Method and system for capturing tiny collision of vehicle and reconstructing trajectory in anti-shielding environment

Abstract

The invention discloses a vehicle tiny collision capture and trajectory reconstruction method and system in an anti-shielding environment, and the method comprises the steps: constructing a vehicle semantic space-time diagram of dynamic shielding perception, and locking a visible rigid connection feature part of a shielding risk target; frequency domain micro-motion analysis is carried out on the feature part based on phase amplification, and vibration fingerprint feature vectors are extracted in combination with a vehicle type-modal frequency fingerprint database; by introducing an impulse-momentum theorem as a neural network of physical constraint, performing collision event reasoning on the vibration fingerprint features and regressing a collision impulse; and repairing and generating a vehicle track conforming to collision dynamics in the shielding region by using a denoising diffusion model under the condition of collision impulse. According to the method, the collision is inversely deduced by analyzing the microscopic vibration of the unshielded part, and the physical principle is integrated into the neural network, so that the problems of difficult detection of the small collision under the shielding and distortion of track restoration are solved, the track with the collision dynamic characteristics can be generated, and powerful evidence is provided for accident liability affirmation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence image processing and traffic monitoring technology, specifically to a method and system for capturing and reconstructing minor vehicle collisions in an occlusion-resistant environment. Background Technology

[0002] In the process of urbanization, with the continuous increase in vehicle density, minor vehicle collisions in complex traffic scenarios such as parking lots and narrow streets are becoming increasingly frequent. These accidents, such as minor scratches or low-speed contact, although causing limited property damage, often lead to disputes in determining liability, resulting in damaged vehicles remaining at the scene for extended periods, causing localized traffic congestion and reducing road efficiency. Existing technologies face technical challenges in addressing these issues.

[0003] First, visual occlusion is the primary factor affecting the effectiveness of surveillance systems. In dense traffic environments, the field of view of surveillance cameras is often obstructed by static objects (such as building pillars and walls) or other moving vehicles. The contact points of minor collisions often occur on the sides or corners of vehicle bodies, locations that are easily within blind spots. Therefore, traditional collision detection algorithms that rely on directly observing deformation at the contact point or debris splashing will fail under occlusion conditions due to the lack of direct visual evidence.

[0004] Secondly, the physical characteristics generated by minor collisions are extremely subtle. Unlike severe collisions, minor contact may only cause instantaneous, minute vibrations in the vehicle body or slight changes in the suspension system's attitude. In surveillance videos at normal frame rates, these changes are difficult to detect in both time and space; their amplitude may be less than a few pixels, easily masked by noise from video compression, changes in ambient lighting, or minor bumps during normal vehicle movement. Existing anomaly detection methods based on macroscopic motion analysis such as optical flow or background subtraction are ineffective at capturing such minute features.

[0005] Finally, current technologies lack refined trajectories that can provide crucial evidence for liability determination. Determining liability in traffic accidents requires not only confirming the occurrence of a collision but also accurately reconstructing the vehicle's trajectory before and after the collision to analyze driving behavior. However, commonly used vehicle trajectory tracking algorithms, such as those based on Kalman filtering or Long Short-Term Memory (LSTM) networks, inherently assume the smoothness and continuity of vehicle motion. When processing collision events, these algorithms tend to treat the nonlinear trajectory abrupt changes generated at the moment of collision as noise and smooth them out, thus erasing crucial evidence reflecting the facts of collision dynamics. This makes trajectory-based liability determination lack objective technical basis. Although some solutions detect collisions using onboard sensors (such as inertial measurement units), these solutions require the vehicles involved to be equipped with specific devices, making them unsuitable for existing traffic scenarios that heavily rely on roadside monitoring. Therefore, there is an urgent need for a technical solution that can effectively capture minor collisions and accurately reconstruct accident trajectories even when external visibility is obstructed and the point of contact cannot be directly observed. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art in that it is difficult to detect minor vehicle collisions under conditions of visual obstruction and weak collision features, and it is unable to reconstruct accident trajectories containing real dynamic features.

[0007] To address the aforementioned technical problems, one aspect of the present invention provides a method for capturing and reconstructing minor vehicle collisions in an environment with resistance to occlusion, comprising the following steps: Step 1: Construct a vehicle semantic spatiotemporal map for dynamic occlusion perception: Real-time parsing of monitoring video streams from roadside monitoring equipment or parking lot monitoring terminals, identification of static occlusions in the scene, identification and tracking of target vehicles, acquisition of their vehicle type classification and mass estimation values, identification of rigid connection feature parts on the vehicle body, and when it is determined that the target vehicle has an occlusion risk, the currently visible rigid connection feature parts are determined as micro-motion observation windows. Preferably, the rigid connection feature refers to a component that is directly connected to the vehicle chassis through a rigid structure, including at least one of the following: roof shark fin antenna, roof rack, A-pillar, B-pillar, or C-pillar upper edge.

