Safety assessment method for airport pick-up area based on video trajectory data

Abstract

The disclosed embodiment provides a method for safety assessment of airport pick-up areas based on video trajectory data, which belongs to the field of data recognition technology and specifically includes: collecting video data of the target pick-up area; extracting trajectory data of vehicles in the target pick-up area based on the video data and preprocessing it; performing lane matching on the preprocessed trajectory data to obtain trajectory information; calculating vehicle conflict indicators based on the trajectory information; inputting the conflict indicators into the target modified CUSBoost model for safety analysis, obtaining interaction trajectory classification based on the conflict indicators and exploring the spatial distribution characteristics and driving behavior characteristics of each classification. Through the solution disclosed in the present invention, a framework from trajectory data collection to safety analysis and unbalanced interaction trajectory data classification is constructed. From the conflict indicators and the safety analysis results of the vehicle interaction process, the causes of conflicts and the changing trends of the conflict indicator values ​​can be studied, thereby improving the efficiency, accuracy and real-time performance of the assessment.

Description

Technical Field

[0001] The present disclosure relates to the field of data recognition technology, and more particularly to a method for assessing the safety of an airport pick-up area based on video trajectory data. Background Art

[0002] Currently, demand for air passenger transport is showing an overall upward trend, placing higher demands on airports' collection and distribution capabilities. Given the constraints of land and environmental resources, how to effectively connect aviation with various forms of transportation has become a critical issue. The airport pick-up area is not only the destination of a vehicle's journey, but also its starting point. A vehicle's entry into the rightmost parking area marks the end of a journey, while its exit marks the beginning of another. Therefore, when a traffic accident occurs, a bottleneck appears on this road section, causing congestion. This section is the only route that subsequent traffic must pass through, and rerouting cannot alleviate the queues. Therefore, conducting a traffic safety analysis of the roads in the airport pick-up area is of great practical significance.

[0003] It can be seen that there is an urgent need for an airport pick-up area safety assessment method based on video trajectory data with high evaluation efficiency, accuracy and real-time performance. Summary of the Invention

[0004] In view of this, the embodiments of the present disclosure provide an airport pick-up area safety assessment method based on video trajectory data, which at least partially solves the problems of poor assessment efficiency, accuracy and real-time performance in the existing technology.

[0005] In a first aspect, an embodiment of the present disclosure provides a method for assessing the safety of an airport pick-up area based on video trajectory data, comprising:

[0006] Step 1: Collect video data of the target pick-up area;

[0007] Step 2: Extracting the trajectory data of vehicles in the target pick-up area based on the video data and preprocessing it;

[0008] Step 3: perform lane matching on the pre-processed trajectory data to obtain trajectory information;

[0009] Step 4: Calculate a vehicle conflict index based on the trajectory information, where the conflict index includes the time required for two vehicles to collide if they continue to run on the same path at their current speed (TTC), the time required for two vehicles to collide if they continue to run on the same path while accelerating or in the process of accelerating (MTTC), the time difference (PET) between the moment the offending vehicle exits the potential collision area and the moment the conflicting vehicle with the right of way arrives at the potential collision point, and the rate (DRAC) at which the following vehicle must decelerate to avoid a collision with the leading vehicle.

[0010] In step 5, the conflict index is input into the target-modified CUSBoost model for safety analysis, and the interaction trajectory classification based on the conflict index is obtained. The spatial distribution characteristics and driving behavior characteristics of each classification are explored.

[0011] According to a specific implementation of the embodiment of the present disclosure, step 2 specifically includes:

[0012] The video is converted into continuous image frames at a preset frame rate and connected to the API interface to obtain the coordinate position and vehicle type of each vehicle. The identification area is specified in the original image, and the behavior of leaving the area is determined based on the vehicle trajectory. The vehicle flow is counted and the statistical value and vehicle tracking box rendering are returned to form trajectory data.

[0013] The pixel coordinates in the trajectory data are converted into real coordinates through homography transformation;

[0014] The Savitzky-Golay filter is selected to smooth the true coordinates of the trajectory points and eliminate abnormal trajectories.

[0015] According to a specific implementation of the embodiment of the present disclosure, step 3 specifically includes:

[0016] Use ArcMap to draw multiple lane layers, and put the lane vector surface features and vehicle trajectory features as input features into Input Features;

[0017] Set the path and name of the output feature, attribute union, minimum tolerance between features and type of output feature to obtain the lane information of the vehicle;

[0018] The lane information is matched with the corresponding trajectory data to obtain the trajectory information.

[0019] According to a specific implementation of the embodiment of the present disclosure, the calculation formula of the TTC value is:

[0020]

[0021] Among them, x L,t and x F,t They represent the positions of the front and rear vehicles at time t, respectively, v L,t and v F,t is the speed of the front and rear vehicles at time t, D L It is the vehicle length of the vehicle in front.

[0022] According to a specific implementation of the embodiment of the present disclosure, the calculation formula of the MTTC value is:

[0023]

[0024] Among them, a L,t and aF,t is the acceleration of the front and rear vehicles at time t.

[0025] According to a specific implementation of the embodiment of the present disclosure, the calculation formula of the PET value is:

[0026] PET t =t F,t -t L,t

[0027] Among them, t F,t is the time when the following vehicle reaches the intrusion line, t L,t It is the time when the front vehicle leaves the intrusion line.

[0028] According to a specific implementation of the embodiment of the present disclosure, the calculation formula of the DRAC value is:

[0029]

[0030] According to a specific implementation of the embodiment of the present disclosure, before step 5, the method further includes:

[0031] The classified and labeled sample indicators are randomly divided into K parts for training. K-1 parts are randomly selected each time as the training set to initially modify the CUSBoost model, and the remaining part is used as the test set. After this round is completed, K-1 parts are randomly selected again to train the data. After multiple rounds, the loss function is selected to evaluate the optimal model and parameters. The average classification accuracy of the K models on the final validation set is used as the performance indicator of the classifier under this K-CV.

