A litter motion analysis system and method applied to a highway

Abstract

The application discloses a litter motion analysis system and method applied to an expressway, and particularly relates to the field of traffic monitoring, and is used for solving the problem that the prior art can only sporadically detect the position of suspected litter under low frame rate monitoring and is difficult to continuously analyze the motion trajectory, is through the following mode: acquiring a low frame rate video sequence to extract a candidate trajectory segment set, calculating a direction fitting degree parameter and an out-of-bound risk parameter based on calibration parameters and road geometry, inputting the decision tree model associated segments, detecting displacement abnormalities to predict and complete unassociated segments, converting image coordinates to generate a motion path, extracting direction crossing and staying positions, checking time continuity and lane boundary constraints to update associated conditions and correct results, so that the motion process of the litter is finely described and fully utilized, and the accuracy and efficiency of the expressway abnormal event processing are improved.

Description

Technical Field

[0001] This invention relates to the field of traffic monitoring, and more specifically, to a system and method for analyzing the motion of projectiles on highways. Background Technology

[0002] Fixed video surveillance equipment is typically deployed along highways to capture road surface images and assist in detecting abnormal situations such as spilled or fallen objects. Existing technologies, such as CN119625663A, "An Algorithm for Low Frame Rate Detection of Spilled Objects on Highways Based on Image Segmentation," extract frames from the surveillance video, automatically identify road areas, and then segment the video into blocks. Foreground detection is only performed when there are significant changes in local image blocks. The foreground results from each image block are then combined into a complete foreground image. Finally, suspected targets are identified and reported. This type of solution can detect and identify spilled objects on highways under conditions of low frame rate monitoring and limited processing resources, determining whether a suspicious spilled object exists at a specific location at a specific time, and achieving a judgment on the presence and approximate location of spilled objects.

[0003] However, existing low frame rate detection methods based on image segmentation mostly focus on identifying which area of ​​a frame contains a suspected object, lacking the ability to continuously analyze the object's movement process. When an object rapidly passes through the monitoring screen and appears in different image blocks at different times, the system only records scattered local foreground results appearing in certain frames. The software lacks a dedicated mechanism to link the multiple appearances of the same object at different times and in different blocks, nor does it organize these scattered detection results into a complete trajectory from entering the monitoring field of view to its movement or stillness. Therefore, in a low frame rate monitoring environment, even if the system can determine that a suspected object has appeared on a certain road segment, it struggles to further support a detailed analysis of the object's movement process. It cannot clearly describe how the object moves along the timeline on the road, how it crosses different lanes, and its final stopping position. Consequently, it fails to meet the requirement of a comprehensive characterization and utilization of the object's movement trajectory for a highway object movement analysis system and method.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a system and method for analyzing the motion of projectiles on highways. This system extracts a set of candidate trajectory segments from low-frame-rate video sequences, inputs these segments into a decision tree model based on calibration parameters and road geometry calculations of direction fit parameters and boundary risk parameters, detects displacement anomalies, predicts and completes unrelated segments, transforms image coordinates to generate motion paths, extracts directions, and identifies crossing and stopping positions. Finally, it verifies temporal continuity and lane boundary constraints, updates the correlation conditions, and corrects the results. This approach addresses the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for analyzing the motion of projectiles on highways, comprising the following steps:

[0008] S1. Obtain low frame rate video sequences from highway monitoring equipment, segment each frame into road regions, extract suspected foreground blocks of thrown objects within the road regions, record time markers, image coordinates, and appearance features to form a set of candidate trajectory segments;

[0009] S2. For the segments to be associated and the end segments of the target temporary trajectory in the candidate trajectory segment set, convert the image coordinates into road coordinates based on the calibration parameters and road geometry, calculate the direction fit parameter and the boundary risk parameter, input the two parameters into the pre-trained machine learning model to obtain the trajectory association confidence index, and according to whether the trajectory association confidence index meets the association conditions, merge the segments to be associated into the target temporary object trajectory or mark them as unassociated segments.

[0010] S3. When the displacement of the object in the road coordinate system of the adjacent frame exceeds the preset threshold, the image coordinate range of the next moment is predicted based on the road coordinates and movement direction of the end segment of the temporary object trajectory. Within the range, the association judgment in S2 is performed first to complete the unassociated segments and generate the object image trajectory.

[0011] S4. Based on the calibration parameters and road geometry, convert the image coordinates in the trajectory of the thrown object into road coordinates to obtain the motion path of the thrown object in the road coordinate system. Determine the motion direction, lane crossing and stopping position of the thrown object from the motion path to generate the motion analysis results of the thrown object.

[0012] S5. Based on the temporal continuity of the motion path in the road coordinate system and the lane boundary constraints, the motion analysis results of the thrown object are verified. When the motion path does not match the candidate trajectory segment in time or position, the trajectory association conditions are updated to re-associate and complete the candidate trajectory segment, thus obtaining the corrected motion analysis results of the thrown object.

[0013] In a preferred embodiment, step S1 includes acquiring a low frame rate video sequence from a highway monitoring device, decomposing the video sequence into an image frame sequence using a video decoding module, applying a road semantic segmentation network to each image frame to generate a mask image to preserve the road area image, calculating a difference image by comparing it with a background reference image using a Gaussian mixture model and extracting suspected foreground blocks of thrown objects through connected component analysis, calculating the geometric center of the minimum bounding box of each foreground block as image coordinates and extracting a high-dimensional feature vector output by a convolutional neural network as an appearance feature vector, integrating the time-stamped image coordinates and appearance feature vectors to form candidate trajectory segments and adding them to the candidate trajectory segment set in chronological order.

[0014] In a preferred embodiment, step S2 specifically includes selecting segments to be associated from the candidate trajectory segment set and maintaining the target temporary trajectory end segments; converting image coordinates into road coordinates based on camera calibration parameters and road geometry information to calculate displacement vectors; calculating direction fit parameters to evaluate the degree of conformity between motion and road direction; calculating boundary risk parameters to quantify the possibility of lane migration; inputting the direction fit parameters and boundary risk parameters into a pre-trained decision tree model to obtain a trajectory association confidence index; and merging the segments to be associated into the temporary thrown object trajectory or marking them as unassociated segments according to whether the trajectory association confidence index meets the association conditions.

[0015] In a preferred embodiment, the calculation of the directional fit parameter includes obtaining the tangent direction vector of the highway centerline at the road coordinate point of the end segment of the target temporary trajectory based on road geometry information as the longitudinal reference vector and constructing its orthogonal lateral reference vector. The displacement vector is projected onto the longitudinal and lateral reference vectors and decomposed by the dot product algorithm to obtain the longitudinal and lateral components. The sign of the longitudinal component is compared with the highway driving direction, and the ratio of the absolute value of the longitudinal component to the lateral component is compared. If they are consistent and the ratio is in the first ratio range, the directional fit parameter is of the first level; if they are consistent and the ratio is in the second ratio range, it is of the second level; otherwise, it is of the third level.