[0008] Step 2: Frequency domain micro-motion feature extraction based on phase Euler amplification: The image sequence of the micro-motion observation window is decomposed into complex manipulable pyramids. The phase time sequence signal is weighted using a weighted frequency band mask generated based on the vehicle classification and quality estimation values ​​to highlight the collision resonance signal and suppress background noise, thereby extracting the vibration fingerprint feature vector. Preferably, the process of generating the weighted frequency band mask in step two includes: pre-constructing a vehicle-modal frequency fingerprint database that stores resonance frequency distribution data of different vehicle body structures; retrieving the corresponding resonance frequency distribution from the vehicle-modal frequency fingerprint database based on the vehicle classification result and mass estimation value of the target vehicle; and generating the weighted frequency band mask, wherein the weighted frequency band mask is configured to assign a first weight to the collision resonance frequency band of the target vehicle and a second weight to the engine idling or background noise frequency band, wherein the first weight is higher than the second weight.

[0009] Furthermore, the vibration fingerprint feature vector includes at least one of the following parameters: vibration energy spectral density, vibration attenuation rate, vibration principal direction, and phase consistency.

[0010] Step 3: Physically Constrained Collision Event Neural Inference: Input the vibration fingerprint feature vector into the physical perception temporal convolutional network, use the impulse-momentum theorem as a physical constraint to determine whether a collision event exists, and regress to obtain the collision impulse vector and the time of collision occurrence. Preferably, the physical-aware temporal convolutional network includes a physical constraint layer, and the physical constraint-driven collision event neural inference step further includes: the physical constraint layer calculating the minimum theoretical impulse required to generate the vibration based on the vibration energy spectral density contained in the vibration fingerprint feature vector and the estimated mass of the target vehicle; and the physical constraint layer comparing the collision impulse vector obtained by regression from the physical-aware temporal convolutional network with the minimum theoretical impulse, and if the magnitude of the collision impulse vector is less than a preset physical threshold, then setting the collision probability to zero.

[0011] Furthermore, when there are two vehicles suspected of colliding in the monitoring video stream, the physical constraint-driven collision event neural reasoning step further includes: only when the difference in the vibration start time of the two vehicles is less than a predetermined time threshold, and the main vibration directions of the two vehicles meet the condition that the action and reaction forces are opposite, is it finally determined to be a two-vehicle collision event.

[0012] Step 4: Refined trajectory repair based on conditional denoising diffusion model: Using the conditional denoising diffusion probability model, the trajectory segments before and after the collision, the moment of the collision, and the collision impulse vector in the dynamic trajectory information are used as guiding conditions to generate the motion trajectory of the target vehicle in the occluded area that conforms to the collision dynamics through the reverse diffusion process. Preferably, the refined trajectory repair step based on the conditional denoising diffusion model further includes: gradually denoising from Gaussian noise through a reverse diffusion process to generate trajectory points in the occlusion region; and, in each denoising step of the reverse diffusion process, introducing a dynamic perturbation field constructed based on the collision impulse vector and the collision occurrence time as an additional conditional bias term, wherein the dynamic perturbation field applies a velocity mutation gradient related to the magnitude and direction of the collision impulse vector to the sampling process as the collision occurrence time approaches, so as to guide the generated trajectory to produce discontinuous trajectory deflection at the collision point.

[0013] Furthermore, the refined trajectory repair step based on the conditional denoising diffusion model also includes: generating multiple candidate trajectories; and using static occlusions in the scene as geometric constraints to filter the multiple candidate trajectories, so as to remove invalid trajectories that cross the static occlusions, and to perform a weighted average on the remaining valid trajectories to output the final reconstructed trajectory.

[0014] Step 5: Accident Liability Map Generation and Evidence Chain Synthesis: Based on the reconstructed motion trajectory and collision physical parameters, generate accident evidence data for accident liability determination.

[0015] Preferably, step five further includes: quantifying the severity of the collision based on the reconstructed motion trajectory and collision physical parameters, generating structured data and visual evidence for accident liability determination, and sending the structured data to the traffic management platform to trigger an accident warning or assist in the liability determination process; the visual evidence includes: a slow-motion video with micro-motion magnification effect superimposed, and a top-down trajectory reconstruction animation showing the complete motion process of the target vehicle in the occluded area.

[0016] Another aspect of the present invention provides a vehicle minor collision capture and trajectory reconstruction system in an unobstructed environment, comprising: The panoramic perception and occlusion processing module is configured to analyze the monitoring video stream from roadside monitoring equipment or parking lot monitoring terminal in real time, identify static occlusions in the scene, identify and track target vehicles, obtain their vehicle type classification and mass estimation value, identify rigid connection feature parts on the vehicle body, and determine the currently visible rigid connection feature parts as micro-motion observation windows when it is determined that the target vehicle has an occlusion risk. The micro-feature frequency domain analysis module is configured to perform complex manipulable pyramid decomposition on the image sequence of the micro-motion observation window, and use the weighted frequency band mask generated based on the vehicle model-modal frequency fingerprint library to weight the phase time sequence signal to highlight the collision resonance signal and suppress background noise, thereby extracting the vibration fingerprint feature vector. The neurophysical inference engine is configured to input the vibration fingerprint feature vector into the physical perception temporal convolutional network, use the impulse-momentum theorem as a physical hard constraint to determine whether a collision event exists, and regress to obtain the collision impulse vector and the time of collision occurrence. The generative trajectory reconstruction server is configured to use a conditional denoising diffusion probability model, with trajectory segments before and after the collision, the time of the collision, and the collision impulse vector as guiding conditions, to generate the motion trajectory of the target vehicle in the occluded area in accordance with collision dynamics through a reverse diffusion process. The intelligent liability determination and evidence collection terminal is configured to quantify the severity of a collision based on the reconstructed motion trajectory and collision physical parameters, generate structured data and visual evidence for accident liability determination, and send the structured data to the traffic management platform to trigger accident warnings or assist in the liability determination process.