[0032] A variety of classification algorithms were selected as the basic algorithms in the first-stage ensemble classifier A and the second-stage classification algorithms in the initial modified CUSBoost model. The optimal model parameters of the ensemble classifier were determined by the traversal optimization method to obtain the target modified CUSBoost model.

[0033] According to a specific implementation of the embodiment of the present disclosure, step 5 specifically includes:

[0034] Extract safe interaction trajectory samples belonging to the majority class and dangerous interaction trajectory samples belonging to the minority class from the conflict index, and use the K-means clustering algorithm to cluster the safe interaction trajectory samples into k categories. Randomly collect samples with the same number of dangerous interaction trajectory samples from the k categories and form a new balanced sample D with the minority class samples. i Then perform binary classification to obtain new dangerous interaction trajectory samples;

[0035] The new dangerous interaction trajectory sample is input into the refined classification model, and the probability vector P of each dangerous category in the sample is output based on the integrated model A. first, and then retrain the supervised clustering algorithm classifier M, and use the classifier to output the probability vector P of each dangerous category in the sample to be processed second ;

[0036] According to the probability vector P of each hazard category first and the probability vector P for each hazard category second Calculate the classification of interaction trajectories and explore the spatial distribution characteristics and driving behavior characteristics of each classification.

[0037] The airport pick-up area safety assessment scheme based on video trajectory data in the embodiment of the present disclosure includes: step 1, collecting video data of a target pick-up area; step 2, extracting and preprocessing trajectory data of vehicles in the target pick-up area based on the video data; step 3, performing lane matching on the preprocessed trajectory data to obtain trajectory information; step 4, calculating a vehicle conflict index based on the trajectory information, wherein the conflict index includes a TTC value required for two vehicles to collide while continuing to travel on the same path at their current speed, a MTTC value required for two vehicles to collide while continuing to travel on the same path while accelerating or in the process of accelerating, a PET value, a time difference between the moment the offending vehicle exits the potential collision area and the moment the conflicting vehicle with the right of way arrives at the potential collision point, and a DRAC value, a rate at which the following vehicle must decelerate to avoid colliding with the leading vehicle; and step 5, inputting the conflict index into a target-modified CUSBoost model for safety analysis, obtaining interactive trajectory classification based on the conflict index, and exploring the spatial distribution characteristics and driving behavior characteristics of each classification.

[0038] The beneficial effects of the embodiments of the present disclosure are as follows: through the scheme of the present disclosure, vehicle trajectories are extracted from the traffic flow video of the airport pick-up area to realize vehicle identification and tracking, the conversion between pixel coordinates and real coordinates is realized through homography transformation, the lane information of the trajectory is matched using ArcMap, and the trajectory is smoothed by Savitzky-Golay filtering. The TTC, MTTC, PET and DRAC traffic conflict indicators are calculated and analyzed in combination with the conflict occurrence process; according to the data characteristics, the vehicle interaction trajectory data, i.e., the vehicle conflict index calculation results, are classified by modifying the CUSBoost model, and the spatial distribution characteristics and driving behavior characteristics of each classification are explored. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0040] Figure 1A flowchart of a method for assessing airport pick-up area safety based on video trajectory data provided by an embodiment of the present disclosure;

[0041] Figure 2 A schematic diagram of the design concept of a method for assessing airport pick-up area safety based on video trajectory data provided by an embodiment of the present disclosure;

[0042] Figure 3 A traffic characteristics diagram provided by an embodiment of the present disclosure;

[0043] Figure 4 A schematic diagram of a pick-up process provided by an embodiment of the present disclosure;

[0044] Figure 5 A schematic diagram of traffic conflict index calculation results provided by an embodiment of the present disclosure;

[0045] Figure 6 A schematic diagram of a conflict indicator value change provided by an embodiment of the present disclosure;

[0046] Figure 7 A schematic diagram of a modified CUSBoost model data processing flow provided in an embodiment of the present disclosure;

[0047] Figure 8 A schematic diagram of an MLP algorithm ROC curve provided in an embodiment of the present disclosure;

[0048] Figure 9 A schematic diagram of the spatial distribution of trajectory points of various categories provided in an embodiment of the present disclosure;

[0049] Figure 10 An average speed distribution of each category provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0050] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0051] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0052] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0053] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0054] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0055] An embodiment of the present disclosure provides a method for assessing airport pick-up area safety based on video trajectory data, which can be applied to road analysis in traffic management scenarios.

[0056] See also Figure 1 , is a flow chart of a method for assessing airport pick-up area safety based on video trajectory data provided by an embodiment of the present disclosure. Figure 1 As shown, the method mainly includes the following steps:

[0057] Step 1: Collect video data of the target pick-up area;

[0058] When it is specifically implemented, the design idea of ​​the present invention is as follows Figure 2 As shown in the figure, it mainly consists of three parts: 1) video data processing module; 2) interactive trajectory data security analysis module; 3) interactive trajectory classification module; 4) airport pick-up area road parking plan improvement module.

[0059] For example, consider Airport A as a research location. The airport's arrival area has three one-way lanes, each 3.5 meters wide and approximately 170 meters long. Due to restrictions on drone flight within the airport area, it's impossible to obtain a full-angle aerial view of the scene, meaning it's difficult to clearly capture traffic conditions and the interactions between traffic participants. Instead, take a bird's-eye view of the arrival area from a corner of the departure level road. The coordinates and image are then converted to a bird's-eye view for easier analysis.

[0060] Step 2: Extracting the trajectory data of vehicles in the target pick-up area based on the video data and preprocessing it;

[0061] Furthermore, the step 2 specifically includes:

[0062] The video is converted into continuous image frames at a preset frame rate and connected to the API interface to obtain the coordinate position and vehicle type of each vehicle. The identification area is specified in the original image, and the behavior of leaving the area is determined based on the vehicle trajectory. The vehicle flow is counted and the statistical value and vehicle tracking box rendering are returned to form trajectory data.

