right-side crush detection based on inner wheel difference trajectory": detects whether the inner wheel of the trailer will sweep into the non-motorized vehicle lane or shoulder; "Forward collision avoidance detection based on relative speed": detects the longitudinal distance between the front of the trailer and the vehicle in front (car); "Side scraping detection based on trailer sway" (core): detects whether the rightmost rear end of the trailer will collide with the small truck on the right; "Entry conflict detection based on blind spot target": analyzes whether the trajectory of the small truck on the right intersects with the trajectory of the unmanned vehicle.

[0104] The car had already entered the lane, followed closely by the minivan, instantly increasing the scene's complexity. The first level (physical rule layer): the system quickly ran a geometric algorithm; within microseconds, it calculated that the physical constraint layer boundary of the trailer had spatially overlapped with the minivan's enclosure box on the right; a preliminary judgment was made: a geometric conflict existed. The second level (statistical learning layer): the system called a classifier (SVM) based on historical data to analyze the minivan's motion characteristics; the classifier determined that the minivan had not used its turn signal and its longitudinal speed was stable, belonging to a "straight-line following mode," showing no intention to swerve to the left, which greatly increased the collision risk. The third level (deep prediction layer): for the most complex "trailer swaying and scraping" scenario, the system activated an LSTM network, inputting the yaw rate of the unmanned vehicle and the road friction coefficient of the past second, to predict the trailer's swaying amplitude within the next two seconds. The results of the three-level models were fused, confirming that the "side scraping detection" scenario had the highest risk level, while the "forward collision avoidance detection" scenario had a lower risk level because the front of the vehicle had already avoided the car. The system ultimately determined a dedicated dynamic detection system for the current slippery road conditions.

[0105] The system constructed a high-fidelity dynamic model of the unmanned vehicle (simulating the inertial lag of a 49-ton mass on a slippery road surface) and a surrounding environment model (including a car and a small truck) in digital space. The model extrapolated the situation 4 seconds ahead. The simulation showed that the front of the unmanned vehicle moved to the left, and the trailer continued to swing to the right under the action of centrifugal force. At T+1.8 seconds, the right rear corner of the trailer and the left front corner of the small truck intersected in space. The system calculated the TTC (Time to Collision) at the intersection point to be 1.8 seconds, which is less than the safety threshold (2.5 seconds). At the same time, the calculated minimum safe distance (MTTC) was negative, indicating that a collision would be unavoidable without intervention. The system output a "risk potential field heat map". In the risk potential field heat map, the area in front of the unmanned vehicle was displayed as green (safe), but the area in the right rear of the trailer showed an explosive bright red hot spot, the position of which corresponded precisely to the small truck on the right.

[0106] refer to Figure 5 In step S14, the specific steps are as follows:

[0107] S141: In this dynamic detection system, the space around the vehicle is determined based on the vehicle's geometric boundary and the enhanced panoramic view. The space around the vehicle is divided into regions and is divided into multiple dynamic detection areas in real time. The multiple dynamic detection areas are the near-field collision zone of the lane, the lateral blind spot and blind angle monitoring zone, the adjacent lane cut-in warning zone, and the far-field intersection conflict zone.

[0108] S142: For each dynamic detection area, analyze the content of each dynamic detection area, identify the type, relative position and motion vector of each target in the dynamic detection area, and predict the collision time and headway in combination with the vehicle's kinematic model. Determine the dynamic warning level of the dynamic detection area based on the collision time, headway and the shape of each target.

[0109] S143: Determine the current driving status of the unmanned vehicle based on multiple driving data of the unmanned vehicle, determine the primary warning logic based on the current driving status of the unmanned vehicle and the corresponding surrounding weather characteristics, and determine the warning control logic of the unmanned vehicle based on the primary warning logic and the multi-level optimization of the dynamic warning level of each dynamic detection area.