[0017] The technical solution provided by this invention effectively solves the technical challenge of detecting minor collisions under visual obstruction conditions by analyzing the microscopic vibrations of unobstructed parts of the vehicle body to infer invisible collision events. The physical principles of rigid body collisions, such as impulse conservation and resonant frequency characteristics, are deeply integrated into the design of the neural network model, ensuring the physical authenticity and logical reliability of the algorithm's inference results and significantly improving the accuracy of collision identification. Based on phase analysis-based micro-motion extraction technology, combined with adaptive frequency weighting for specific vehicle models, this method can detect sub-pixel-level micro-displacements, exhibiting high detection sensitivity and robustness against environmental interference for minor scratches and other incidents. Finally, the trajectory generated by the dynamically guided diffusion model retains the non-smooth motion abrupt changes at the moment of collision, realistically reproducing the accident process and providing strong technical support and evidence for traffic accident liability determination, especially in handling complex cases such as staged accidents and hit-and-runs. Attached Figure Description

[0018] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0019] Figure 1 This is a flowchart of a method for capturing and reconstructing minor vehicle collisions in an environment with resistance to occlusion, provided by an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of dynamic occlusion sensing and micro-motion observation window determination in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the structure of the Physical Aware Temporal Convolutional Network (Pa-TCN) in an embodiment of the present invention.

[0022] Figure 4This is a schematic diagram illustrating the principle of refined trajectory repair based on the conditional denoising diffusion probability model (cDDPM) in an embodiment of the present invention.

[0023] Figure 5 This is a visualization of the trajectory generation process based on the conditional denoising diffusion probability model in an embodiment of the present invention.

[0024] Figure 6 This is a bird's-eye view of the deployment of a parking lot obstruction scenario in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0026] Example 1 Please see Figure 1 This embodiment provides a method for capturing and reconstructing minor vehicle collisions in environments with strong visual obstruction. This method can be applied to urban traffic monitoring systems, parking management systems, or intelligent security platforms. The method aims to address the difficulties in detecting minor collisions and the unclear determination of liability caused by visual obstruction and weak collision features.

[0027] Step S100: Construct a vehicle semantic spatiotemporal graph for dynamic occlusion perception.

[0028] In a specific implementation scenario, such as Figure 2 As shown, the surveillance camera is facing a parking lot entrance containing building pillars. This step first receives and processes the real-time video stream from the camera. A pre-trained deep semantic segmentation network, such as Mask R-CNN or U-Net, is used to perform pixel-level parsing on each video frame. The parsing results divide the scene into a static background layer and a dynamic foreground layer. In the static background layer, a multi-frame background modeling algorithm stably identifies static occlusions such as pillars and walls, generating their corresponding static masks. In the dynamic foreground layer, an instance segmentation algorithm identifies each vehicle in the scene and assigns it a unique identifier. Simultaneously, a fine-grained vehicle classifier categorizes each vehicle, for example, identifying it as a sedan, SUV, or minivan. Based on a pre-built vehicle physical parameter library, the system assigns an average quality estimate, denoted as m_est, to each identified vehicle. This parameter library can be obtained by collecting publicly available vehicle specification data or through statistical modeling.

[0029] Furthermore, for each identified foreground vehicle, for example Figure 2 For vehicles A and B, the system uses an instance segmentation network to further identify rigidly connected features (RCFs) on their vehicle bodies. These features are preferably the shark fin antenna on the roof, roof rack, A-pillar, B-pillar, or the upper edge of the C-pillar. These features are chosen as observation points because they are rigidly connected to the vehicle chassis, enabling efficient and high-fidelity transmission of impact vibrations from collisions. Furthermore, their higher position reduces the probability of them being occluded by other vehicles in densely populated scenes. Subsequently, the system calculates the intersection-over-union (IoU) ratio between the bounding box of the target vehicle (e.g., vehicle A) and the masks of all other objects in the scene (including static pillars and dynamic vehicle B). When the calculated occlusion rate exceeds a preset threshold, such as 50%, or when the system predicts that vehicle A's future trajectory will enter a potential collision area that is occluded, the system marks vehicle A as an "occlusion risk target." At this point, the system locks one or more rigidly connected features of vehicle A that are still visible as micro-motion observation windows (ROIs), such as... Figure 2 The ROI region is shown in the diagram, and the image sequence data of that window is passed to the next step for processing.

[0030] Step S200: Frequency domain micro-motion feature extraction based on phase Euler amplification.