[0063] The pixel coordinates in the trajectory data are converted into real coordinates through homography transformation;

[0064] The Savitzky-Golay filter is selected to smooth the true coordinates of the trajectory points and eliminate abnormal trajectories.

[0065] To call the API for vehicle detection, the video is first converted into image frames, with a frame rate set to 5fps. Once connected to the API, a continuous sequence of video frames is input to detect and track vehicles, returning the coordinates and type of each vehicle. A recognition area is specified in the original image, and vehicle exit behavior is determined based on vehicle trajectories. Traffic flow is then counted, and the statistical value and a rendered image of the vehicle tracking box are returned.

[0066] After tracking a vehicle, we can obtain its trajectory data. The data is stored in the format (vehicle_ID,frame,height,left,top,width), which represents the vehicle ID, frame number, detection frame height, detection frame left horizontal coordinate, detection frame top vertical coordinate, and detection frame width, respectively. To effectively extract traffic conflicts, we need to remove abnormal trajectories.

[0067] The position and trajectory information obtained during vehicle target monitoring and tracking is calibrated based on the pixel coordinates of the original video image. A conversion between pixel coordinates and actual coordinates is required to describe the relative position of the vehicle target in the image. Although there are issues with the shooting angle when shooting overhead videos, the pixel coordinates of the vehicle trajectory cannot reflect the actual road traffic conditions. However, the coordinates can be mapped from one plane to another through a homography transformation to obtain the actual relative geographic location coordinates. At the same time, the image frame of the video is subjected to an image perspective transformation. The transformation projects the image onto a new viewing plane under the condition that the projection center, the original image point, and the target image point are collinear. The converted lane and vehicle coordinates are the overhead plane.

[0068] After the vehicle trajectory coordinates are converted, there is a certain deviation between the obtained trajectory point data and the actual trajectory data due to the interference of some random factors and system errors in the detection process that cannot be avoided by optimization algorithms. The distribution of trajectory points on the left and right of the actual trajectory has a greater impact on the speed. The results calculated from the original trajectory data affect the calculation of traffic conflict analysis indicators. Therefore, the trajectory data needs to be smoothed.

[0069] To ensure a close fit to the actual trajectory, retain subtle mutation information, and smooth the vehicle trajectory for optimal analysis results, a Savitzky-Golay filter is used to smooth the trajectory point coordinates. This filter is used to smooth and remove noise from the data stream. It uses a local polynomial least squares fit in the time domain to eliminate noise while maintaining the shape and width of the fluctuation signal.

[0070] Step 3: perform lane matching on the pre-processed trajectory data to obtain trajectory information;

[0071] Based on the above embodiment, step 3 specifically includes:

[0072] Use ArcMap to draw multiple lane layers, and put the lane vector surface features and vehicle trajectory features as input features into Input Features;

[0073] Set the path and name of the output feature, attribute union, minimum tolerance between features and type of output feature to obtain the lane information of the vehicle;

[0074] The lane information is matched with the corresponding trajectory data to obtain the trajectory information.

[0075] In specific implementation, lane matching means adding lane information to vehicle trajectory information. Three lane layers are drawn using ArcMap, and the lane vector surface features and vehicle trajectory features are placed as input features in "Input Features". The second item is the path and name of the output feature. The third item is the attribute union, which determines which fields in the attribute table of each feature in the original input feature continue to appear in the output feature. The fourth item is the minimum tolerance between features. The fifth item is the type of output feature, thereby obtaining trajectory information.

[0076] Step 4: Calculate a vehicle conflict index based on the trajectory information, where the conflict index includes the time required for two vehicles to collide if they continue to run on the same path at their current speed (TTC), the time required for two vehicles to collide if they continue to run on the same path while accelerating or in the process of accelerating (MTTC), the time difference (PET) between the moment the offending vehicle exits the potential collision area and the moment the conflicting vehicle with the right of way arrives at the potential collision point, and the rate (DRAC) at which the following vehicle must decelerate to avoid a collision with the leading vehicle.

[0077] Furthermore, the calculation formula of the TTC value is:

[0078]

[0079] Among them, x L,t and x F,t They represent the positions of the front and rear vehicles at time t, respectively, v L,t and v F,t is the speed of the front and rear vehicles at time t, D L It is the vehicle length of the vehicle in front.

[0080] Furthermore, the calculation formula of the MTTC value is:

[0081]

[0082] Among them, a L,t and a F,t is the acceleration of the front and rear vehicles at time t.

[0083] Furthermore, the calculation formula of the PET value is:

[0084] PET t =t F,t -t L,t

[0085] Among them, t F,t is the time when the following vehicle reaches the intrusion line, t L,t It is the time when the front vehicle leaves the intrusion line.

[0086] Furthermore, the calculation formula of the DRAC value is:

[0087]

[0088] In the specific implementation, considering the special traffic conditions in the airport pick-up area, the vehicle needs to complete the parking process and then start on a straight road of about 170 meters. The speed of the vehicle is very low. According to the video data, the average speed is 7.29km / h. Figure 3 As shown, (a) represents the average speed, (b) represents the average acceleration, and (c) represents the average deceleration. The horizontal axis represents the position from 0 to 170 meters, and the vertical axis represents the lane number. Lane 1 is the lane closest to the terminal. Lane 3 has the highest average speed, followed by Lane 2. Lane 1 has the lowest speed because it is the parking area. The middle part of the three lanes near the terminal exit is slower than the sides, which is caused by congestion in the waiting area.

[0089] The time that airport pickup vehicles spend in the rightmost lane of the airport arrival area is a key indicator of traffic efficiency. Excessive dwell time can affect the parking options of subsequent vehicles, leading to queues, traffic congestion, and increased risk of traffic conflicts. In the 13-minute case video, 91 vehicles entered and exited the study area, with an average dwell time of 1.09 minutes.