[0110] In the embodiments of this application, the dynamic detection system determines the vehicle's surrounding space based on the vehicle's geometric boundaries and the enhanced panoramic image. The vehicle's surrounding space is then divided into multiple dynamic detection areas in real time. These multiple dynamic detection areas are the near-field collision zone of the current lane, the lateral blind spot and blind angle monitoring zone, the adjacent lane cut-in warning zone, and the far-field intersection conflict zone. This system takes into account both the vehicle's geometric boundaries and the enhanced panoramic image, ensuring the accuracy of the vehicle's surrounding space.

[0111] At this point, the system constructs a precise oriented bounding box (OBB) based on the geometric parameters (length, width, height, wheelbase) of the unmanned vehicle, and expands outward from this core. Combined with the "enhanced panoramic view" generated by S131, the system evaluates the effective detection range and confidence level of the sensors (LiDAR, camera). The system delineates a dynamic "vehicle periphery space" centered on the unmanned vehicle. This space not only eliminates blind spots that the sensors cannot perceive (such as behind buildings), but also eliminates low-risk areas at extremely long distances, forming a closed three-dimensional spatial field that is close to the vehicle and rich in effective perception information.

[0112] The system does not perform uniform segmentation, but rather non-uniform segmentation based on vehicle kinematic characteristics (such as minimum turning radius and braking distance). The system divides the space determined by S141-1 into logical topologies according to longitudinal distance (near / middle / far) and lateral angle (front / side / rear). This segmentation ensures that high-interaction-frequency areas (such as front and rear) have higher spatial resolution, while low-interaction-frequency areas use larger grid cells, thereby optimizing the allocation of computing power.

[0113] Based on spatial segmentation, the system specifically instantiates four key "dynamic detection areas" and assigns specific detection task attributes to each area: Near-field collision zone: defined as a fan-shaped area in front of the vehicle within 0 to the TTC (Time to Collision) threshold. This area requires extremely high refresh rates and recognition accuracy to address rear-end collision risks; Lateral blind spot and blind angle monitoring zone: covering the blind spots on both sides of the autonomous vehicle, the B-pillar / C-pillar blind spots, and the near-ground blind angle below the front of the vehicle; specifically, for the "inner wheel difference" area when heavy trucks turn right, this area is separately marked as a high-voltage zone; Adjacent lane cut-in warning zone: defined within the left and right adjacent lanes, in areas parallel to and a certain distance ahead and behind the autonomous vehicle, used to monitor whether adjacent vehicles have a trajectory trend of merging into this lane; Far-field intersection conflict zone: defined at intersections, merging points, etc., at a relatively far distance ahead, used for macro-level traffic flow prediction and path planning.

[0114] Furthermore, for each dynamic detection area, the content of each dynamic detection area is analyzed to identify the type, relative position, and motion vector of each target within the dynamic detection area. The collision time and headway are predicted in conjunction with the vehicle's kinematic model. Based on the collision time, headway, and the shape of each target, the dynamic warning level of the dynamic detection area is determined. This approach takes into account the overall consideration of collision time, headway, and the shape of each target, ensuring the accuracy of the dynamic warning level of the dynamic detection area.

[0115] At this point, the system traverses the data within each dynamic detection area; it uses deep learning networks (such as PointPillars or Centernet) to cluster and segment the point cloud and image features in the enhanced panoramic image, filters out ground clutter and static backgrounds (such as guardrails and road markings), and locks in potential targets of interest (ROIs).

[0116] Refined attribute extraction is performed on the locked target, including: type identification: determining the target category (such as car, pedestrian, non-motorized vehicle, large obstacle); relative position: calculating the three-dimensional coordinates (x, y, z) of the target's centroid in the unmanned vehicle coordinate system and its geometric dimensions; motion vector: using Kalman filtering or multi-target tracking algorithms (such as DeepSORT), combined with historical frame data, to calculate the target's velocity vector and acceleration vector, especially its relative velocity component with respect to the unmanned vehicle.