[0016] In a preferred embodiment, the calculation of the boundary crossing risk parameter includes dividing the lateral coordinates into lane intervals according to the lane boundary geometry, determining the lane number of each of the two road coordinate points, calculating the minimum distance from each point to the nearest lane boundary, using the line connecting the two points as an approximate line segment of the motion trajectory to detect the intersection of the line segment with the lane boundary line. If the lane numbers are the same and the distance is in the first distance interval and there is no intersection, the boundary crossing risk parameter is of the first level. If they are the same and the distance is in the second distance interval or there is contact with a single boundary, it is of the second level. If they are different or there is intersection with multiple boundaries, it is of the third level.

[0017] In a preferred embodiment, step S3 specifically includes checking the longitudinal displacement of adjacent segments of the road coordinate system in the temporary debris trajectory to determine the missing segments, constructing a motion prediction vector to simulate the expected motion, calculating the predicted road coordinate points to estimate the next position, mapping the predicted road coordinate points to image coordinates to form a search area, filtering unrelated segments in the search area and performing association judgment to complete the temporary debris trajectory, and generating a continuous debris image trajectory.

[0018] In a preferred embodiment, step S4 specifically includes reading the image coordinates in the trajectory of the thrown object in chronological order, converting the image coordinates into road coordinates to obtain the motion path, calculating the longitudinal displacement of adjacent road coordinate points from the motion path to determine the motion direction, statistically analyzing the crossing situation based on the road coordinate points and lane boundary positions to identify lane crossings, selecting the road coordinate point with the largest time stamp as the stopping position, and combining the motion path, motion direction, lane crossing situation, and stopping position into the thrown object motion analysis result.

[0019] In a preferred embodiment, step S5 specifically includes comparing the time stamps of adjacent road coordinate points to check the temporal continuity of the motion path, analyzing the positional relationship between the connecting line segment and the lane boundary to perform spatial consistency verification, scanning the marked area to determine if the time or position is inconsistent, updating the trajectory association conditions to adjust the confidence index threshold parameter constraints and weights, reselecting candidate trajectory segments within the abnormal time range to perform association judgment and complete the segments, and converting the new image coordinates to form the corrected motion path and the analysis results of the falling object motion.

[0020] A projectile motion analysis system for use on highways includes:

[0021] Trajectory Segment Extraction Module: Segment road regions from low frame rate video sequences and extract foreground blocks of suspected thrown objects, and record time-stamped image coordinate appearance features to form a set of candidate trajectory segments;

[0022] The segment association judgment module calculates the direction fit parameter and the boundary risk parameter based on the calibration parameters, road geometry transformation image coordinates, and inputs them into the machine learning model to obtain the confidence index. Based on the confidence index, the segments to be associated are merged into the temporary falling object trajectory or marked as unassociated.

[0023] Missing Completion Prediction Module: When anomalies in the trajectory displacement of a temporary falling object are detected, the image coordinate range is predicted based on the direction of motion of the end coordinate. Within the range, association judgment is performed to complete the unassociated segments and generate the falling object image trajectory.

[0024] Path feature analysis module: The motion path is obtained by transforming the image coordinates of the road geometry using calibration parameters. The direction of the thrown object's movement, the position of the lane it crosses and stops at are determined from the path, and motion analysis results are generated.

[0025] Consistency result correction module: Based on the motion path time continuity lane boundary constraint verification analysis results, when they do not match the candidate trajectory segments, update the association conditions and re-associate to complete the corrected motion analysis results.

[0026] The technical effects and advantages of the projectile motion analysis system and method applied to highways according to the present invention are as follows:

[0027] This invention unifies scattered detection results into a complete trajectory in resource-constrained environments, avoiding the limitations of existing technologies that only locate suspected targets in a single frame. This ensures a comprehensive depiction of the spatiotemporal dynamics of the object from its appearance to its rest, improving the timeliness and accuracy of anomaly detection. Secondly, through predictive completion and iterative verification mechanisms, it compensates for fragment loss and correlation errors caused by low frame rates, making the analysis results highly consistent with actual road behavior, reducing misjudgments and optimizing traffic safety management decisions. Finally, the coordinate transformation and boundary constraint application under overall coordination provide risk assessment of object deviation and lane migration sequence information, supporting subsequent uses such as early warning reporting or path optimization, significantly improving the efficiency and reliability of highway monitoring systems. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating a method for analyzing the motion of projectiles on highways according to the present invention.

[0029] Figure 2 This is a schematic diagram of the structure of a projectile motion analysis system applied to highways according to the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example 1: Figure 1 This invention provides a method for analyzing the motion of projectiles on highways, comprising:

[0032] S1. Obtain low frame rate video sequences from highway monitoring equipment, segment each frame into road regions, extract suspected foreground blocks of thrown objects within the road regions, record time markers, image coordinates, and appearance features to form a set of candidate trajectory segments;

[0033] S2. For the segments to be associated and the end segments of the target temporary trajectory in the candidate trajectory segment set, convert the image coordinates into road coordinates based on the calibration parameters and road geometry, calculate the direction fit parameter and the boundary risk parameter, input the two parameters into the pre-trained machine learning model to obtain the trajectory association confidence index, and according to whether the trajectory association confidence index meets the association conditions, merge the segments to be associated into the target temporary object trajectory or mark them as unassociated segments.

[0034] S3. When the displacement of the object in the road coordinate system of the adjacent frame exceeds the preset threshold, the image coordinate range of the next moment is predicted based on the road coordinates and movement direction of the temporary object trajectory end segment. Within the range, the association judgment in step two is performed first to complete the unassociated segments and generate the object image trajectory.

[0035] S4. Based on the calibration parameters and road geometry, convert the image coordinates in the trajectory of the thrown object into road coordinates to obtain the motion path of the thrown object in the road coordinate system. Determine the motion direction, lane crossing and stopping position of the thrown object from the motion path to generate the motion analysis results of the thrown object.

[0036] S5. Based on the temporal continuity of the motion path in the road coordinate system and the lane boundary constraints, the motion analysis results of the thrown object are verified. When the motion path does not match the candidate trajectory segment in time or position, the trajectory association conditions are updated to re-associate and complete the candidate trajectory segment, thus obtaining the corrected motion analysis results of the thrown object.

[0037] Highway monitoring equipment captures road surface images via fixed video feeds to monitor for abnormal events such as falling objects in real time, thereby maintaining traffic safety. However, existing low frame rate detection technologies mainly rely on image segmentation methods, which can only identify the approximate location of suspected falling objects within a single frame, lacking the ability to continuously track the movement of falling objects. When a falling object rapidly crosses the frame, the system only records scattered foreground blocks, unable to correlate these scattered results into a complete trajectory, resulting in the inability to analyze its direction of movement, lane crossing, and stopping position. This limitation hinders a comprehensive assessment of falling object events and fails to meet the needs of highway management for refined motion analysis. Therefore, step S1 aims to construct spatiotemporally continuous basic data by acquiring video sequences and extracting candidate trajectory segments, providing reliable input for subsequent trajectory association and motion path generation, and ensuring improved dynamic monitoring of falling objects even in resource-constrained environments.

