A two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics

Abstract

The application discloses a two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics, and comprises the following steps: collecting original trajectory point data of vehicles at a road intersection through a data acquisition device; preprocessing trajectory point data in the original trajectory point data; grouping the preprocessed trajectory point data according to vehicle IDs, each vehicle ID being associated with a group of data, so as to determine the trajectory point data of each vehicle; and adopting a two-stage joint denoising method to denoise the trajectory point data of each vehicle.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of transportation engineering, and particularly relates to a two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics. BACKGROUND

[0002] For the denoising problem of vehicle trajectory data detected by radar at road intersections, the core idea of the existing technology is mostly based on Kalman filtering and S-G filtering to achieve data denoising, and related algorithms such as rule definition are usually used for data denoising. The existing technology includes using fuzzy subtraction clustering to identify noise data, and using singular value filtering method to process noise data. For the characteristics of single-source traffic data, the existing technology uses a data-driven method to identify and repair traffic flow fault data (missing data, redundant data, abnormal data, etc.). The existing technology also combines an improved fruit fly algorithm to predict network traffic data under different Gaussian noise distribution data.

[0003] Kalman filtering is essentially a data fusion algorithm under the assumption of linearity and Gaussian noise, which uses observation information, state transition and observation model to smooth, filter and predict data characteristics, so as to obtain a more accurate measurement parameter. By establishing a signal and noise corresponding state space model, the optimal estimation of the current state can be determined according to the state of the last time and the possible measurement value. The main steps of the denoising method using Kalman filtering technology in the existing technology are as follows: in the iteration process of "prediction-update", if the absolute deviation of the observation value and the predicted value is obviously greater than the average deviation value, then the data should be marked as noise trajectory point in preliminary screening. S-G filtering is a polynomial regression method based on local characteristics of a curve. By using S-G filtering to fit the local trajectory data of the target vehicle, noise trajectory points can be more accurately identified through the motion characteristics of the vehicle trajectory, and continuous trajectory smoothing can be performed.

[0004] The existing technology uses a sliding window to move the fitted data on the original data sequence one by one until the entire data sequence is traversed. However, there are often a large number of abnormal parking state trajectory points at intersections, which will seriously affect the filtering effect of the existing technology. SUMMARY

[0005] In view of the problems of difficult filtering denoising and poor trajectory fitting accuracy caused by vehicle state trajectory point interference at intersections in the existing technology, the present application proposes a two-stage joint denoising algorithm based on Kalman filtering and improved S-G filtering based on driving characteristics. For noise data, based on the kinematic characteristics of vehicles in different regions of the intersection, a two-stage joint denoising algorithm based on Kalman filtering and improved S-G filtering is proposed to achieve efficient identification and rapid processing of trajectory point noise data, effectively solving the shortcomings of the existing technology.

[0006] The application provides a two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics, which comprises the following steps:

[0007] Step 1: collecting original trajectory point data of vehicles at a road intersection by a data acquisition device;

[0008] Step 2: preprocessing the trajectory point data in the original trajectory point data;

[0009] Step 3: grouping the preprocessed trajectory point data according to vehicle ID, each vehicle ID being associated with a group of data, so as to determine the trajectory point data of each vehicle;

[0010] Step 4: for the trajectory point data of each vehicle, a two-stage joint denoising method is used for denoising, wherein the two-stage joint denoising comprises the following steps:

[0011] Step 5: filtering the trajectory point data of each vehicle by using a Kalman filtering method to generate a first filtering result;

[0012] Step 6: filtering the first filtering result by using an improved S-G filtering method to generate a second filtering result, and according to the second filtering result, performing noise trajectory point identification and processing to obtain a denoising result,

[0013] wherein the improved S-G filtering method comprises the following steps:

[0014] Step 7: identifying the trajectory points in each sliding window in an abnormal parking state, and integrating the trajectory points in the abnormal parking state into one trajectory point;

[0015] Step 8: filtering the integrated trajectory points in each sliding window by using an S-G filtering method.

[0016] Further, the preprocessing comprises data cleaning, wherein the data cleaning comprises: processing missing data, abnormal data and repeated data in the trajectory point data.

[0017] Further, the method further comprises the following steps: sorting the preprocessed trajectory point data according to sampling time; grouping the sorted trajectory point data according to sampling time; for each group of data after grouping, calculating the IoU value of any two data in each group of data and removing repeated target data according to the IOU value.