[0117] The system does not only calculate geometric distances, but also calls the current kinematic model of the unmanned vehicle (including the current load mass, tire-road adhesion coefficient, and braking system response delay time). Considering that the braking distance of a fully loaded heavy truck is 2-3 times that of an unloaded truck, the model makes a conservative estimate of the future deceleration trajectory. At this time, the time to collision (TTC) is calculated based on relative distance and relative speed, with the constraint term of the maximum braking deceleration of the unmanned vehicle added to ensure that the prediction is the "remaining time for a physically avoidable collision". At the same time, the time difference between the front of the unmanned vehicle reaching the current position of the target (or the target passing through the reference point) is calculated. The headway (THW) is mainly used to evaluate the safety of the following distance. For heavy trucks, the safety threshold is usually set at 2-3 seconds, which is much higher than the 1.5 seconds for cars.

[0118] The system weights and fuses the acquired "target shape" (such as rigid metal objects, flexible human bodies, and size) with the calculated collision time and headway. For example, when facing pedestrians and crash barriers at the same headway, pedestrians have a higher risk weight; when facing cars and trucks at the same distance, trucks have a higher risk level because of their larger mass and more severe collision consequences.

[0119] Using multi-level decision trees or fuzzy logic algorithms, the fused risks are mapped to discrete "dynamic warning levels," typically categorized as follows: Level 0 (Safe): TTC > 5.5s, no collision risk; Level 1 (Attention): TTC between 4.0s and 5.5s, attention is required; Level 2 (Warning): TTC between 2.5s and 4.0s, preparation for braking is required; Level 3 (Danger): TTC < 2.5s and the vehicle is in the same lane, or THW < 1.8s, immediate braking or evasive action is required.

[0120] Therefore, the current driving state of the unmanned vehicle is determined based on multiple driving data. The initial warning logic is determined based on the current driving state of the unmanned vehicle and the corresponding surrounding weather characteristics. The warning and control logic of the unmanned vehicle is determined based on the initial warning logic and the multi-level optimization of the dynamic warning level of each dynamic detection area. This approach is compatible with the overall consideration of the initial warning logic and the multi-level optimization of the dynamic warning level of each dynamic detection area, ensuring the accuracy of the warning and control logic of the unmanned vehicle.

[0121] At this time, the system collects multiple driving data from the chassis CAN bus in real time, including longitudinal speed, lateral acceleration, steering wheel angle, brake master cylinder pressure, current gear, and load status (estimated by air suspension pressure). The system uses Kalman filtering or a state observer to fuse the above data, eliminate sensor noise, and identify the vehicle's current driving state. This state is not only a simple constant speed or deceleration, but also includes complex dynamic states, such as "full-load steady-state cruise", "unloaded sharp turn", "engine braking on a long downhill", or "low-traction slippage tendency". This state vector directly determines the vehicle's current maneuverability and braking limits.

[0122] The system acquires "surrounding weather characteristics," such as light intensity (night / day), precipitation (rain / snow), visibility (fog / haze), and road surface temperature. These characteristics are converted into road surface adhesion coefficient estimates and sensor confidence correction factors. The system couples and matches the determined "current driving state" with the weather characteristics. For example, when the state is "fully loaded" and the weather is "heavy rain," the road surface adhesion coefficient decreases, and the braking distance is significantly extended. At this time, the "primary warning logic" determined by the system will automatically increase its sensitivity: increase the safe following distance (increase the TTC threshold), reduce the maximum speed limit, and shorten the alarm response time. This is a "global safety baseline" based on physical constraints, ensuring that the vehicle is in a defensive driving mode before any local risks occur.

[0123] The system takes the "dynamic warning level" (Level 0-3) of each area calculated by S142 as input and dynamically corrects the generated primary warning logic. This usually adopts priority-based preemptive scheduling or weighted fusion algorithm. If a Level 3 (extremely high risk) warning occurs in a local area, the signal will have the highest priority and directly override the regular parameters in the primary logic, triggering the emergency braking strategy. If it is Level 1 or Level 2, the smooth intervention mode of the primary logic (such as sound prompts and intermittent braking) is followed.

[0124] The system determines the final "early warning control logic", which includes: HMI level: volume, frequency, and color of the audible and visual alarm (such as red flashing); chassis intervention level: braking deceleration magnitude, steering assist correction magnitude, and preset parameters of the electronic stability system (ESC); communication level: sending collision warning signals (V2X) to the following vehicle.