[0038] S1.1 Video Sequence Acquisition and Decoding.

[0039] Highway monitoring equipment uses low frame rate video sequences as input, capturing road surface dynamics. The video data stream is read directly from the monitoring equipment and processed frame-by-frame by a video decoding module, decomposing the video sequence into independent image frame sequences, each corresponding to a specific time point. The decoding process uses a standard video codec to process the raw compressed data, outputting an image matrix in red-green-blue or grayscale format, with the matrix size matched to the monitoring equipment's resolution. After decoding, all image frames are arranged chronologically to form an ordered frame sequence for road area segmentation.

[0040] S1.2 Road area segmentation.

[0041] Based on the decoded image frame sequence, the road surface needs to be isolated to focus on falling object detection. The current image frame is extracted from the ordered frame sequence, and a road semantic segmentation network (pre-trained on a highway scene dataset) is applied to perform pixel-level classification of the entire image. This network can distinguish between road pixels and non-road pixels, outputting a binary mask image where road pixels are marked with a value of one and non-road pixels with a value of zero. This mask image is then multiplied pixel-by-pixel with the original image frame, retaining only pixel regions with a mask value of one, thus removing irrelevant parts such as the sky, guardrails, and green belts, generating a cropped image containing only the road region. This cropped image maintains the original coordinate system for easy subsequent coordinate calculation. After segmentation, the road region image serves as the input for foreground detection.

[0042] S1.3 Foreground detection and connected component analysis.

[0043] In the road area image generated in the previous step, potential changes need to be identified to extract suspected debris. A background reference image of the road area is established, and the current road area image is compared pixel by pixel using a Gaussian mixture model algorithm to calculate the difference image. Locations where the pixel difference exceeds a preset threshold are marked as foreground candidate regions. The difference image is calculated by taking the absolute value of the difference between the pixel intensity value of the current road area image and the corresponding pixel intensity value of the background reference image.

[0044] A connected component analysis algorithm is applied to the foreground candidate regions. The difference image is traversed to find connected components formed by adjacent pixels. The area and contour shape features of each component are calculated. Only connected components with an area greater than a minimum area threshold and a contour rectangularity close to the typical shape of a thrown object are retained as suspected foreground blocks. The contour rectangularity is calculated by dividing the area of ​​the connected component pixels by the product of the width and height of its enclosing rectangle. These foreground blocks represent possible locations of thrown objects, avoiding noise interference. After analysis, the foreground block list is prepared for feature extraction.

[0045] S1.4 Foreground block coordinates and appearance feature extraction.

[0046] For each suspected foreground object block, its position and description are further quantified to support trajectory association. A minimum bounding box is calculated for each foreground block, defined as the smallest rectangular region covering all pixels of the foreground block. The geometric center of the minimum bounding box is taken as the image coordinates of that foreground block, expressed in pixels, with the horizontal axis as the x-axis and the vertical axis as the y-axis. Simultaneously, appearance features are extracted. A convolutional neural network outputs a high-dimensional feature vector from the pixel region of the foreground block. This vector captures information such as grayscale distribution, color distribution, and texture pattern; the dimension is fixed to a preset value to ensure consistency. After extraction, each foreground block is associated with its image coordinates and high-dimensional feature vector.

[0047] S1.5 Candidate trajectory segment formation and set maintenance.

[0048] The timestamp of the current image frame is integrated with the image coordinates and appearance feature vectors of each foreground block to form an independent candidate trajectory segment. This segment serves as the basic unit for recording the appearance information of the thrown object in a single frame. The timestamp is derived from the video frame metadata and represents an absolute timestamp. All image frames are processed frame by frame, and the generated candidate trajectory segments are added to the candidate trajectory segment set in ascending order of timestamp. This set is stored in a list structure for easy traversal and association in subsequent steps. The maintenance process ensures that the set completely covers the video sequence without any missing segments. After the segment set is constructed, it is provided to step S2 as a data source to be associated.

[0049] Step S1 involves decoding images frame-by-frame from a low frame rate video sequence, performing road region segmentation to focus on road surface pixels, and extracting suspected foreground blocks of thrown objects using background modeling and connected component analysis. The time stamp, image coordinates, and appearance feature vector of each block are recorded, forming an ordered set of candidate trajectory segments. This set comprehensively captures the spatiotemporal information of the thrown object in the video, avoiding misjudgments due to non-road interference and noise, and laying the foundation for trajectory association. The final output serves as the input for step S2, ensuring seamless integration with subsequent road coordinate-based association judgments, enabling continuous analysis of the thrown object's movement process, and improving the accuracy and efficiency of handling abnormal events on highways.

[0050] Step S1 has generated a set of candidate trajectory segments, recording time-stamped image coordinates and appearance feature vectors, providing basic data for trajectory construction. However, in low frame rate environments, these segments remain fragmented, lacking a unified association mechanism, and cannot form temporary object trajectories. Therefore, step S2 calculates direction fit parameters and boundary crossing risk parameters through coordinate transformation, and inputs them into a decision tree model to obtain a trajectory association confidence index. This aims to achieve accurate segment attribution, ensure the continuity of temporary object trajectories, support subsequent prediction completion and motion path analysis, and improve the comprehensiveness of object event assessment.

[0051] S2.1 Selection of segments to be associated and maintenance of temporary trajectories.

[0052] Based on the candidate trajectory segment set generated in step S1, potential matches need to be filtered to construct a continuous trajectory. From the candidate trajectory segment set, unassigned segments to be associated are selected in time-stamped order. These segments contain time-stamped image coordinates and appearance feature vectors. Simultaneously, a target temporary trajectory end segment is maintained for each emerging temporary projectile trajectory. This end segment is the most recently associated candidate trajectory segment, recording its time-stamped image coordinates and appearance feature vector. If no existing temporary projectile trajectory exists, a new trajectory is initialized with the earliest unassigned segment and set as the target temporary trajectory end segment. A linked list structure is used to store the temporary projectile trajectories to ensure fast access to the end segments. After selection and maintenance, pairs of segments to be associated and target temporary trajectory end segments are provided for coordinate transformation.

[0053] S2.2 Coordinate Transformation and Displacement Vector Calculation.

[0054] Based on the selected segments to be associated and the target temporary trajectory's end segment, a unified coordinate system is needed to quantify the motion. Using camera calibration parameters and road geometry information, the image coordinates of the segment to be associated and the target temporary trajectory's end segment are mapped to a road coordinate system. This coordinate system uses the highway centerline as the longitudinal axis and the vertical direction as the lateral axis. Each mapping generates a road coordinate point, including longitudinal and lateral coordinates. The displacement vector between the two road coordinate points is calculated by subtraction. Its longitudinal component is the difference between the longitudinal coordinates of the road coordinate point of the segment to be associated and the longitudinal coordinate of the road coordinate point of the target temporary trajectory's end segment; its lateral component is the difference between the lateral coordinates of the road coordinate points of the segment to be associated and the lateral coordinates of the road coordinate point of the target temporary trajectory's end segment. After the transformation and calculation, the road coordinates and displacement vector are ready for parameter calculation.