[0090] When a pick-up vehicle enters or leaves the pick-up area, it first slows down to the rightmost road to stop and wait for the passenger, then departs from the waiting area and accelerates to leave the area. During the interaction between vehicles, the driver's response to avoid conflict will also cause changes in speed. The average deceleration of vehicles in the pick-up area is -0.15m / s2, and the average acceleration is 0.12m / s2. (b) and (c) show the average deceleration and acceleration of each area. The largest acceleration and deceleration are at the entrance of the pick-up area, where there is an intention to pass or stop. The overall pick-up process is as follows Figure 4 shown.

[0091] Traffic conflicts in airport arrival areas are primarily characterized by low-speed collisions, frequent lane changes, and short distances between vehicles. Although speeds are low, the limited space and time make it difficult to take appropriate collision avoidance measures in dangerous situations. To assess traffic conflicts, we introduce the TTC, MTTC, PET, and DRAC metrics.

[0092] TTC is defined as “the time required for two vehicles to collide if they continue traveling on the same path at their current speed.” Generally speaking, the smaller the TTC value, the higher the risk of collision.

[0093]

[0094] Among them, x L,t and x F,t Respectively represent the positions of the front and rear vehicles at time t; v L,tand v F,t is the speed of the front and rear vehicles at time t (m / s); D L It is the vehicle length of the vehicle in front.

[0095] TTC is measured only when the conflicting vehicles are on a collision course, meaning the rear vehicle's speed is higher than the front vehicle's. To account for other situations where the conflicting vehicles are not on a collision course but could collide due to differences in acceleration or deceleration, the MTTC metric relaxes the constant speed assumption during the collision course and takes into account the acceleration or deceleration of the vehicles.

[0096]

[0097] Among them, a L,t and a F,t is the acceleration of the front and rear vehicles at time t.

[0098] PET is defined as the time difference between the moment the "offending" vehicle exits the potential collision zone and the moment the "conflicting" vehicle, which has the right of way, arrives at the potential collision point. For rear-end collisions, it is the conflict line, not the area, that is used to measure PET. PET calculation does not require speed-related parameters or extrapolation of future positions.

[0099] PET t =t F,t -t L,t (3)

[0100] Among them, t F,t is the time when the following vehicle reaches the intrusion line, t L,t It is the time when the front vehicle leaves the intrusion line.

[0101] DRAC is the rate at which the following vehicle must decelerate to avoid a collision with the preceding vehicle. The larger the DRAC value, the higher the risk of collision. If DRAC exceeds the maximum available deceleration, a collision will occur.

[0102]

[0103] like Figure 5 As shown in Figure 1, (a) shows the conflict index distribution, and (b) shows the conflict index box plot. The conflict index distribution in (a) shows that the three traffic conflict indicators, MTTC, PET, and TTC, are all concentrated around 10 seconds, with their frequencies first increasing and then decreasing. The solid line in the box plot in (b) represents the median, and the dashed line represents the mean. In the airport pickup area, when driving at low speeds or with some vehicles stationary, the conflict index values ​​can, to a certain extent, reflect the road conflict situation.

[0104] Extract TTC, MTTC, PET less than 3s and DRAC greater than 3m / s from the calculation results of the conflict indicators 2The trajectory information is combined with the vehicle interaction in the video frame for analysis, see Table 1. The numbers in the table represent how many trajectory points belong to a certain conflict process under this conflict indicator. For example, 164 represents that 164 trajectory points with a TTC of less than 3s in the trajectory data are in the process of the rear vehicle changing lanes. In the case of a rear vehicle changing lanes, the number of TTC and MTTC is the largest. Since MTTC takes into account the impact of vehicle acceleration (or deceleration) compared with TTC, the number of trajectories of these two indicators is also different in this case. PET has the largest proportion in the following vehicle scenario because it considers more the time difference between the leading and following vehicles arriving at the same conflict section. DRAC best represents the scenario where a vehicle in the side lane is inserted in front of the vehicle. The driver needs to react in time from the process of planning the lane change to the lane change and inserting in front of the vehicle in the lane.

[0105] Drivers should brake when necessary to avoid a collision with the vehicle in front.

[0106]

[0107] Table 1

[0108] In order to observe the numerical changes of conflict indicators in the same scenario, this embodiment extracts specific scenarios such as Figure 6 Figure 1 shows a scenario where (a) a rear vehicle changes lanes, (b) a rear vehicle follows, and (c) a vehicle changing lanes and the vehicle behind it in the target lane. (a) shows the TTC value changes when a rear vehicle changes lanes. As the rear vehicle changes lanes, the distance between the rear vehicle and the leading vehicle decreases, and the TTC value shows a downward trend until the rear vehicle completes the lane change. Simultaneously, as can be seen in the figure, another vehicle is also changing lanes behind. Due to the small distance between vehicles, drivers should always pay attention to the changes in the operating status of the vehicle ahead to avoid traffic accidents.

[0109] (b) shows a scenario of a car following behind. The distance between the leading and trailing vehicles is very short, and the trailing vehicle has a tendency to change lanes. The overall PET value also shows a downward trend. The characteristics of the airport pickup area create a short following distance. Especially when the leading and trailing vehicles need to change lanes, the changes in distance and speed between the leading and trailing vehicles lead to different levels of conflict.

[0110] (c) illustrates a conflict scenario between the lane-changing vehicle and the vehicle behind it in the target lane. When a vehicle initiates a lane change, the DRAC value is maximum, and the following vehicle must immediately decelerate to accommodate the lane change to avoid a collision. As the preceding vehicle's lane change progresses, the DRAC value gradually decreases. When a vehicle in an adjacent lane changes lane in front of the vehicle in the target lane, more traffic conflicts occur due to limited vision and varying reaction speeds.

[0111] In step 5, the conflict index is input into the target-modified CUSBoost model for safety analysis, and the interaction trajectory classification based on the conflict index is obtained. The spatial distribution characteristics and driving behavior characteristics of each classification are explored.