[0125] refer to Figure 6 In step S15, the specific steps are as follows:

[0126] S151: Mark the tractor and trailer in the unmanned vehicle, determine the corresponding relative sway angle based on the dynamic detection of the tractor and trailer, deeply fuse the relative sway angle, the current driving data of the unmanned vehicle and the corresponding road friction coefficient, and determine the corresponding variable features during the fusion process. The variable features reflect the content of the unmanned vehicle in the critical instability state.

[0127] S152: Input the variable characteristics and the early warning and control logic of the unmanned vehicle into the multi-level inference decision engine, and perform comprehensive analysis in the multi-level inference decision engine to output multiple emergency features. Based on the multiple emergency features, the on-board load data of the unmanned vehicle and the corresponding current driving status, determine the corresponding multimodal emergency measures.

[0128] In the embodiments of this application, the tractor and trailer of the unmanned vehicle are marked, and the corresponding relative sway angle is determined based on the dynamic detection of the tractor and trailer. The relative sway angle, the current driving data of the unmanned vehicle, and the corresponding road friction coefficient are deeply fused, and the corresponding variable features are determined during the fusion process. The variable features reflect the content of the unmanned vehicle in the critical instability state, which is compatible with the overall consideration of the dynamic detection of the tractor and trailer, and ensures the accuracy of the corresponding relative sway angle.

[0129] At this point, the system uses lidar and visual sensors to perform semantic segmentation of the unmanned vehicle's own structure in point clouds and image pixels; the system accurately marks the rear end features of the tractor (cab) and the front end features of the trailer, and establishes rigid body models for them in the vehicle coordinate system.

[0130] Based on inertial measurement unit (IMU) data from the tractor and trailer, as well as observations from external sensors, the system tracks their heading angles in real time. By calculating the angle between the projections of the tractor's longitudinal axis and the trailer's longitudinal axis onto the horizontal plane, the system calculates the "relative sway angle" (i.e., articulation angle). Simultaneously, the system calculates the angular velocity of this angle to capture the dynamic trend of the sway, which is the direct basis for determining whether the vehicle has "turned over" or "fishtailed".

[0131] The system uses the acquired "relative yaw angle" as the main variable, and simultaneously incorporates the "current driving data" of the unmanned vehicle (including longitudinal velocity, yaw rate, lateral acceleration, and steering wheel angle) and the "road friction coefficient" (μ value) estimated by the sensors. It employs extended Kalman filtering (EKF) or multi-model adaptive estimation algorithms for deep fusion. The fusion process is not a simple weighting, but a nonlinear state estimation based on vehicle dynamics models (such as three-wheeled models or dual-track models). At the same time, the road friction coefficient is included as a constraint in the fusion equation. For example, under low friction coefficients, the same relative yaw angle may mean that the tire is close to the adhesion limit. The fusion algorithm will adjust the state covariance matrix accordingly to increase the weight of sensitivity to abnormal states.

[0132] Based on the fused state, the system identifies and calculates "variable features" reflecting the vehicle's "critical instability state." These features go beyond a single position or angle and touch upon the boundaries of physical stability. Core features include: Lateral Load Transfer Rate (LTR): Real-time calculation of the ratio of vertical loads on the left and right wheels. When the LTR approaches 1, it means that the inner wheel is about to leave the ground, and the vehicle is at the critical point of rollover. Center of gravity sideslip angle: Especially for trailers, when it exceeds the stability domain (such as the phase plane stability boundary), it indicates that the vehicle is about to sideslip. Phase plane features: Constructing the phase plane trajectory of sideslip angle and sideslip rate to determine whether the current state converges to the stable focus.