[0031] This step receives the image sequence of the Region of Interest (ROI) output from step S100, aiming to extract subtle vibration signals, imperceptible to the naked eye, caused by minute collisions. Specifically, a complex manipulable pyramid decomposition technique is first employed to decompose the image sequence within the ROI region into a series of complex subbands in the complex domain at multiple scales and in multiple directions. Compared to directly analyzing pixel brightness values, analyzing local phase changes in the complex subbands can more robustly capture subpixel-level minute displacements because it is insensitive to changes in illumination and has an approximately linear correspondence with the actual displacement.

[0032] To improve the signal-to-noise ratio of collision signal detection, this step introduces an adaptive resonant frequency weighting network based on a vehicle-modal frequency fingerprint database. This fingerprint database pre-stores the main resonant frequency distributions of different vehicle models (e.g., sedans, SUVs) under impact, obtained through finite element simulations or actual impact experiments. When processing the ROI signal, a frequency-spatial attention module receives the vehicle classification result and mass estimate m_est provided in step S100 as input. Based on the input information, this module retrieves or dynamically generates an optimal weighted frequency band mask from the fingerprint database. The mask is designed to assign high gain weights to the resonant frequency range most likely to occur for the vehicle model during a collision, while assigning low weights or zero weights to known background noise frequencies, such as low-frequency vibrations (typically 10-25 Hz) generated by the engine at idle or vibrations caused by common road surface irregularities. Before phase amplification, this mask is applied to each decomposed frequency sub-band signal to achieve targeted enhancement of the potential collision signal and effective noise suppression.

[0033] Subsequently, the weighted phase timing signal is multiplied by a preset amplification factor α. The value of α can be dynamically adjusted according to the scene's signal-to-noise ratio. For example, in a well-lit indoor parking lot with a stable background, α can be set to a higher value, such as 40; while in outdoor scenes with significant wind noise or tree shadow interference, α can be appropriately lowered to 15 to avoid over-amplifying noise. The amplified phase information is then re-fused with the original signal to generate a "micro-motion enhanced video stream." In this video stream, previously invisible collision vibrations are visually amplified into clearly discernible vehicle body swaying. Finally, a short-time Fourier transform (STFT) is performed on the filtered and weighted phase timing signal, and a set of characteristic parameters that comprehensively describe the vibration characteristics are extracted, including but not limited to vibration energy spectral density (ESD), vibration energy distribution in a specific frequency band, vibration attenuation rate time constant, vibration principal direction vector, and multi-point phase consistency. The vibration principal direction vector is defined as the direction of the pixel displacement principal axis determined by the histogram of the complex sub-band phase change gradient within the micro-motion observation window. In the specific calculation, principal component analysis (PCA) is performed on the instantaneous phase change vectors of all valid pixels within the observation window. The direction of the eigenvector corresponding to the first principal component is taken as the principal vibration direction on the image plane, and it is projected onto the world coordinate system through the camera extrinsic matrix to characterize the instantaneous displacement direction caused by the force. These parameters are combined into a high-dimensional "vibration fingerprint feature vector" as the input for the next step.

[0034] Step S300: Neural inference of collision events driven by physical constraints.

[0035] This step aims to analyze the vibration fingerprint feature vector extracted in step S200 to determine whether a collision has occurred and to estimate the key physical parameters of the collision. For this purpose, a pre-trained Physics-Aware Temporal Convolutional Network (Pa-TCN) is used; its structural diagram can be found in [link to diagram]. Figure 3 The core of this network is a feature extraction module consisting of multiple layers of dilated causal convolutions, such as... Figure 3 As shown, the module contains cascaded dilated convolutional layers with dilation rates of 1, 2, and 4, and fuses multi-scale features through residual connections. This structure can capture long-distance temporal dependencies in vibration signals with less computational resources, and is particularly effective for analyzing the generation and decay process of collision impacts.

[0036] Specifically, this Pa-TCN embeds a physical constraint layer to ensure that the network's inference results conform to basic physics principles, namely the impulse-momentum theorem. The network's output layer includes not only a Sigmoid activation function node that outputs the "collision probability," but also nodes for regressing the collision impulse vector J, the collision contact point location, and the collision time t_impact. The physical constraint layer works as follows: when the network initially determines that a collision may occur, this layer calculates the minimum theoretical impulse required to generate the vibration based on the input vibration energy spectral density and the estimated vehicle mass m_est. The specific calculation process is as follows: First, based on the law of conservation of energy, the kinetic energy generated by the collision will be partially converted into vibration energy of the vehicle body. During the calculation, the first step is to process the input vibration energy spectral density... In the preset vehicle resonance frequency band By performing frequency domain integration, the vibration energy density per unit volume can be obtained. The second step involves combining the preset vibration waves and the equivalent vehicle body volume coefficient. Calculate the total vibration energy The third step is to apply the momentum-kinetic energy relationship formula. (in For momentum, (for quality), and introduces an impulse transfer loss coefficient. (Used to compensate for inelastic deformation and heat loss during the collision process, typically ranging from 1.1 to 1.5), to deduce the minimum theoretical impulse required to generate this total vibration energy. The minimum theoretical impulse The calculation formula is:

[0037] in, This represents the minimum theoretical impulse; Indicates the impulse transmission loss coefficient; This represents the estimated mass of the target vehicle. This represents the total vibrational energy; Indicates the vibrational energy spectral density; and These represent the lower and upper limits of the preset vehicle resonance frequency band, respectively. This indicates the vibration wave and the equivalent vehicle body volume coefficient.