[0112] Based on the above embodiment, before step 5, the method further includes:

[0113] The classified and labeled sample indicators are randomly divided into K parts for training. K-1 parts are randomly selected each time as the training set to initially modify the CUSBoost model, and the remaining part is used as the test set. After this round is completed, K-1 parts are randomly selected again to train the data. After multiple rounds, the loss function is selected to evaluate the optimal model and parameters. The average classification accuracy of the K models on the final validation set is used as the performance indicator of the classifier under this K-CV.

[0114] A variety of classification algorithms were selected as the basic algorithms in the first-stage ensemble classifier A and the second-stage classification algorithms in the initial modified CUSBoost model. The optimal model parameters of the ensemble classifier were determined by the traversal optimization method to obtain the target modified CUSBoost model.

[0115] Furthermore, the step 5 specifically includes:

[0116] Extract safe interaction trajectory samples belonging to the majority class and dangerous interaction trajectory samples belonging to the minority class from the conflict index, and use the K-means clustering algorithm to cluster the safe interaction trajectory samples into k categories. Randomly collect samples with the same number of dangerous interaction trajectory samples from the k categories and form a new balanced sample D with the minority class samples. i Then perform binary classification to obtain new dangerous interaction trajectory samples;

[0117] The new dangerous interaction trajectory sample is input into the refined classification model, and the probability vector P of each dangerous category in the sample is output based on the integrated model A. first , and then retrain the supervised clustering algorithm classifier M, and use the classifier to output the probability vector P of each dangerous category in the sample to be processed second ;

[0118] According to the probability vector P of each hazard category first and the probability vector P for each hazard category second Calculate the classification of interaction trajectories and explore the spatial distribution characteristics and driving behavior characteristics of each classification.

[0119] In the specific implementation, considering the need to better examine the impact of conflict indicators on the risk level of the same trajectory, the status of the two vehicles at that moment is comprehensively evaluated, and the TTC, MTTC, PET, and DRAC indicators of each trajectory data are classified. For trajectory data with TTC, MTTC, and PET greater than 60s or empty values, they are marked as 60s; for empty values ​​of DRAC, they are marked as 0m / s. 2 . Due to the uneven distribution of safe driving and dangerous driving samples in the trajectory data, some samples were first manually classified and labeled. Class 0 samples were set as safe driving, that is, the trajectory data performed relatively well on the four conflict indicators, and the indicator values ​​were within the safe range; Class 1 samples mostly showed conflicting states on the MTTC and PET indicators, and were relatively safe overall on the TTC and DRAC indicators; Class 2 samples showed dangerous performance on the MTTC indicator, while other indicators were relatively safe; Class 3 samples showed dangerous performance on the TTC, PET, and DRAC indicators, while being safe on the MTTC indicator.

[0120] To classify interaction trajectory data, we introduce Modified CUSBoost, a multi-classification algorithm for imbalanced samples targeting interaction trajectory categories. Conventional machine learning methods for imbalanced data classification tend to maximize the accuracy of the majority class, while often underperforming on the minority class. However, in this research problem, we are more interested in the characteristics of the minority class—dangerous interaction trajectories.

[0121] Methods for handling imbalanced data classification problems include sampling methods, cost-sensitive learning methods, and ensemble learning methods. CUSBoost, a clustering-based undersampling boosting method, is highly effective in handling imbalanced data classification problems. The algorithm divides the imbalanced dataset into two parts: the majority class and the minority class. The K-means clustering algorithm is used to cluster the majority class into multiple classes, and the majority class is selected from each cluster to form a balanced dataset with the minority class. CUSBoost uses the C4.5 algorithm to consider a series of decision trees and combines the votes of each tree to classify new instances.

[0122] The characteristics of dangerous interaction trajectories are not single, so they can be subdivided into multiple categories. The modified cusboost algorithm solves the imbalance problem of multi-classification. It is a combination of cluster-based sampling and Adaboost algorithm. Figure 7As shown in the figure, the algorithm is divided into two stages. The first is a binary classification model for safe and dangerous interaction trajectories: safe interaction trajectory samples belonging to the majority class and dangerous interaction trajectory samples belonging to the minority class are extracted from the original data, and the majority class samples are clustered into k categories using the K-means clustering algorithm. Samples with the same number of minority class samples are uniformly randomly collected from the k categories (samples of non-majority classes in the original data are all regarded as minority class samples), and a new balanced sample D is formed with the minority class samples. i , parameter k is determined by traversal. Use the integrated supervised decision algorithm classifier A and combine each algorithm A i The votes of x are used to classify the new instance. i The classification difficulty is determined by the weight w of the sample i To reflect, the sample weight w i The samples that are difficult to classify will be adjusted according to their classification method, with higher weights given to samples that are easy to classify and lower weights given to samples that are easy to classify. i The classification error rate is:

[0123]

[0124] Among them, if an instance x i is misclassified, then error(x i ) is 1. Otherwise, error(x i ) is 0, and during the iteration, if a classifier converts x in each iteration i Correct classification, then x i The weight multiplied by Normalize the weights of all instances (including misclassified instances). Here, if model A i If the error rate exceeds 0.5, then give up A. i , and generate a new sub-dataset D by i To derive a new A i . Use the integrated classifier to output the majority and minority category labels of the samples to be processed.

[0125] Then, a refined classification model for dangerous interaction trajectories is established. Based on the model in the previous stage, dangerous interaction trajectory samples R are extracted, and these samples are used to output the probability vector P of each dangerous category in the sample based on the integrated model A. first , and then retrain the supervised clustering algorithm classifier M, and use this classifier to output the probability vector P of each dangerous category in the sample to be processed second The probability vector of each hazard category in the sample to be processed is determined using the following equation:

[0126] P=P first*accuracy(A) +P second *accuracy(M) (6)

[0127] Where accuracy(A) represents the accuracy of the first-stage ensemble classifier, and accuracy(M) represents the accuracy of the second-stage classifier.