[0133] Furthermore, the variable features and the early warning and control logic of the unmanned vehicle are input into a multi-level inference decision engine, where they are comprehensively analyzed to output multiple emergency features. Based on these multiple emergency features, the unmanned vehicle's onboard load data, and the corresponding current driving state, corresponding multimodal emergency measures are determined. This approach incorporates a holistic consideration of multiple emergency features, the unmanned vehicle's onboard load data, and the corresponding current driving state, ensuring the accuracy of the corresponding multimodal emergency measures. At the same time, the variable features are further controlled, enabling multi-level inference of the variable features and the unmanned vehicle's early warning and control logic, thereby improving the accuracy of the multimodal emergency measures.

[0134] At this point, the system simultaneously inputs the "variable features" (such as high lateral load transfer rate LTR and large centroid sideslip angle) extracted by S151 reflecting critical instability and the "early warning and control logic" (such as sharp left turn to avoid collision and emergency braking) formulated by S143 into the "multi-level inference decision engine". This engine usually adopts a hybrid architecture based on rule base and model predictive control (MPC). The lower rule base handles fast reactions (such as directly triggering ABS), and the upper MPC performs multi-objective optimization. The engine will perform nonlinear solution to the current dynamic state and weigh various driving needs.

[0135] Through game theory analysis, the engine outputs a set of high-dimensional "emergency features" to describe the specific nature of the current crisis. These features include: "Jackknifing Risk", "Rollover Imminent", and "Lane Deviation Margin". These features are no longer simple data, but a qualitative description of the conflict between the vehicle's current physical state and the control objective.

[0136] The system acquires the "onboard load data" (real-time total weight and axle load distribution calculated by air suspension pressure sensors) and "current driving status" (vehicle speed, gear, throttle opening) of the unmanned vehicle. This data is crucial because the rotational inertia and center of gravity height of a fully loaded 49-ton truck and an unloaded 10-ton truck are vastly different, and the control strategies are completely different.

[0137] Based on the above parameters, the system matches and determines "multimodal emergency measures" from the control library. This is not a single action, but a coordinated operation of multiple chassis systems: Differential braking: applying braking force to the wheels on a specific side, using the generated yaw moment to correct the attitude of the tractor or trailer; Power limiting / cut-off: reducing or cutting off drive torque through the engine management system (EMS) to reduce drive wheel slippage; Active self-centering / damping control: if it is a semi-active suspension system, adjusting suspension damping to suppress body roll; if it is steer-by-wire, fine-tuning the front wheel angle to stabilize the center of gravity; Independent trailer braking: performing independent braking control on the trailer axle to directly suppress trailer sway.

[0138] Please see Figure 7 , Figure 7 This is a schematic diagram of the structural composition of a vehicle dynamic detection system according to an embodiment of the present invention; the vehicle dynamic detection system includes:

[0139] The dynamic driving stereoscopic image module 21 is used for unmanned vehicles to drive on the road. It combines multiple road images and point cloud data detected by LiDAR for spatial alignment, and combines the driving data of the unmanned vehicle to determine the dynamic driving stereoscopic image of the unmanned vehicle.

[0140] The occlusion dynamic target module 22 is used to perform multimodal learning on the dynamic driving stereo image and the driving text information of the unmanned vehicle, and to mark the visual occlusion area of ​​the unmanned vehicle, and to determine the occlusion dynamic target based on the compensation detection of the visual occlusion area.

[0141] The dynamic detection system module 23 is used to determine the enhanced panoramic image based on the occluded dynamic target and the dynamic driving stereo image; mark multiple key geometric features of the unmanned vehicle, and construct the dynamic detection system based on the multiple key geometric features, the current driving data of the unmanned vehicle and the enhanced panoramic image;

[0142] The early warning and control logic module 24 is used to output multiple dynamic detection areas by the dynamic detection system and determine the early warning and control logic of the unmanned vehicle based on the area content of each dynamic detection area, the corresponding dynamic early warning level and the driving status of the unmanned vehicle.

[0143] The multimodal emergency response module 25 is used to mark the relative sway angle between the tractor and trailer in the unmanned vehicle, determine the corresponding variable features based on the relative sway angle and the current driving data of the unmanned vehicle, and determine the corresponding multimodal emergency response based on the variable features and the multi-level reasoning of the early warning and control logic of the unmanned vehicle.