[0038] If the magnitude of the impulse vector J regressed by the network is significantly smaller than the minimum theoretical impulse, or if the impulse is insufficient to cause the vehicle to produce a rigid body displacement of more than one pixel, the physical constraint layer will determine that the signal is more likely to originate from a non-collision event (such as the vibration of a car door closing), and will force the final output collision probability value to be cleared to zero or apply a large penalty factor, thereby effectively filtering out false signals and reducing the false alarm rate.

[0039] Furthermore, such as Figure 3 As shown in the dual-vehicle association verification module at the bottom, when two suspected collision objects are detected simultaneously in the scene, the system performs dual-vehicle association verification. This verification process includes two dimensions of physical consistency checks: First, a time dimension check, calculating the time difference t_impact between the two vehicles at the moment of impact and determining whether it is less than a preset time threshold (e.g., 50ms) to confirm the synchronicity of the collision; second, a spatial dimension check, checking whether the directions of the collision impulse vectors J of the two vehicles satisfy Newton's third law, i.e., they are opposite to each other. Only when both the time difference and the impulse direction satisfy the above physical constraints is it finally determined to be a dual-vehicle collision event.

[0040] Step S400, refined trajectory repair based on conditional denoising diffusion model, the principle of which is illustrated in the diagram below. Figure 4 As shown.

[0041] like Figure 4 As shown, this step first defines a set of guiding conditions, specifically including: the pre-collision trajectory segment (T_pre), the post-collision trajectory segment (T_post), the collision time (t_impact) obtained by regression from step S300, and the collision impulse vector (J). These guiding conditions are input into a conditional denoising diffusion probability model (cDDPM) to control the direction of trajectory generation.

[0042] Specifically, the backdiffusion process of cDDPM begins with a pure Gaussian noise distribution. The model execution starts with a denoising step t=T, proceeds through t=T / 2, and continues iteratively until t=1. To ensure that the generated trajectory conforms to the laws of physical collision, this step constructs a dynamic perturbation field based on the collision impulse vector (J). At each step of the backdiffusion (each step injects a perturbation bias), this dynamic perturbation field is injected as an additional gradient term into the sampling process. Especially at the time step close to the collision moment (t_impact), this perturbation field forces a velocity abrupt change in direction and magnitude proportional to the impulse vector J, thereby breaking the tendency of conventional diffusion models to generate smooth curves and ensuring that the generated trajectory exhibits a physically accurate inflection point at the collision point.

[0043] After the diffusion process is complete, the model outputs multiple candidate trajectories in parallel. Subsequently, the system performs a scene geometric constraint filtering step. Using the spatial location information of static occluders (such as walls and pillars) identified in step S100, it checks whether each candidate trajectory spatially overlaps with an occluder. Trajectories that cross occluders are considered invalid and discarded. Finally, a weighted average is applied to all remaining valid candidate trajectories that meet the geometric constraints, resulting in the final refined reconstructed trajectory (including collision inflection points), which fills the motion gaps within the visual occlusion area.

[0044] like Figure 5 As shown, the complete process of the conditional denoising diffusion probability model gradually generating the repair trajectory in two-dimensional space is illustrated. The figure contains four time step panels, each corresponding to a different stage of the denoising process. At the initial time t=T, the noise trajectory points 301 are scattered in a Gaussian random distribution throughout the parking lot space, and the trajectory information is completely submerged by noise. At time t=T / 2, the noise points begin to gather towards the spatial region where the real trajectory is located, but still maintain a large degree of dispersion, only vaguely showing a trend direction. At time t=T / 4, the trajectory shape is initially formed, with scattered points closely arranged along the repair trajectory, and the position of the collision inflection point 303 is vaguely discernible. At the final time t=0, the model outputs the final repair trajectory 304, which is represented by a dashed line, completely connecting the known trajectory segment T_pre(306) before the collision and the known trajectory segment T_post(307) after the collision, and showing a clear discontinuous deflection feature at the collision inflection point 303. The dynamic perturbation field 302 is represented in the form of a gray-scale arrow field, with the arrows dense near the collision point and sparse in the distance, reflecting the guiding role of physical constraints on the denoising process. The occlusion column 305 and its projected shadow area mark the spatial extent of visual occlusion. The entire process visually demonstrates the denoising mechanism that gradually "crystallizes" the trajectory from pure noise into a precise trajectory containing collision dynamics characteristics.

[0045] Step S500: Accident responsibility map generation and evidence chain synthesis.

[0046] This step, based on the refined trajectory reconstructed in step S400 and the collision physics parameters regressed in step S300, automatically generates a structured accident analysis report and visualized evidence. First, the system quantifies the severity and morphology of the collision by calculating the spatial overlap area and relative velocity vector of the vehicles involved on the reconstructed trajectory at the moment of collision (t_impact). Next, based on the morphological characteristics of the reconstructed trajectory before and after the collision, it identifies driving behavior semantics such as "forced lane change" and "sudden reversing," and, combined with the direction of the collision impulse, determines the active and passive parties in the collision. Finally, the system synthesizes and outputs a multi-dimensional evidence package, which may include the original surveillance video, a slow-motion video overlaid with micro-motion magnification to visually demonstrate vehicle vibration, and a top-down trajectory reconstruction animation to clearly show the complete motion process of the vehicle within the obscured area and its dynamic behavior at the moment of collision.