[0055] S2.3 Directional fit parameter calculation.

[0056] Based on the road coordinates obtained in the previous step, in order to assess whether the segment to be associated conforms to the typical motion pattern of a falling object on a highway, that is, mainly moving along the road direction rather than deviating significantly, it is necessary to calculate the directional fit parameter to quantify the degree of fit between the displacement vector and the road direction, thereby providing a basis for directional consistency for trajectory association.

[0057] First, at the road coordinate points of the end segment of the target temporary trajectory, the tangent direction vector of the highway centerline is obtained based on road geometry information, serving as the longitudinal reference vector, and its orthogonal lateral reference vector is constructed. The displacement vector is projected onto the longitudinal and lateral reference vectors, and the dot product algorithm is used to decompose them into longitudinal and lateral components. The longitudinal component is calculated by the dot product of the displacement vector and a unit longitudinal reference vector, and the lateral component is calculated by the dot product of the displacement vector and a unit lateral reference vector. Then, the sign of the longitudinal component is compared with the highway driving direction, and the ratio of the absolute value of the longitudinal component to the lateral component is compared: if they are consistent and the ratio is in the first ratio range, the direction fit parameter is at level one; if they are consistent and the ratio is in the second ratio range, it is at level two; otherwise, it is at level three. This parameter is stored as an integer and is dimensionless. After calculation, it is input into the model along with the boundary risk parameter.

[0058] The first ratio interval represents cases where the directions are highly consistent. This first interval is defined as starting from zero and extending to an upper limit less than half of the predicted direction, for example, no more than about 30% of the total range. It covers the range of samples observed in actual tests where the directions are visually consistent. The second ratio interval represents cases where the directions are still largely consistent but have deviated somewhat. This second interval is defined as a range greater than the upper limit of the first interval but less than or equal to about half of it. It covers samples where the deviation is acceptable in actual tests but has affected the judgment of the motion trend. The specific upper limits and dividing points of the two ratio intervals were determined by calculating the angle ratio between the actual and predicted directions after collecting a large number of highway debris trajectory samples, and by combining statistical distribution and manually labeled degree of consistency.

[0059] S2.4 Calculation of boundary crossing risk parameters.

[0060] For road coordinates and lane boundaries, the probability of lane migration needs to be quantified. First, based on the geometric description of the lane boundaries, the lateral coordinates are divided into lane intervals. The lane numbers of each road coordinate point are determined, and the minimum distance from each point to the nearest lane boundary is calculated using the Euclidean distance algorithm. Then, the line connecting the two points is used as an approximate line segment of the trajectory, and the intersection of the line segment with the lane boundary line is detected: if the lane numbers are the same, the distance is in the first distance interval, and there is no intersection, the risk parameter for lane migration is at level one; if they are the same, the distance is in the second distance interval, or there is contact with a single boundary, it is at level two; if they are different or there is intersection with multiple boundaries, it is at level three. This parameter is recorded as an integer level and is dimensionless. After the calculation is completed, the model input is prepared.

[0061] The first distance interval is used to characterize situations where the thrown object is close to the lane boundary. The normalized distance for the first interval is defined as starting from zero and extending to a small upper limit, for example, not exceeding approximately 30% of the normalized lane width. This corresponds to scenarios where the thrown object is close to the lane boundary or has a potential risk of crossing the boundary. The second distance interval is used to characterize situations where the thrown object is at a moderate distance from the lane boundary. The normalized distance for the second interval is defined as being greater than the upper limit of the first interval but less than or equal to approximately half of it. This corresponds to scenarios where the thrown object is still within the current lane but has clearly approached the boundary. The specific upper limits of the first and second distance intervals are determined by collecting lane widths and thrown object trajectories from different road segments, normalizing the minimum lateral distance between the trajectory and the lane boundary according to the lane width, and then performing distribution analysis based on accident statistics and on-site video playback results to select boundary points that can effectively distinguish between low-risk approach, medium-risk approach, and high-risk boundary crossing or single boundary contact.

[0062] S2.5 Machine Learning Model Input and Trajectory Association Confidence Index Acquisition.

[0063] Based on the directional fit parameter and the cross-boundary risk parameter, it is necessary to predict the association confidence level. These two parameters are fed as input vectors into a pre-trained decision tree model. During the training phase, the model uses historical trajectory samples to label whether trajectories are the same, learning the decision path between parameters and association relationships. During runtime, the model traverses the branches and outputs a trajectory association confidence index based on the parameter levels. This index is a dimensionless value between zero and one level, representing the probability of association. Once obtained, it is used for association determination.

[0064] The decision tree model consists of a root node, internal nodes, and leaf nodes. The root node is the initial split point, internal nodes represent feature splitting decisions, and leaf nodes output classification results or confidence indices. In this invention, the decision tree model adopts a binary tree structure. Each internal node splits based on input features, which are orientation fit parameters and out-of-bounds risk parameters, both of which are dimensionless integer values. The model has no fixed depth; instead, the height and connections of the tree are dynamically determined during training. Each node connects to two child nodes until a stopping condition is met.

[0065] The training process of the decision tree model is divided into three stages: data preparation, tree construction, and optimization. First, in the data preparation stage, samples of thrown object trajectories are collected from historical surveillance videos of highways. These samples are extracted from actual road events, including the complete sequence of thrown objects from their appearance to their resting state. For each pair of adjacent trajectory segments, their directional fit parameter and boundary crossing risk parameter are calculated as input features, and they are manually labeled as belonging to the same thrown object trajectory as output labels. The labels are binary values ​​indicating association or non-association. The number of samples is no less than a preset scale to cover various thrown object speeds, directions, and lane migration scenarios, ensuring the model's generalization ability.

[0066] Secondly, during the tree construction phase, the dataset is recursively split starting from the root node. When selecting splitting features, information gain is used as the criterion, calculating the purity improvement of the subset after splitting for each feature. Information gain is obtained through entropy subtraction: parent node entropy minus the weighted entropy of child nodes. For the orientation fit parameter and the out-of-bounds risk parameter, all possible levels are enumerated as splitting thresholds, and the splitting point with the maximum information gain is selected. Stopping conditions include the node depth reaching the maximum depth threshold, the number of node samples being less than the minimum sample splitting threshold, or all sample labels being the same. The maximum depth threshold is set to an integer value to prevent overfitting, and the minimum sample splitting threshold is set to an integer value to ensure node reliability.

[0067] Finally, in the optimization phase, the constructed decision tree is post-pruned using a cost-complexity pruning method. A regularization parameter alpha is introduced, and its value is determined through cross-validation. Subtrees with small contributions are removed to improve the model's robustness in low-frame-rate falling object scenarios. Training is performed iteratively using batch samples until the tree structure stabilizes.