[0128] During model training, K-fold cross validation (K-fold cross validation) is used. The sample data is randomly divided into K parts (usually equally divided). K-1 parts are randomly selected each time as the training set, and the remaining part is used as the test set. After each round, K-1 parts are randomly selected again for training. After several rounds (less than K), a loss function is selected to evaluate the optimal model and parameters. The average classification accuracy of these K models on the final validation set is used as the performance indicator of the classifier under this K-fold cross validation. Here, the value of k is 5.

[0129] In addition, six highly efficient classification algorithms were selected as the basis for the first-stage ensemble classifier A and the second-stage classification algorithm mentioned in the previous section. These six algorithms are the decision tree generation algorithm (C4.5), the Gaussian Naive Bayes algorithm (GaussianNB), the K-nearest neighbor classification algorithm (KNN), the logit regression classification algorithm (LR), the multilayer perceptron algorithm (MLP), and the support vector machine classification algorithm (SVC). The optimal model parameters of the ensemble classifier were determined using a traversal optimization method.

[0130] The ROC curve plots the True Positive Rate (TPR) along the Y-axis and the False Positive Rate (FPR) along the X-axis. A perfect classifier would have an area under the curve (AUC) of 1, where TPR = 1 and FPR = 0. The curve shows the classifier's TPR and FPR at different cutoff thresholds.

[0131]

[0132]

[0133] In the case of imbalanced interaction trajectory classification, placing special emphasis on the prediction accuracy of the minority class while retaining the accuracy of the majority class will result in high TPR and low FPR, which is reflected in high AUC values. Therefore, AUC is used to compare the performance of each classification algorithm.

[0134] For multi-classification problems, each test sample has a probability vector corresponding to n different categories and a label vector. For example, if n = 3, the label vector for the sample belonging to the second category is [0, 1, 0]. Therefore, for all samples, based on the obtained probability matrix P and label matrix L, the false positive rate (FPR) and true positive rate (TPR) at various thresholds can be calculated, thereby plotting a receiver operating characteristic (ROC) curve. This results in a total of n ROC curves. Finally, the n ROC curves are averaged to obtain the final ROC curve. The area under the curve (AUC) is also calculated. The final classifier accuracy is calculated using the aforementioned K-fold cross-validation method.

[0135] Among the six classification algorithms, MLP has the highest accuracy of 0.937211. Table 2 illustrates the AUC values ​​of different classification models. Figure 8 The ROC curve of the MLP is shown.

[0136]

[0137] Table 2

[0138] Unlabeled samples are fed into the model trained on labeled samples to obtain data labels, i.e., classification results. Table 3 shows the classification results, including the number and percentage of each category, as well as the minimum, maximum, mean, and variance of each metric within each category. The classification results show that the performance of each category on the conflict metric is consistent with the performance of the samples used in data preparation.

[0139]

[0140] Table 3

[0141] Analyzing the spatial distribution of each type of interaction trajectory can help us better understand the causes of traffic conflicts and lay the foundation for subsequent improvement measures. Figure 9 As shown in Figure 1, (a) to (d) show the spatial distribution of the four categories. Category 0, which accounts for the largest share, is concentrated in the rear of Lanes 1 and 2, where parked vehicles begin to leave the airport pick-up area. Category 1 is concentrated in Lanes 2 and the rear of Lane 1, which is also where lane changes and congestion are most frequent. Category 2 is mainly concentrated in Lane 1 near the terminal exit, which is an area where parked vehicles and passengers gather. Category 3 is mainly concentrated in the rear section of Lane 1, where there are more vehicle entry and exit behaviors and more intersections where vehicle trajectories overlap.

[0142] The spatial distribution characteristics of velocity also vary in different classifications. Figure 10The average speeds for different areas of each lane within each category are displayed, with the speedometer indicating 0 km / h in areas without track points. The speed distribution for Category 0 shows higher speeds at the entrances and exits of lanes 2 and 3, and lower speeds in other areas. Vehicles at the entrances and exits of the airport pickup area travel smoothly, with good spacing between vehicles. The spatial speed distributions for Categories 1 and 3 are similar, due to their lane-changing and stopping behaviors. Both categories have higher speeds at the entrances to lane 2 and between lanes 3. Category 2 vehicles have lower overall speeds, with higher speeds concentrated in lane 3.

[0143] At the same time, in practical applications, after obtaining the classification of conflict indicators and exploring the spatial distribution characteristics and driving behavior characteristics of each classification, targeted improvement measures can be proposed to improve the safety and efficiency of the roads in the airport pick-up area.

[0144] For example, Airport A has a designated parking lot for ride-hailing companies, but its utilization rate is very low. After landing, it's inconvenient for passengers to walk to the designated parking lot with their luggage. Passengers who choose to leave the airport hub by ride-hailing company or private car are more likely to enter the pick-up area through the terminal exit and wait for their vehicle to arrive. Therefore, improvement measures should focus on the docking area design and docking rules in the pick-up area. The following improvement measures can be designed:

[0145] Option 1: Parking in different areas

[0146] Vehicles entering the airport pickup area have varying waiting times. Parking areas are assigned based on the length of their parking time, with shorter parking times assigned closer to the front of the area. Furthermore, because the proportion of vehicles with different parking times varies, the number of parking spaces in each area also varies. The parking lot is divided into three distinct zones. No parking space attraction rates are set, but instead different types of vehicles are assigned to specific zones. The three types of vehicles entering the parking lot are then used in routing decisions, and the resulting efficiency and safety changes are compared to a scenario without zone division.

[0147] Option 2: Attract parking

[0148] The parking area of ​​this scheme is not divided, and the attraction rate of parking spaces is set from large to small from front to back. That is, the closer the parking space is to the passenger exit, the greater the attraction. If there is a parking space in front, the vehicle will park there. If the parking space is full, the vehicle that needs to park will wait in the middle lane. The changes in efficiency and safety of this scheme are compared with the initial random scheme.