[0047] Example 2 This embodiment provides a vehicle minor collision detection and trajectory reconstruction system in an occlusion-resistant environment. This system can be a software system deployed on a cloud server, or a hardware or software-hardware hybrid system integrated into an edge computing device. The system is used to execute the method described in Embodiment 1, and its internal structure and functional modules include: The panoramic perception and occlusion processing module is responsible for executing step S100 in the aforementioned method. Specifically, it is internally configured with computing units for running deep semantic segmentation and instance segmentation algorithms, capable of parsing the input monitoring video stream in real time and establishing a semantic spatiotemporal graph of the scene. This module also includes a database or memory structure for indexing and managing vehicle occlusion status, vehicle model information, quality estimates, and the locations of rigid connection features.

[0048] The micro-feature frequency domain analysis module is responsible for executing step S200 in the aforementioned method. Its configuration is used to perform parallel phase decomposition on the micro-motion observation window region determined by the panoramic perception and occlusion processing module. This module integrates or calls a vehicle model-modal frequency fingerprint library and runs an adaptive resonant frequency weighted network to filter and enhance the signal, ultimately extracting and generating a high-dimensional vibration fingerprint feature vector.

[0049] The neurophysical inference engine is responsible for executing step S300 in the aforementioned method. Its core consists of deploying one or more pre-trained physical perception temporal convolutional network models. This engine receives vibration fingerprint feature vectors as input, infers through model inference, and incorporates physical conservation laws such as the impulse-momentum theorem as hard constraints for verification. Finally, it outputs judgments regarding key physical parameters such as whether a collision event occurred, the time of occurrence, and the impulse vector.

[0050] The generative trajectory reconstruction server is responsible for executing step S400 in the aforementioned method. A conditional denoising diffusion probability model is deployed on it. Based on the collision event information output by the neurophysical inference engine, the server generatively repairs the vehicle trajectory within the occluded area, ensuring that the generated trajectory conforms to collision dynamics near the collision point and outputs high-precision vehicle trajectory data.

[0051] The intelligent liability determination and evidence collection terminal is responsible for executing step S500 in the aforementioned method. It can be a software application or web service with a graphical user interface. This terminal receives and integrates the reconstructed trajectory and collision parameters generated by other modules, generates a structured accident liability map through a built-in rule engine or algorithm, and synthesizes a visual evidence package including the original video, micro-motion enhanced video, and trajectory reconstruction animation, for traffic management personnel or relevant parties to review and use. Upon receiving the evidence package, the traffic control center can quickly determine liability and instruct vehicles at the scene to quickly evacuate, thereby restoring traffic flow.

[0052] Example 3 This embodiment illustrates the specific execution process of the method of the present invention in a particular application scenario with severe visual occlusion. For example... Figure 6 As shown, this embodiment uses an underground parking lot as a typical application scenario, demonstrating the complete spatial deployment and occlusion relationships from a bird's-eye view. The parking lot has multiple parking spaces 406 arranged along the top, bottom, left, and right sides, each marked with a diagonal line. A driving lane 407 is located in the center of the parking lot, with its center line marked by a dashed line, allowing for two-way traffic. Multiple load-bearing columns 402 are evenly distributed throughout the parking lot; their cross-sections are square and filled with dark gray, forming static occlusions in the scene. Four surveillance cameras 401 are installed on the ceiling at the four corners of the parking lot, with the field of view of each camera represented by a fan-shaped area. Due to the physical occlusion of the load-bearing columns 402, a blind spot 405 is formed behind the central column; this area is marked with a diagonal line, indicating a visual blind spot that the surveillance cameras cannot effectively cover.

[0053] In this scenario, a car (i.e. Figure 6 Vehicle A (marked 403) reverses out of the bottom parking space, its trajectory indicated by the dashed arrow, moving in an arc towards the driving lane. Simultaneously, a SUV (i.e., Figure 6Vehicle B (marked as 404) is traveling left along the main driving lane, its trajectory marked with a dotted-line arrow. The trajectories of the two vehicles intersect near the blind spot, forming a potential collision zone 408, emphasized by a dashed circle and a cross mark. This visual obstruction prevents the surveillance video from directly recording any collision or scratch. Due to the lack of direct visual evidence and the weak physical characteristics of low-speed scratches, traditional detection methods based on image deformation or motion anomalies fail in this scenario, making it impossible to determine whether an accident has occurred and attribution of responsibility. Therefore, the method of this invention is necessary.