[0068] In the highway debris motion analysis scenario of this invention, a decision tree model is combined with specific applications: the input consists of directional fit parameters and boundary risk parameters extracted from a set of candidate trajectory segments. These parameters originate from road coordinate transformation and geometric calculations, reflecting the physical characteristics of debris motion; the output trajectory association confidence index is directly used to determine segment affiliation, supporting the construction of temporary debris trajectories. This combination ensures that the model can handle actual road surface dynamics. The input data settings emphasize parameter hierarchy to match tree splitting, and the output is set with dimensionless values ​​for threshold comparison, enabling technicians to reproduce the model based on records and achieve continuous analysis of debris trajectories.

[0069] Through the training logic described above, the decision tree model can be effectively integrated into step S2, providing reliable association judgments and laying the foundation for the subsequent generation of the trajectory of the falling object.

[0070] S2.6 Trajectory Association Judgment and Update.

[0071] Based on the trajectory association confidence index obtained in the previous step, it is necessary to determine the segment assignment. The trajectory association confidence index is compared with preset association conditions. If it exceeds a threshold, the segment to be associated is appended to the corresponding temporary object trajectory, and the end segment of the target temporary trajectory is updated to this segment; otherwise, it is marked as an unassociated segment and retained in the candidate trajectory segment set. A threshold comparison algorithm is used for this determination. After the update, the temporary object trajectory is more complete and ready for use in step S3.

[0072] In step S2, segments to be associated and the end segments of the target temporary trajectory are selected from the candidate trajectory segment set. Road coordinate transformation and displacement vector calculation are performed, directional fit parameters and boundary crossing risk parameters are quantified, and a trajectory association confidence index is output using a decision tree model. Based on the confidence index, the segment is either incorporated into the temporary thrown object trajectory or marked as unassociated. This process incorporates road geometry and lane boundary constraints to improve association accuracy and avoid misjudgments of missing segments at low frame rates. The final temporary thrown object trajectory serves as the input for step S3, ensuring seamless subsequent displacement detection and prediction completion, achieving complete generation of the thrown object image trajectory, and improving the reliability of highway motion analysis.

[0073] Step S2 has associated a set of candidate trajectory segments using a decision tree model to form a temporary trajectory for the thrown object. However, in low frame rate environments, there may still be missing segments, resulting in discontinuous trajectories and an inability to accurately capture the complete motion of the thrown object. Therefore, step S3 introduces a displacement detection mechanism. When the displacement between adjacent frames exceeds a threshold, the position at the next moment is calculated using motion prediction vectors, and unassociated segments are prioritized for completion within the image coordinate range. This aims to repair trajectory gaps, ensure the spatiotemporal continuity of the thrown object image trajectory, support subsequent motion path conversion and analysis, and improve the accuracy of thrown object event processing.

[0074] S3.1 Temporary falling object trajectory inspection and displacement detection.

[0075] Based on the temporary debris trajectory formed in step S2, it is necessary to identify missing segments caused by low frame rates to ensure trajectory integrity. Each temporary debris trajectory is traversed, and the longitudinal displacement of adjacent segments in the road coordinate system is checked in time-stamped order. The displacement distance is calculated as the Euclidean distance between two road coordinate points. If this distance exceeds a preset threshold, a missing trajectory segment is determined. The displacement distance is obtained by calculating the square root of the difference between the longitudinal coordinates of the next road coordinate point and the longitudinal coordinates of the previous road coordinate point, plus the square of the difference between the lateral coordinates of the next road coordinate point and the lateral coordinates of the previous road coordinate point. This threshold is set based on the typical speed and frame rate of the debris and is used to detect abnormal jumps. After detection, missing trajectories are marked for prediction processing.

[0076] S3.2 Motion prediction vector construction.

[0077] In the trajectory of a temporary object with detected displacement gaps, a motion prediction vector needs to be constructed to simulate the expected motion of the object in order to infer the location of the missing segment. Based on the road coordinates of the final segment of the temporary object trajectory and the motion directions of multiple previous segments, an average direction vector is calculated as the basis for prediction. This vector represents the longitudinal and lateral components, and is combined with the estimated displacement to form a complete motion prediction vector. The displacement is obtained from the average displacement of historical segments. After construction, the motion prediction vector is used for coordinate deduction.

[0078] S3.3 Predicting the calculation of road coordinate points.

[0079] Using the motion prediction vector, the next moment's position needs to be estimated in the road coordinate system to guide the search. Starting from the road coordinate point at the end of the temporary debris trajectory segment, the motion prediction vector is added to that point to obtain the predicted road coordinate point. This point includes vertical and horizontal coordinates, representing the expected arrival position of the debris. The vertical coordinate value of the predicted road coordinate point is the sum of the vertical coordinate value of the road coordinate point at the end of the temporary debris trajectory segment and the vertical component value of the motion prediction vector; the horizontal coordinate value is the sum of the horizontal coordinate value of the road coordinate point at the end of the temporary debris trajectory segment and the horizontal component value of the motion prediction vector. The calculation uses a vector addition algorithm. After the calculation is completed, the predicted road coordinate point is ready for image projection.

[0080] S3.4 Image coordinate range is formed.

[0081] Based on the predicted road coordinates, an image plane needs to be returned to define the search area. The predicted road coordinates are mapped back to image coordinates through an inverse transformation of the camera calibration parameters, yielding predicted image points. A rectangular search area is then expanded around these predicted image points; the size of this area is determined by considering the typical scattering range of the projectiles, covering potentially off-center locations. This is achieved using a boundary offset method. Once the search area is formed, it is provided for segment selection.

[0082] The specific process of expanding the rectangular search area centered on the predicted image point begins with the image coordinates of the predicted image point. These coordinates are obtained from the predicted road coordinates through the inverse transformation of the camera calibration parameters, containing both horizontal and vertical coordinates, and serve as the geometric center point of the search area. Next, the width and height dimensions of the search area are determined. These dimensions are directly combined with the typical scattering range of the projectile. This scattering range is statistically derived from historical highway monitoring data and represents the spatial deviation distribution of projectiles in low frame rate video due to motion uncertainties. For example, the longitudinal scattering range corresponds to displacement variations in the road's driving direction, and the lateral scattering range corresponds to lateral offsets within the lane width. Specifically, the longitudinal value of the typical projectile scattering range is mapped to the height of the search area, and the lateral value is mapped to the width. A scaling factor is used to match the image resolution; for example, the width is set to twice the lateral value of the typical projectile scattering range to cover 95% of the deviation probability, and the height is set to twice the longitudinal value to ensure depth search. Then, the left boundary is obtained by subtracting half the width from the x-coordinate value of the predicted image point, the right boundary is obtained by adding half the width to the x-coordinate value, the upper boundary is obtained by subtracting half the height from the y-coordinate value, and the lower boundary is obtained by adding half the height to the y-coordinate value. This forms an axis-aligned rectangular search area, the boundary of which is represented in pixels. This area is used to limit the range of image coordinates for unrelated segments, thereby covering possible positional deviations of the projectiles caused by wind, road friction, or low frame rate sampling errors.

[0083] S3.5 Unrelated segment filtering and completion.