[0144] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for dynamic detection of a vehicle, characterized in that, include: When an unmanned vehicle is driving on the road, multiple road images and point cloud data detected by LiDAR are combined and spatially aligned, and the driving data of the unmanned vehicle is combined to determine the dynamic driving stereo map of the unmanned vehicle. Multimodal learning is performed on dynamic driving stereo images and driving text information of unmanned vehicles, and visual occlusion areas of unmanned vehicles are marked. Occlusion dynamic targets are determined based on compensation detection of these visual occlusion areas. This includes: collecting driving text information of unmanned vehicles; introducing a multimodal learning mechanism into the dynamic driving stereo images and driving text information of unmanned vehicles; and parsing the road scene of unmanned vehicles through cross-modal semantic association to mark visual occlusion areas that cannot be directly observed by visual sensors; for visual occlusion areas, the beam penetration characteristics of lidar are used to extract point cloud echoes behind the occluded objects, and combined with the temporal information of the dynamic driving stereo images, Kalman filtering is used to track the motion trend around the occlusion area to determine multiple dynamic features; and the occlusion dynamic targets are determined based on the characteristic shape, relative position, and corresponding point cloud echoes of the multiple dynamic features, and the specific position and motion state of the occlusion dynamic targets under occlusion conditions are marked. An enhanced panoramic image is determined based on the occluded dynamic target and the dynamic driving stereo image; multiple key geometric features of the unmanned vehicle are marked, and a dynamic detection system is constructed based on the multiple key geometric features, the current driving data of the unmanned vehicle, and the enhanced panoramic image. This includes: determining the position coordinates and attribute information of the occluded dynamic target based on the detection of the occluded dynamic target, and loading the position coordinates and attribute information of the occluded dynamic target into the dynamic driving stereo image. At this time, information gaps in the dynamic driving stereo image are filled through semantic alignment and edge smoothing processing, and a corresponding enhanced panoramic image is generated. The enhanced panoramic image includes both visible targets and occluded dynamic targets. The dynamic detection system outputs multiple dynamic detection areas, and determines the warning and control logic of the unmanned vehicle based on the area content of each dynamic detection area, the corresponding dynamic warning level, and the driving status of the unmanned vehicle. The relative sway angle between the tractor and trailer in the unmanned vehicle is marked. Based on the relative sway angle and the current driving data of the unmanned vehicle, the corresponding variable features are determined. Based on the variable features and the multi-level reasoning of the early warning and control logic of the unmanned vehicle, the corresponding multimodal emergency measures are determined.

2. The vehicle dynamic detection method according to claim 1, characterized in that, The unmanned vehicle travels on the road, and multiple road images and point cloud data detected by LiDAR are combined and spatially aligned. This, combined with the unmanned vehicle's driving data, determines a dynamic 3D view of the unmanned vehicle's movement, including: The system monitors the road driving process of unmanned vehicles in real time and collects multiple road images from different positions of the unmanned vehicles based on multiple cameras of the unmanned vehicles. The pixel coordinate system of the multiple road images is transformed into a unified LiDAR vehicle coordinate system. In this LiDAR vehicle coordinate system, the multiple road images and the point cloud data detected by the LiDAR are combined and aligned in the same spatial dimension to generate an RGB-D fusion map. This RGB-D fusion map contains color texture and depth information. The system accesses the chassis CAN bus data of the unmanned vehicle in real time and extracts multiple driving data from the unmanned vehicle. It then uses these multiple driving data to perform frame-by-frame motion distortion correction on the RGB-D fusion image and outputs the corrected primary driving stereo image. Finally, it aligns the corrected primary driving stereo image with the vehicle status information with timestamps in the temporal dimension to construct a dynamic driving stereo image of the unmanned vehicle.