[0054] First, the system executes step one, processing the video stream of the scene. The system identifies the load-bearing pillar as a static obstruction through semantic segmentation, and performs instance segmentation on vehicle A (identified as a sedan, with a mass estimate m_est_A) and vehicle B (identified as a SUV, with a mass estimate m_est_B). When the two vehicles approach the pillar and their occlusion rate exceeds a preset threshold, the system marks them as "occlusion risk targets." Subsequently, the system identifies the visible shark fin antenna on the roof of vehicle A and the visible end of the roof rack on the roof of vehicle B as their respective rigid connection feature parts (RCF), and establishes corresponding micro-motion observation windows (ROIs).

[0055] Next, the system executes step two, analyzing the image sequences of the two ROIs. The system performs complex manipulable pyramid decomposition on each ROI and then proceeds to the phase analysis process. At this point, the adaptive resonance frequency weighted network based on the vehicle model modal fingerprint database comes into play. This network utilizes the vehicle model and mass information obtained in step one to generate a weighted mask for vehicle A (sedan) that emphasizes the higher resonance frequencies of its body structure, and a weighted mask for vehicle B (SUV) that emphasizes the lower resonance frequencies corresponding to its larger body mass. Through this adaptive weighting process, the system can directionally enhance the weak vibration signals transmitted from a potential collision to the two different vehicle bodies while suppressing the background noise of different frequencies generated by the respective engine idling speeds. After processing, the system extracts a vibration event with similar start times, low energy, but a clear signal-to-noise ratio from the phase time-series signals of both vehicles and constructs their respective vibration fingerprint feature vectors.

[0056] Subsequently, in step three, the two vibration fingerprint feature vectors are fed into a Physical Aware Temporal Convolutional Network (Pa-TCN) for inference. The network first determines that both vectors indicate a high probability of a collision event. The physical constraint layer verifies this determination, confirming that the vibration energy and the theoretical collision impulse corresponding to the estimated mass of the two vehicles are within a reasonable range. Crucially, the network performs a correlation check on the two events, finding that the difference between the start timestamps of the two vibration signals is less than a preset 40-millisecond threshold. Furthermore, analysis of the main vibration direction shows that the vibration direction of vehicle A's antenna and the vibration direction of vehicle B's roof rack are essentially opposite, conforming to the law of action and reaction. Based on this, the system ultimately determines that a collision occurred between the two vehicles at time t_impact and regresses the corresponding collision impulse vector J.

[0057] In step four, the system uses the Conditional Denoising Diffusion Probability Model (cDDPM) to repair the trajectories of the two vehicles within the pillar obstruction area. The model's input conditions include: the visible trajectory segment T_pre before the two vehicles enter the obstruction area, the visible trajectory segment T_post after they leave the obstruction area, the collision time t_impact, and the collision impulse vector J. During the inverse denoising trajectory generation process, a dynamic perturbation field constructed from the impulse J is applied in the sampling step. This causes a non-smooth inflection point at time t_impact in the generated trajectory. The angle and position changes of this inflection point accurately reflect the instantaneous change in the motion state of the two vehicles due to the collision impulse. The repaired trajectory clearly shows that during reversing, vehicle A's rear profile encroaches on vehicle B's normal driving path and contacts the side of vehicle B at this point.

[0058] Finally, in step five, the system integrates all information to generate a chain of evidence. The output includes an original video with the ROI labeled, a slightly magnified video (showing that the antenna of vehicle A and the roof rack of vehicle B vibrate synchronously at the moment of collision), and a top-down trajectory reconstruction animation. This animation clearly demonstrates the complete spatiotemporal relationship and collision point of the two reconstructed trajectories within the obstructed area, providing objective and physically consistent technical evidence for determining that vehicle A bears primary responsibility.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for capturing and reconstructing vehicle minor collisions in an environment with resistance to occlusion, characterized in that, Includes the following steps: Step 1: Construct a vehicle semantic spatiotemporal graph for dynamic occlusion perception, and determine the micro-motion observation window based on the graph: real-time analysis of roadside or parking lot monitoring video streams to identify static occlusions in the scene; Identify and track target vehicles to obtain their dynamic trajectory information; simultaneously, identify the vehicle type, estimated mass, and rigid connection features on the vehicle body; spatiotemporally correlate static occlusion information with the target vehicle's dynamic trajectory information to construct a semantic spatiotemporal graph describing the dynamic relationships between elements in the scene; when it is determined based on this graph that the target vehicle is at risk of being occluded, the currently visible rigid connection features are identified as micro-motion observation windows; Step 2: Frequency domain micro-motion feature extraction based on phase Euler amplification: The image sequence of the micro-motion observation window is decomposed into complex manipulable pyramids to obtain the phase time-series signal. The phase time-series signal is weighted using a weighted frequency band mask generated based on the vehicle type classification and quality estimation value to extract the vibration fingerprint feature vector. Step 3: Physically Constrained Collision Event Neural Inference: Input the vibration fingerprint feature vector into the physical perception temporal convolutional network, use the impulse-momentum theorem as a physical constraint to determine whether a collision event exists, and regress to obtain the collision impulse vector and the time of collision occurrence. Step 4: Refined trajectory repair based on conditional denoising diffusion model: Using the conditional denoising diffusion probability model, the trajectory segments before and after the collision in the dynamic trajectory information and the collision impulse vector are used as conditions to generate the motion trajectory of the target vehicle in the occluded area that conforms to the collision dynamics. Step 5: Accident Liability Map Generation and Evidence Chain Synthesis: Based on the reconstructed motion trajectory and collision physical parameters, generate accident evidence data for liability determination.