[0084] Within the image coordinate range established in the previous step, potential matching segments need to be searched to fill the gaps. From the unassociated segments marked in step S2, segments whose time markers fall within the expected time interval are selected, and their image coordinates are checked to see if they are within the range. For matching segments, the association logic of step S2 is re-executed, including coordinate transformation to calculate the direction fit parameter and the boundary risk parameter, as well as the trajectory association confidence index output by the decision tree model. If the confidence index meets the association conditions, it is merged into the temporary projectile trajectory, and the end segment is updated. After filtering and completion, the trajectory is more continuous.

[0085] S3.6 Generation of projectile image trajectories.

[0086] The above completion operations require the integration of fragments to form the final image trajectory. The completed temporary projectile trajectory is sorted by time stamp, and the image coordinates of all fragments are extracted to construct a temporally continuous projectile image trajectory. This trajectory records the sequence of image positions of the projectile from its entry to its stationary state. This generation is achieved using a list merging method. After generation, the projectile image trajectory is used for conversion in step S4.

[0087] Step S3 involves displacement checks on the temporary debris trajectory, constructing motion prediction vectors to calculate predicted road coordinates, mapping these to image coordinate ranges to filter out unrelated segments, and reapplying association logic to complete missing parts, forming a continuous debris image trajectory. This process addresses the low frame rate frame skipping problem by providing a prediction completion mechanism, incorporating road coordinates and scattering range constraints to improve trajectory integrity. The final generated debris image trajectory serves as input for step S4, ensuring that subsequent coordinate transformation based on camera calibration parameters can generate a reliable motion path, achieving comprehensive determination of the debris's direction of lane crossing and stopping position, and improving the overall effectiveness of highway anomaly monitoring.

[0088] Step S3 has generated a continuous trajectory of the thrown object image through predictive completion, providing a spatiotemporal sequence on the image plane. However, this trajectory is still limited to pixel coordinates and cannot directly reflect the actual road geometry and physical motion. Therefore, step S4 uses camera calibration parameters to convert the image coordinates into road coordinates, forming a motion path, and extracting the direction of motion, lane crossing, and stopping position. This aims to achieve a comprehensive physical description of the thrown object event, support subsequent verification, ensure that the analysis results meet the actual needs of highway scenarios, and improve the practicality of anomaly handling.

[0089] S4.1 Reading the trajectory of a falling object.

[0090] Based on the projectile image trajectories generated in step S3, each trajectory needs to be processed individually to extract spatiotemporal information. For each trajectory, all trajectory points are traversed in ascending order of time stamp, and the image coordinates of each point are read. These coordinates, expressed in pixels, represent the horizontal and vertical axis values, indicating the position of the projectile on the image plane. The reading process uses a loop to iterate over the trajectory list, ensuring coverage of all points from the earliest to the latest. After reading, the image coordinate sequence is ready for coordinate transformation.

[0091] S4.2 Convert image coordinates to road coordinates.

[0092] Based on the read image coordinate sequence, it needs to be mapped to physical space to reflect the actual road movement. Using camera calibration parameters and road geometry information, each image coordinate point is transformed into a road coordinate system. This system uses the highway centerline as the vertical axis and the vertical direction as the horizontal axis, generating a time-ordered sequence of road coordinate points, each point including both vertical and horizontal coordinate values. The transformation uses a homography matrix projection method. After the transformation, the road coordinate point sequence forms the movement path of the thrown object in the road coordinate system.

[0093] S4.3 Motion Direction Analysis.

[0094] To determine the overall trajectory of the projectile, its dynamic characteristics need to be described. In the road coordinate system, the displacement direction of adjacent road coordinate points along the longitudinal axis is calculated. By accumulating the longitudinal components of all adjacent displacements, if the overall positive direction increases along the highway's travel direction, it is determined to be forward motion; if the overall negative direction decreases or reverses, it is determined to be reverse rolling. The accumulated longitudinal components are obtained by summing the longitudinal coordinate differences of each pair of adjacent points from the first to the second-to-last road coordinate point. The analysis process traverses all path point pairs to ensure complete trajectory coverage. After the analysis is completed, the direction of motion is included as a component of the result.

[0095] S4.4 lane crossing recognition.

[0096] Based on road coordinates along the motion path, lane changes need to be tracked to quantify migration behavior. According to the relative position of each road coordinate point to the lane boundary, geometric intersection detection is used to determine the lane, and the number and order of trajectory points crossing lane boundaries are counted to generate migration sequence information of the projectile from the starting lane to the ending lane. Recognition uses a line segment and boundary intersection method. After recognition, lane crossing information is prepared for integration.

[0097] S4.5 Stop position determined.

[0098] Based on the identified lane crossings, the final point of rest of the thrown object needs to be located to complete the event description. The road coordinate point with the largest time stamp is selected from the movement path, and its location is designated as the resting point. This location includes the specific lane number and road coordinate values, representing the physical point where the thrown object stopped rolling or came to rest. The determination process directly indexes the end point of the path. Once determined, the resting point is added to the analysis results.

[0099] S4.6 The results of the projectile motion analysis are generated.

[0100] Based on the above analysis, information needs to be summarized to form a complete event record. The movement path, direction of movement, lane crossing details, and stopping location are combined into structured data as the result of the object movement analysis. This result describes the entire process of the object from its appearance to its stopping point. Specifically, the movement path is used as the main sequence data field, the direction of movement as the category label field, the lane crossing details as the ordered list field, and the stopping location as the endpoint data field, organized into a single record object using key-value pairs. This generation is achieved using a data dictionary merging method. After generation, the object movement analysis result is used for verification in step S5.

[0101] In step S4, the image coordinate sequence is read from the trajectory of the thrown object and converted into a motion path in the road coordinate system. The longitudinal displacement of adjacent points is analyzed to determine the direction of motion, lane boundary crossings are statistically analyzed to identify crossing behavior, and the endpoint is selected as the stopping position. This information is then combined to form the thrown object motion analysis result. This process incorporates road geometric constraints, providing an accurate depiction of the thrown object's dynamics and avoiding the influence of image distortion. The final generated thrown object motion analysis result serves as the input for step S5, ensuring that subsequent checks based on temporal continuity and lane boundaries can effectively detect discrepancies and trigger updates to the associated conditions, achieving the corrected result and improving the accuracy and reliability of thrown object motion analysis under low frame rate monitoring.

[0102] Step S4 has generated the projectile motion analysis results through coordinate transformation and feature extraction, providing a description of the movement path direction, crossing points, and stopping positions. However, low frame rate correlation may introduce errors, causing the results to differ from the original data and compromising the accuracy of the analysis. Therefore, step S5 introduces temporal continuity and lane boundary constraint verification. When the movement path is abnormal in time interval or location distribution, the trajectory correlation conditions are updated and the segments are re-completed. This aims to correct discrepancies, ensure that the results are highly consistent with the candidate trajectory segment set, improve the reliability of projectile motion characterization, and support the effective utilization of abnormal events on highways.

[0103] S5.1 Time continuity check.