[0149] Option 3: Attract parking by area

[0150] To ensure that parking areas are spaced appropriately for vehicles with different parking time distributions, the zoned attraction parking system divides parking spaces into different zones. The number of parking spaces in each zone is determined based on the proportion of vehicles with different parking time distributions. Furthermore, the attraction level of parking spaces is set for different parking zones, with attraction decreasing from the front to the back, to compare the efficiency and safety changes compared to other solutions.

[0151] Using VISS IM software, we simulated the airport arrival area based on video trajectory data, ensuring the simulation's authenticity and accuracy. Using VISS IM's traffic flow and traffic control modules, we set up detectors to calculate road network evaluation metrics such as delays, queue lengths, and travel times for each road segment. The steps are as follows:

[0152] 1. Build a road network. Draw each road segment (including the two outer lanes and the innermost parking lane) in VISSIM software based on the actual road lane distribution, and connect them with connectors to form a road network.

[0153] 2. Model calibration and verification. To ensure simulation accuracy, various behavioral parameters in the simulation program are adjusted based on the results of trajectory data processing and analysis. The simulation operation parameters, traffic flow characteristic parameters, and driving behavior in the system are verified.

[0154] 3. Basic simulation data input: Input environmental variables, traffic facility geometry parameters, driving behavior control type, vehicle parameters, operating criteria, and other data.

[0155] 4. Set up detectors for parameter evaluation. Set up data collectors, travel time detectors, queue counters, and other devices on each road section in the airport pickup area. After setting the evaluation parameters, run the simulation program to obtain simulation results for each road section.

[0156] When running the VI SS IM simulation, the simulation parameters were calibrated based on the trajectory data of the arrival level pick-up area of ​​Terminal 2 at Airport A.

[0157] 1. Vehicle composition

[0158] Based on the trajectory data, the simulation system configured the vehicle composition and expected speed distribution for the airport pickup area's roads. By analyzing the video trajectories, the road's vehicle composition was determined to include only cars. Three vehicle types were set, each comprising 70% of vehicles requiring parking in the parking area. Vehicles were categorized by parking duration and assigned different color types, with a ratio of 1:2:1.

[0159] 2. Expected speed and driving behavior settings

[0160] Based on actual road conditions, the simulation system uses the Widemann74 model in VISSIM as the vehicle behavior model. Furthermore, based on actual traffic conditions, the vehicle's lateral driving behavior is configured to observe adjacent lanes and allow vehicles in the same lane to overtake from the left.

[0161] According to trajectory data statistics, the expected speed distribution in the airport pick-up area is between 2km / h and 43km / h, and 80% of the vehicles are traveling between 2km / h and 20km / h.

[0162] 3. Path Decision

[0163] Vehicle routing involves setting the vehicle's travel path. Routing requires selecting a starting point and end point to form a decision segment. In this scenario, a static route is first set for the segment, from the airport pickup area entrance to the exit. A parking route is then defined based on the parking lot's location. The parking route is included within the static route, starting near the entrance and ending at the parking area. Of the three types of vehicles, only those with parking needs are included in the routing decision to enter the parking lot. The remaining vehicles follow the static route, entering the airport pickup area and exiting through the exit.

[0164] 4. Parking area

[0165] A parking area is set up on the innermost road of the airport pickup area, with each parking space being 6 meters. Based on the road length and actual parking conditions, a total of 21 parking spaces are set up. Every vehicle entering the pickup area with a parking demand must enter the parking lot and wait until it picks up its passenger, that is, until the parking time setting is reached, before it can leave. In addition, when the parking area is fully loaded, the status of the remaining vehicles is set to waiting. The distribution of vehicle parking time follows the pre-set waiting time, which follows a normal distribution. Based on the actual parking time of vehicles extracted from the trajectory data, the mean parking time of the three types of vehicles is 30 seconds, 60 seconds, and 90 seconds, respectively, with a standard deviation of 5 seconds. In order to simulate some drivers who are not proficient in entering the parking area for parallel parking, the vehicle road occupation time is set to follow a normal distribution with a mean of 5 seconds.

[0166] 5. Model Validation

[0167] In order to verify the validity of the model settings in the system, the two sets of data are compared between the trajectory data and the system simulation model operation. If the simulation data and the actual data are consistent with each other, the simulation model is considered to meet the requirements.

[0168] Each simulation experiment lasted 3600 seconds, and was run five times in succession. The first 300 seconds of the simulation was used as the clearing time, and the system began collecting data from the 300th second, with a data collection interval of 300 seconds. The effectiveness of the traffic flow on the roads in the airport pick-up area was verified.

[0169] First, a two-sample F-test was used to verify whether the variances of the two groups of data were equal. The F value obtained through the F-test was 2.38, and the one-tailed critical value of F was 2.82. Since F < F one-tailed critical, there was no significant difference between the two sets of sample data at a 95% confidence level, proving that the variances of the actual data and the simulation data were equal.

[0170] A T-sample equal variance hypothesis test was used to verify whether the means of the actual data and the simulation data were equal, and it was verified that the means of the samples of the actual data and the simulation data were the same.

[0171] Therefore, it can be considered that the setting of the simulation scenario is more in line with the traffic operation state of the actual scenario, and the results output by the simulation scenario can reflect the real situation to a certain extent.

[0172] When comparing Plan 1 (zoned parking) with the current situation, the results showed that the number of conflicts caused by lane changes decreased. Although Plan 1 can distinguish different types of vehicles according to the parking time, it cannot effectively utilize the empty parking spaces, increasing the burden of traffic congestion.

[0173] Compared with the current situation plan, Plan 2 (attracting parking) significantly improved road traffic safety, increased the traffic safety index value, and significantly reduced the number, severity, and distribution range of traffic conflicts. Vehicles will preferentially choose the parking spaces near the exit of the pick-up area to park and wait, reducing the walking distance and time of passengers and increasing the turnover rate of parking spaces.