2. The method according to claim 1, characterized in that, The rigid connection feature refers to a component that is directly connected to the vehicle chassis through a rigid structure, including at least one of the following: roof shark fin antenna, roof rack, A-pillar, B-pillar, or C-pillar upper edge.

3. The method according to claim 1, characterized in that, The physical-aware temporal convolutional network includes a physical constraint layer, and the physical constraint-driven collision event neural inference steps further include: The physical constraint layer calculates the minimum theoretical impulse required to generate the vibration based on the vibration energy spectral density characterized by the vibration fingerprint feature vector and the estimated mass of the target vehicle; and, The physical constraint layer compares the collision impulse vector obtained by the regression of the physical-aware temporal convolutional network with the minimum theoretical impulse. If the magnitude of the collision impulse vector is less than a preset physical threshold, it is determined that the collision event does not exist.

4. The method according to claim 3, characterized in that, When there are two vehicles suspected of colliding in the surveillance video stream, the physical constraint-driven collision event neural inference step further includes: Only when the difference in the start time of vibration between the two vehicles is less than a predetermined time threshold, and the main vibration directions indicated by the vibration fingerprint feature vectors of the two vehicles are opposite, will it be ultimately determined as a two-vehicle collision event.

5. The method according to claim 1, characterized in that, The refined trajectory repair step based on the conditional denoising diffusion model further includes: Through a reverse diffusion process, Gaussian noise is gradually denoised to generate trajectory points within the occluded region; and, In each step of the reverse diffusion process, a dynamic perturbation field based on the collision impulse vector and the moment of collision is introduced as an additional conditional bias term. When the dynamic perturbation field approaches the moment of collision, it applies a velocity abrupt gradient related to the magnitude and direction of the collision impulse vector to the sampling process, so as to guide the generated trajectory to produce a discontinuous trajectory deflection at the collision point.

6. The method according to claim 5, characterized in that, The refined trajectory repair step based on the conditional denoising diffusion model also includes: Generate multiple candidate trajectories; and, Using the static occlusion as a geometric constraint, the multiple candidate trajectories are filtered to eliminate invalid trajectories that cross the static occlusion, and the remaining valid trajectories are weighted and averaged to output the final reconstructed trajectory.

7. The method according to claim 1, characterized in that, The process of generating the weighted frequency band mask in step two includes: Pre-build a vehicle-modal frequency fingerprint library that stores resonance frequency distribution data of different vehicle body structures; Based on the vehicle type classification results and mass estimation values ​​of the target vehicle, the corresponding resonant frequency distribution is retrieved from the vehicle type-modal frequency fingerprint database; and, The weighted frequency band mask is generated, wherein the weighted frequency band mask is configured to assign a first weight to the collision resonance frequency band of the target vehicle and a second weight to the engine idling or background noise frequency band, wherein the first weight is higher than the second weight.

8. The method according to claim 1, characterized in that, The vibration fingerprint feature vector includes at least one of the following parameters: vibration energy spectral density, vibration attenuation rate, vibration principal direction, and phase consistency.

9. The method according to claim 1, characterized in that, Step five further includes: quantifying the severity of the collision based on the reconstructed motion trajectory and collision physical parameters, generating structured data and visual evidence for accident liability determination, and sending the structured data to the traffic management platform to trigger accident warnings or assist in the liability determination process; The visual evidence includes: a slow-motion video overlaid with a micro-motion magnification effect, and a top-down trajectory reconstruction animation showing the complete movement of the target vehicle within the occluded area.

10. A vehicle minor collision capture and trajectory reconstruction system in an environment with resistance to occlusion, characterized in that, include: The panoramic perception and occlusion processing module is configured to analyze roadside or parking lot monitoring video streams in real time, identify static occlusions in the scene, identify and track target vehicles to obtain dynamic trajectory information, vehicle type classification and mass estimation value of target vehicles, identify rigid connection feature parts on the vehicle body, and determine the currently visible rigid connection feature parts as micro-motion observation windows when it is determined that the target vehicle has an occlusion risk. The micro-feature frequency domain analysis module is configured to perform complex manipulable pyramid decomposition on the image sequence of the micro-motion observation window to obtain the phase time sequence signal, and to perform weighted processing on the phase time sequence signal using a weighted frequency band mask generated based on the vehicle classification and quality estimation value to extract the vibration fingerprint feature vector; The neurophysical inference engine is configured to input the vibration fingerprint feature vector into the physical perception temporal convolutional network, use the impulse-momentum theorem as a physical constraint to determine whether a collision event exists, and regress to obtain the collision impulse vector and the time of collision occurrence. A generative trajectory reconstruction server is configured to use a conditional denoising diffusion probability model, based on the trajectory segments before and after the collision in the dynamic trajectory information and the collision impulse vector, to generate a motion trajectory of the target vehicle within the occluded area that conforms to collision dynamics; and... The intelligent liability determination and evidence collection terminal is configured to generate accident evidence data for determining accident liability based on the reconstructed motion trajectory and collision physical parameters.

Images