[0018] The application has the following beneficial effects:

[0019] By designing a two-stage joint denoising algorithm based on the driving characteristics of Kalman filtering and improved S-G filtering, the main idea is that firstly, Kalman filtering is used for preliminary noise recognition and processing, and the trajectory points with obvious errors are removed, and then according to the trajectory characteristics of the vehicles at the intersection entrance, the improved S-G polynomial filtering algorithm is used for noise recognition and processing of the trajectory points, when the window fitting point is to the trajectory point of the parking state, a series of parking trajectory points are regarded as a point for fitting, if the parking point is an abnormal value, the fitting value is used to replace a series of position-invariant trajectory points. By integrating the abnormal parking state points into one point for subsequent processing, the influence of the abnormal parking state trajectory points on the overall motion trend fitting can be significantly reduced.

[0020] Other advantages, objects, and features of the present application will be apparent to those skilled in the art from the following specification, in conjunction with the accompanying drawings, and certain aspects of the application will be explained in greater detail below. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the specification and claims. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings, in which:

[0022] Figure 1 is a schematic flow chart of a two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics according to an embodiment of the present application;

[0023] Figure 2 is a schematic diagram of the device layout of a road intersection;

[0024] Figure 3 is a schematic diagram of IoU calculation;

[0025] Figure 4 (a) and (b) respectively represent the data in the sliding window in the traditional S-G filtering and the improved S-G filtering;

[0026] Figure 5 is the prior estimation, posterior estimation and measurement value result of a certain trajectory after the data is filtered by Kalman filtering;

[0027] Figure 6 (a) and (b) are the filtering of longitudinal distance and transverse distance using S-G polynomial filtering and improved S-G polynomial filtering;

[0028] Figure 7 is a schematic diagram showing the fitting of a certain parking vehicle trajectory using S-G polynomial filtering and improved S-G polynomial filtering;

[0029] Figure 8 is the noise trajectory point condition after two-stage joint filtering;

[0030] Figure 9 is a schematic diagram of a physical area of a road intersection. DETAILED DESCRIPTION

[0031] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments are only for illustrating the present application, and are not intended to limit the protection scope of the present application.

[0032] The present application proposes a two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics, that is, first, Kalman filtering is used for preliminary noise identification and processing to remove trajectory points with obvious errors, and then an improved S-G polynomial filtering algorithm is used for noise identification and processing of trajectory points according to the trajectory characteristics of vehicles on the approach road of the intersection.

[0033] Figure 1 is a schematic flowchart of a two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics according to an embodiment of the present application. In this method, the original trajectory point data of vehicles at the road intersection can be first collected by a data collection device. The data collection device can be a radar. Figure 2 is a schematic diagram of the device layout of a road intersection. Figure 2 The radar shown in the above figure is arranged on the approach road of the intersection. It should be noted that the position of the radar is only schematic and does not limit the present application. The arrangement of the radar can be arbitrarily set according to actual needs.

[0034] The original trajectory point data contains the ID number of the vehicle, the length and width of the vehicle, the speed of the vehicle (lateral speed, longitudinal speed, etc.), position information (lateral position, longitudinal position), and sampling time information, etc. Under normal circumstances, the original trajectory point data collected by the radar will contain data missing, data redundancy, and data anomalies, etc. Therefore, the original trajectory point data also needs to be preprocessed. The preprocessing can include data cleaning, that is, by discarding, filling, replacing, and / or deduplicating the data set, etc., so as to achieve the purposes of removing anomalies, correcting errors, and supplementing missing data.

[0035] In some embodiments, the original trajectory point data also contains repeated target data. The repeated target data can refer to data with the same sampling time, ID number, speed, etc. information, but the position information is similar in value. Therefore, such data also needs to be processed, i.e. removing the repeated target data. In some embodiments, the original data can be processed (e.g. data cleaning, etc.) according to the distribution characteristics of the sampling interval time to obtain a data set after removing the repeated sampling data, and then the processed data set is taken as the input of repeated target data identification processing to remove the repeated target data.

[0036] In some embodiments, the repeated data can be evaluated by calculating the IOU value (i.e. Jaccard index) between the data. IoU is one of the most commonly used evaluation indicators in the field of object detection, also known as jaccard index, which is the most commonly used measure to compare the similarity between two arbitrary shapes A and B. If the two objects overlap greatly, the similarity of the two data is greater, and if the two objects do not overlap, the value of IoU will be zero, i.e. non-repeated data. IoU can be expressed as:

[0037]

[0038] Figure 3 is a schematic diagram of IoU calculation.