[0104] Based on the projectile motion analysis results generated in step S4, the temporal reliability of the motion path needs to be verified to identify potential anomalies. For the motion path in the road coordinate system, the time stamps of adjacent road coordinate points are compared sequentially to confirm whether they are monotonically increasing, and to assess whether adjacent time intervals fall within a reasonable range preset based on frame rate and projectile velocity. If an interval is higher than the average level, that segment is marked as a potential missing trajectory segment or an abnormal correlation area. The check uses an element-wise traversal method to process all point pairs. After the check is completed, the time anomaly markers are prepared for overall verification.

[0105] S5.2 Spatial Consistency Analysis.

[0106] Building upon the time-based check in the previous step, the spatial rationality of the motion path needs to be examined to match the actual lane constraints. For each pair of adjacent road coordinate points, a connecting line segment is constructed, and the positional relationship between this line segment and the lane boundary is detected in the road coordinate system. Geometric intersection is used to determine whether it crosses multiple lanes. Simultaneously, the image coordinate distribution of the candidate trajectory segment set within the corresponding time period is compared. If the motion path shows crossing multiple lanes but the image coordinates are concentrated in a single lane, or if the motion path point falls outside the lane but the image coordinate is inside the lane, it is marked as a positional mismatch. Specifically, the intersection count of the motion path line segment is numerically compared with the lane assignment statistics of the image coordinates. If the number of intersections is greater than one but the lane number of the image coordinate remains unchanged, or if the lateral coordinate of the motion path exceeds the lane boundary but the lateral value of the image coordinate is within the boundary, a mismatch is determined. The analysis is implemented using a line segment intersection detection method. After the analysis, the spatial mismatch marker and the time marker are used together for mismatch determination.

[0107] S5.3 Non-compliance detection.

[0108] For markers indicating temporal continuity and spatial consistency, a comprehensive evaluation is required to trigger a correction mechanism. All marked areas along the motion path are scanned; if discrepancies in time or location are found, the object motion analysis result is deemed invalid. This determination is based on a simple threshold comparison where the marker count exceeds zero. The total marker count is obtained by summing the indicator values ​​of all marked areas, where each indicator value is 1 if there is a discrepancy in that area and 0 otherwise. After discrepancy detection is completed, the associated conditions are directly updated.

[0109] S5.4 Trajectory Association Condition Update.

[0110] In cases where inconsistent projectile motion analysis results are detected, parameters need to be adjusted to optimize re-association. For the trajectory association conditions in step S2, the update method involves fine-tuning specific parameters for abnormal segments by either lowering the trajectory association confidence index threshold to allow more segments to be incorporated, strengthening the directional fit parameter constraint to prioritize road-consistent motion, or increasing the weight of the boundary risk parameter to suppress high-risk crossings. Specifically, the integration method involves setting the confidence index threshold to a percentage reduction from its original value, setting the directional fit parameter constraint to accept only the first and second levels, and setting the boundary risk parameter weight to double its original value. A new condition set is formed by merging parameter vectors. The update is implemented using an adaptive parameter adjustment method. After the update, the new conditions are applied to the abnormal time range.

[0111] S5.5 Re-association and Completion.

[0112] Based on the updated trajectory association conditions, the process needs to be rerun to repair the trajectory. Within the time range corresponding to the abnormal road segment, segments from the candidate trajectory segment set are reselected, and the association judgment in step S2 is performed, including coordinate transformation parameter calculation and the confidence index output by the decision tree model. This is combined with the prediction completion operation in step S3 to fill in the missing parts and construct a new projectile image trajectory. The rerun process iterates segment by segment to ensure coverage of all discrepancies. After the rerun is complete, the new image trajectory is ready for mapping.

[0113] S5.6 Modify motion path formation and result generation.

[0114] Based on the newly constructed projectile image trajectory, remapping is required to obtain a consistent path. The new image coordinates are converted into a motion path in a road coordinate system. Consistency is verified by comparing the path with the temporal location of a candidate trajectory segment set, resulting in a corrected projectile motion analysis result. This result integrates updated path direction, crossing, and dwell information. Specifically, the integration method uses the motion path as the core sequence field, and direction, crossing, and dwell as additional attribute fields, organized into a complete object through nested data structures. This is achieved using coordinate mapping and data verification methods.

[0115] Step S5 processes the falling object motion analysis results by checking temporal continuity and spatial consistency, comprehensively detecting discrepancies, updating trajectory association conditions such as adjusting confidence index thresholds and parameter constraints, and re-associating and completing segments within the anomaly range to form a corrected motion path and result. This process provides an iterative verification mechanism for low frame rate errors, incorporating lane boundary and time interval constraints to improve result consistency. The final corrected falling object motion analysis result serves as the final output of the invention, ensuring a complete description of the falling object dynamic process and achieving refined and practical monitoring of falling objects on highways.

[0116] Example 2: Figure 2 This invention provides a projectile motion analysis system for highways, comprising:

[0117] The trajectory segment extraction module segments the road region from the low frame rate video sequence and extracts the foreground blocks of suspected thrown objects, and records the appearance features of the time-stamped image coordinates to form a set of candidate trajectory segments.

[0118] The segment association judgment module calculates the direction fit parameter and the boundary risk parameter based on the calibration parameters, road geometry transformation, and image coordinates. It inputs these parameters into the machine learning model to obtain the confidence index. Based on the confidence index, the segments to be associated are merged into the temporary falling object trajectory or marked as unassociated.

[0119] Missing Completion Prediction Module: When anomalies in the trajectory displacement of a temporary object are detected, the image coordinate range is predicted based on the direction of motion of the end coordinate. Within the range, association judgment is performed to complete the unassociated segments and generate the trajectory of the object image.

[0120] Path Feature Analysis Module: The motion path is obtained by transforming the image coordinates using calibration parameters and road geometry. The direction of the thrown object's movement, the position of the lane it crosses, and the stopping position are determined from the path, generating motion analysis results.

[0121] Consistency result correction module: Based on the motion path time continuity lane boundary constraint verification analysis results, when they do not match the candidate trajectory segments, update the association conditions and re-associate to complete the corrected motion analysis results.

[0122] Specifically, the above description is only a preferred embodiment of this application and is not intended to limit this application.