[0174] Compared with the current situation, Plan 3 (zoned attracting parking) did not have an obvious improvement effect. The utilization rate of parking spaces did not reach its maximum capacity. The traffic conflicts between waiting vehicles and departing vehicles and the rear-end collisions caused by subsequent vehicles were not alleviated either.

[0175] In terms of road network performance, the road network evaluation index of Plan 2 was close to the current situation plan; in terms of traffic safety, Plan 2 had an obvious improvement effect, and both the number and severity of conflicts were improved. Therefore, Plan 2 had the best improvement effect on the traffic safety of the airport pick-up area. The trajectory overlap and interweaving between vehicles entering and leaving the pick-up area were the root causes of serious traffic conflicts, and the trajectory overlap range in Plan 2 was reduced, so the improvement effect on traffic conflicts was more obvious.

[0176] This embodiment provides an airport pickup area safety assessment method based on video trajectory data. Vehicle trajectories are extracted from airport pickup area traffic videos using the Baidu API interface to achieve vehicle identification and tracking. A homography transformation is used to convert pixel coordinates to real-world coordinates. Lane information is matched to the trajectories using ArcMap. The trajectories are smoothed using a Savitzky-Golay filter. Traffic conflict indicators such as TTC, MTTC, PET, and DRAC are calculated and analyzed in conjunction with the conflict occurrence process. Based on data characteristics, the vehicle interaction trajectory data, i.e., the calculated vehicle conflict indicators, are classified using a modified CUSBoost model. The spatial distribution characteristics and driving behavior characteristics of each classification are explored.

[0177] The units involved in the embodiments described in this disclosure may be implemented by software or hardware.

[0178] It should be understood that various parts of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0179] The above description is merely a specific embodiment of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. A method for assessing the safety of an airport pick-up area based on video trajectory data, characterized in that: include: Step 1: Collect video data of the target pick-up area; Step 2: Extracting the trajectory data of vehicles in the target pick-up area based on the video data and preprocessing it; Step 3: perform lane matching on the pre-processed trajectory data to obtain trajectory information; Step 4: Calculate a vehicle conflict index based on the trajectory information, where the conflict index includes the time required for two vehicles to collide if they continue to run on the same path at their current speed (TTC), the time required for two vehicles to collide if they continue to run on the same path while accelerating or in the process of accelerating (MTTC), the time difference (PET) between the moment the offending vehicle exits the potential collision area and the moment the conflicting vehicle with the right of way arrives at the potential collision point, and the rate (DRAC) at which the following vehicle must decelerate to avoid a collision with the leading vehicle. Step 5: Input the conflict index into the target-modified CUSBoost model for safety analysis, obtain the interaction trajectory classification based on the conflict index, and explore the spatial distribution characteristics and driving behavior characteristics of each classification; Before step 5, the method further includes: The classified and labeled sample indicators are randomly divided into K parts for training. K-1 parts are randomly selected each time as the training set to initially modify the CUSBoost model, and the remaining part is used as the test set. After this round is completed, K-1 parts are randomly selected again to train the data. After multiple rounds, the loss function is selected to evaluate the optimal model and parameters. The average classification accuracy of the K models on the final validation set is used as the performance indicator of the classifier under this K-CV. A variety of classification algorithms were selected as the basic algorithms in the first-stage ensemble classifier A and the second-stage classification algorithms in the initial modified CUSBoost model. The optimal model parameters of the ensemble classifier were determined by traversal optimization method to obtain the target modified CUSBoost model. The step 5 specifically includes: Extract safe interaction trajectory samples belonging to the majority class and dangerous interaction trajectory samples belonging to the minority class from the conflict index, and use the K-means clustering algorithm to cluster the safe interaction trajectory samples into k categories. Then, uniformly and randomly collect the same number of samples as the dangerous interaction trajectory samples from the k categories and form a new balanced sample with the minority class samples. Then perform binary classification to obtain new dangerous interaction trajectory samples; The new dangerous interaction trajectory samples are input into the refined classification model, and the probability vector of each dangerous category in the sample is output based on the integrated model A. , and then retrain the supervised clustering algorithm classifier M, and use the classifier to output the probability vector of each dangerous category in the sample to be processed ; According to the probability vector of each hazard category and the probability vector of each hazard category Calculate the classification of interaction trajectories and explore the spatial distribution characteristics and driving behavior characteristics of each classification.

2. The method according to claim 1, characterized in that , the step 2 specifically includes: The video is converted into continuous image frames at a preset frame rate and connected to the API interface to obtain the coordinate position and vehicle type of each vehicle. The identification area is specified in the original image, and the behavior of leaving the area is determined based on the vehicle trajectory. The vehicle flow is counted and the statistical value and vehicle tracking box rendering are returned to form trajectory data. The pixel coordinates in the trajectory data are converted into real coordinates through homography transformation; The Savitzky-Golay filter is selected to smooth the true coordinates of the trajectory points and eliminate abnormal trajectories.

3. The method according to claim 2, characterized in that , the step 3 specifically includes: Use ArcMap to draw multiple lane layers, and put the lane vector surface features and vehicle trajectory features as input features into Input Features; Set the path and name of the output feature, attribute union, minimum tolerance between features and type of output feature to obtain the lane information of the vehicle; The lane information is matched with the corresponding trajectory data to obtain the trajectory information.

4. The method according to claim 3, characterized in that , the calculation formula of the TTC value is ； in, and Respectively represent the positions of the front and rear vehicles at time t, and is the speed of the front and rear vehicles at time t, It is the vehicle length of the vehicle in front.

5. The method according to claim 4, characterized in that , the calculation formula of the MTTC value is: ； in, and is the acceleration of the front and rear vehicles at time t.

6. The method according to claim 5, characterized in that , the calculation formula of the PET value is ； in, is the time when the following vehicle reaches the intrusion line, It is the time when the front vehicle leaves the intrusion line.

7. The method according to claim 6, characterized in that , the calculation formula of the DRAC value is 。

Images