3. The vehicle dynamic detection method according to claim 1, characterized in that, The process of determining an enhanced panoramic image based on the occluded dynamic target and the dynamic driving stereoscopic image; marking multiple key geometric features of the unmanned vehicle; and constructing a dynamic detection system based on the multiple key geometric features, the current driving data of the unmanned vehicle, and the enhanced panoramic image also includes: The overall shape of the unmanned vehicle is collected, and several key geometric features are determined based on the overall shape of the unmanned vehicle, the corresponding driving data and attitude data. These key geometric features cover the steering wheel difference trajectory, turning radius and outer contour sway value. A physical constraint layer for the unmanned vehicle during driving is constructed based on these key geometric features. The physical constraint layer is combined with the current driving data of the unmanned vehicle and the enhanced panoramic image. During the combination process, multiple dynamic detection items are constructed. Based on the multi-level training of each dynamic detection item, the dynamic detection system of the unmanned vehicle is determined. This dynamic detection system constructs a real-time digital twin model and dynamically presents the collision risk between the vehicle body and obstacles in the enhanced panoramic image.

4. The vehicle dynamic detection method according to claim 1, characterized in that, The dynamic detection system outputs multiple dynamic detection zones. Based on the content of each dynamic detection zone, the corresponding dynamic warning level, and the driving status of the unmanned vehicle, the system determines the warning and control logic for the unmanned vehicle, including: In this dynamic detection system, the space surrounding the vehicle is determined based on the vehicle's geometric boundaries and the enhanced panoramic image. The space surrounding the vehicle is then divided into multiple dynamic detection areas in real time. These multiple dynamic detection areas are the near-field collision zone of the current lane, the lateral blind spot and blind angle monitoring zone, the adjacent lane cut-in warning zone, and the far-field intersection conflict zone.

5. The vehicle dynamic detection method according to claim 4, characterized in that, The dynamic detection system outputs multiple dynamic detection areas. Based on the content of each dynamic detection area, the corresponding dynamic warning level, and the driving status of the unmanned vehicle, the system determines the warning and control logic for the unmanned vehicle. It also includes: For each dynamic detection area, the content of each dynamic detection area is analyzed, the type, relative position and motion vector of each target in the dynamic detection area are identified, and the collision time and headway are predicted by combining the vehicle's kinematic model. The dynamic warning level of the dynamic detection area is determined based on the collision time, headway and the shape of each target. The current driving status of the unmanned vehicle is determined based on multiple driving data. A primary warning logic is determined based on the current driving status of the unmanned vehicle and the corresponding surrounding weather characteristics. The warning and control logic of the unmanned vehicle is determined based on the primary warning logic and the multi-level optimization of the dynamic warning level of each dynamic detection area.

6. The vehicle dynamic detection method according to claim 1, characterized in that, The relative sway angle between the tractor and trailer in the marked unmanned vehicle is used to determine corresponding variable features based on this relative sway angle and the current driving data of the unmanned vehicle. Based on these variable features and multi-level reasoning of the unmanned vehicle's early warning and control logic, corresponding multimodal emergency measures are determined, including: The tractor and trailer of the unmanned vehicle are labeled, and the corresponding relative sway angle is determined based on the dynamic detection of the tractor and trailer. The relative sway angle, the current driving data of the unmanned vehicle, and the corresponding road friction coefficient are deeply fused, and the corresponding variable features are determined in the fusion process. These variable features reflect the content of the unmanned vehicle in the critical instability state.

7. The vehicle dynamic detection method according to claim 6, characterized in that, The relative sway angle between the tractor and trailer in the marked unmanned vehicle is used to determine corresponding variable features based on the relative sway angle and the current driving data of the unmanned vehicle. Based on these variable features and multi-level reasoning of the unmanned vehicle's early warning and control logic, corresponding multimodal emergency measures are determined. The method also includes: The variable characteristics and the early warning and control logic of the unmanned vehicle are input into the multi-level inference decision engine, and comprehensive analysis is performed in the multi-level inference decision engine to output multiple emergency features. Based on the multiple emergency features, the on-board load data of the unmanned vehicle and the corresponding current driving status, the corresponding multimodal emergency measures are determined.

8. A dynamic detection system for a vehicle, characterized in that, The vehicle dynamic detection system is applied to the vehicle dynamic detection method as described in any one of claims 1-7.

Images