[0123] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0124] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for analyzing the motion of projectiles on highways, characterized in that, Including the following steps: S1. Obtain low frame rate video sequences from highway monitoring equipment, segment each frame into road regions, extract suspected foreground blocks of thrown objects within the road regions, record time markers, image coordinates, and appearance features to form a set of candidate trajectory segments; S2. For the segments to be associated and the end segments of the target temporary trajectory in the candidate trajectory segment set, convert the image coordinates into road coordinates based on the calibration parameters and road geometry, calculate the direction fit parameter and the boundary risk parameter, input the two parameters into the pre-trained machine learning model to obtain the trajectory association confidence index, and according to whether the trajectory association confidence index meets the association conditions, merge the segments to be associated into the target temporary object trajectory or mark them as unassociated segments. The calculation of the directional fit parameter includes obtaining the tangent direction vector of the highway centerline at the road coordinate point of the end segment of the target temporary trajectory based on road geometry information, using it as the longitudinal reference vector, and constructing its orthogonal lateral reference vector. The displacement vector is projected onto the longitudinal and lateral reference vectors and decomposed using a dot product algorithm to obtain the longitudinal and lateral components. The sign of the longitudinal component is compared with the highway driving direction, and the ratio of the absolute value of the longitudinal component to the lateral component is compared. If they are consistent and the ratio is in the first ratio range, the directional fit parameter is at level one; if they are consistent and the ratio is in the second ratio range, it is at level two; otherwise, it is at level three. The displacement vector is obtained by the difference between the road coordinate point of the segment to be associated and the road coordinate point of the end segment of the target temporary trajectory. The first ratio range is limited to the range from zero to 30% of the total range; the second ratio range is limited to the range greater than the upper limit of the first ratio range and less than or equal to half of the total range. The calculation of the boundary crossing risk parameter includes dividing the lateral coordinates into lane intervals based on the lane boundary geometry, determining the lane numbers of the coordinate points of the two roads, calculating the minimum distance from each point to the nearest lane boundary, and using the line connecting the two points as an approximate line segment of the motion trajectory to detect the intersection of the line segment with the lane boundary line. If the lane numbers are the same and the distance is in the first distance interval and there is no intersection, the boundary crossing risk parameter is at level one; if they are the same and the distance is in the second distance interval or there is contact with a single boundary, it is at level two; if they are different or there is intersection with multiple boundaries, it is at level three. The first distance interval is limited to the range from zero to 30% of the normalized lane width; the second distance interval is limited to the range greater than the upper limit of the first distance interval and less than or equal to half of the normalized lane width. S3. When the displacement of the object in the road coordinate system of the adjacent frame exceeds the preset threshold, the image coordinate range of the next moment is predicted based on the road coordinates and movement direction of the end segment of the temporary object trajectory. Within the range, the association judgment in S2 is performed first to complete the unassociated segments and generate the object image trajectory. S4. Based on the calibration parameters and road geometry, convert the image coordinates in the trajectory of the thrown object into road coordinates to obtain the motion path of the thrown object in the road coordinate system. Determine the motion direction, lane crossing and stopping position of the thrown object from the motion path to generate the motion analysis results of the thrown object. S5. Based on the temporal continuity of the motion path in the road coordinate system and the lane boundary constraints, the motion analysis results of the thrown object are verified. When the motion path does not match the candidate trajectory segment in time or position, the trajectory association conditions are updated to re-associate and complete the candidate trajectory segment, thus obtaining the corrected motion analysis results of the thrown object.

2. The method for analyzing the motion of projectiles on highways according to claim 1, characterized in that: Step S1 specifically includes acquiring a low frame rate video sequence from highway monitoring equipment, decomposing the video sequence into an image frame sequence through a video decoding module, applying a road semantic segmentation network to each image frame to generate a mask image to preserve the road area image, calculating the difference image by comparing it with the background reference image using a Gaussian mixture model and extracting the suspected foreground block of the object through connected component analysis, calculating the geometric center of the minimum bounding box of each foreground block as the image coordinates and extracting the high-dimensional feature vector output by the convolutional neural network as the appearance feature vector, integrating the time-stamped image coordinates and appearance feature vectors to form candidate trajectory segments and adding them to the candidate trajectory segment set in chronological order.

3. The method for analyzing the motion of projectiles on highways according to claim 2, characterized in that: Step S2 specifically includes selecting segments to be associated from the candidate trajectory segment set and maintaining the target temporary trajectory end segments; converting image coordinates into road coordinates based on camera calibration parameters and road geometry information to calculate displacement vectors; calculating direction fit parameters to evaluate the degree of conformity between motion and road direction; calculating boundary risk parameters to quantify the possibility of lane migration; inputting the direction fit parameters and boundary risk parameters into a pre-trained decision tree model to obtain a trajectory association confidence index; and merging the segments to be associated into the temporary thrown object trajectory or marking them as unassociated segments based on whether the trajectory association confidence index meets the association conditions.

4. The method for analyzing the motion of projectiles on highways according to claim 3, characterized in that: Step S3 specifically includes checking the longitudinal displacement of adjacent road coordinate systems in the temporary debris trajectory to determine the missing segments, constructing motion prediction vectors to simulate the expected motion, calculating predicted road coordinate points to estimate the next position, mapping the predicted road coordinate points to image coordinates to form a search area, filtering unrelated segments within the search area and performing association judgment to complete the temporary debris trajectory, and generating a continuous debris image trajectory.

5. The method for analyzing the motion of projectiles on highways according to claim 4, characterized in that: Step S4 specifically includes reading the image coordinates in the trajectory of the thrown object in chronological order, converting the image coordinates into road coordinates to obtain the movement path, calculating the longitudinal displacement of adjacent road coordinate points from the movement path to determine the direction of movement, statistically analyzing the crossing situation based on the road coordinate points and lane boundary positions to identify lane crossings, selecting the road coordinate point with the largest time stamp as the stopping position, and combining the movement path, direction of movement, lane crossing situation, and stopping position into the thrown object movement analysis result.

6. The method for analyzing the motion of projectiles on highways according to claim 5, characterized in that: Step S5 specifically includes comparing the time stamps of adjacent road coordinate points to check the temporal continuity of the motion path, analyzing the positional relationship between the connecting line segment and the lane boundary to perform spatial consistency verification, scanning the marked area to determine if the time or position is inconsistent, updating the trajectory association conditions to adjust the confidence index threshold parameter constraints and weights, reselecting candidate trajectory segments within the abnormal time range to perform association judgment and complete the segments, and converting the new image coordinates to form the corrected motion path and the analysis results of the falling object motion.

7. A system for analyzing the motion of projectiles on highways, used to implement the method for analyzing the motion of projectiles on highways as described in any one of claims 1-6, characterized in that, include: Trajectory Segment Extraction Module: Segment road regions from low frame rate video sequences and extract foreground blocks of suspected thrown objects, and record time-stamped image coordinate appearance features to form a set of candidate trajectory segments; The segment association judgment module calculates the direction fit parameter and the boundary risk parameter based on the calibration parameters, road geometry transformation image coordinates, and inputs them into the machine learning model to obtain the confidence index. Based on the confidence index, the segments to be associated are merged into the temporary falling object trajectory or marked as unassociated. Missing Completion Prediction Module: When anomalies in the trajectory displacement of a temporary falling object are detected, the image coordinate range is predicted based on the direction of motion of the end coordinate. Within the range, association judgment is performed to complete the unassociated segments and generate the falling object image trajectory. Path Feature Analysis Module: The motion path is obtained by transforming the image coordinates using calibration parameters and road geometry. The motion direction, lane crossing, and stopping position of the thrown object are determined from the path to generate motion analysis results. Consistency Result Correction Module: The motion path time continuity and lane boundary constraint verification analysis results are used. When they do not match the candidate trajectory segments, the association conditions are updated and re-associated to obtain corrected motion analysis results.

Images