[0039] In Figure 3 , (x1, x2) and (y1, y2) represent the position coordinates in the two target data, l1 and l2 represent the body length of the two vehicles respectively; w1 and w2 represent the body width of the two vehicles respectively; l a = x min + l xmin - x max , w a = y min + w ymin - y max , wherein,

[0040] x min = min(x1, x2),

[0041] x max = max(x1, x2),

[0042] y min = min(y1, y2),

[0043] y max = max(y1, y2),

[0044]

[0045]

[0046] Then, the IOU between two data can be calculated by the following formula

[0047]

[0048] wherein, l b = max(l a , 0), w b = max(w a , 0).

[0049] In some embodiments, the pre-processed trajectory point data can also be sorted according to the sampling time; the sorted trajectory point data is grouped according to the sampling time; then the grouped data set is traversed, for each group of data, the IOU value of any two data in each group of data is calculated and marked according to the IOU value (for example, data with IOU value greater than 0 can be recorded and marked), until the traversal of the grouped data set is completed.

[0050] After preprocessing, the pre-processed trajectory point data can be grouped according to the vehicle ID, each vehicle ID is associated with a group of data, so as to determine the trajectory point data of each vehicle, i.e. the driving trajectory of each vehicle.

[0051] Then, the two-stage joint denoising method is used to denoise the trajectory point data of each vehicle by traversing the ID grouped data set, until all the trajectory point data of the vehicles are identified and processed, and finally the data set of the denoised and processed intersection vehicle trajectory point data is obtained.

[0052] The two-stage joint denoising refers to the two-stage joint denoising of Kalman filter and improved S-G filter. First, Kalman filter is used for preliminary noise identification and processing to remove the trajectory points with obvious errors to obtain the first filtering result, and then improved S-G polynomial filter algorithm is used for filtering according to the trajectory characteristics of intersection vehicles to obtain the second filtering result.

[0053] The process of Kalman filter is described below.

[0054] The trajectory model based on vehicle dynamics in Kalman filter can be expressed by the following formula:

[0055]

[0056] V t = V t-1 + a t-1 · ΔT

[0057] wherein, S t and S t-1Vehicle position at time t and time t-1, V t and V t-1 Vehicle velocity at time t and time t-1, a t-1 Vehicle acceleration at time t-1, ΔT is the interval between time t and time t-1.

[0058] The system state equation and the observation equation in Kalman filter are as follows:

[0059] G t = AG t-1 + BU t-1 + W t , W t ~ N(0, M)

[0060] Z t = HG t + N t , N t ~ N(0, L)

[0061] Wherein, G and G t-1 are the system state at time t and time t-1; U t-1 is the state control quantity at time t-1; A is the state transition matrix, which converts the state at time t-1 to the state at time t; B is the control input matrix, which maps the effect of the state control quantity U t-1 to the state vector; W is the process noise subject to Gaussian distribution and expected to be 0, and the covariance is M; Z t is the observation value at time t; H is the measurement system parameter; N t is the measurement noise subject to Gaussian distribution and expected to be 0, and the covariance is L.

[0062] Combined with the kinematic model of the vehicle, the system state equation and the observation equation can be expressed as follows:

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] Wherein, p x,t and p x,t-1 represent the coordinates in the x direction at time t and time t-1, respectively;

[0069] p y,t and p y,t-1respectively represent the coordinates of the t time and the t-1 time in the y direction;

[0070] v x,t and v x,t-1 respectively represent the velocities of the t time and the t-1 time in the x direction;

[0071] v y,t and v y,t-1 respectively represent the velocities of the t time and the t-1 time in the y direction;

[0072] a x,t-1 and a y,t-1 respectively represent the accelerations of the t-1 time in the x direction and the y direction.

[0073] The Kalman filter uses continuous iterations of two processes, prediction and update, to estimate the true value of the system state and calculates the covariance matrix to represent the uncertainty of the prediction.

[0074] The iterative process of the Kalman filter is represented by the formula:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] wherein, represents the state prediction value of the t-1 time to the t time; represents the posterior state estimation value of the t-1 time; represents the prediction value of the system error covariance of the t-1 time to the t time; represents the posterior state estimation value of the t time; K t represents the Kalman gain coefficient calculated according to the covariance; L represents the covariance matrix of N t ; I represents the unit matrix.

[0081] The first filtering result generated after filtering the trajectory point data of each vehicle by the Kalman filtering method can be further filtered for the second time by the improved S-G filtering method.

[0082] The improved S-G filtering method is described below.

[0083] The SG filter design is based on polynomial fitting using a sliding window. The center point of the window is replaced by the center point of the polynomial to produce a smooth signal output. Traditional SG polynomial filtering uses a window to sequentially move the fitted data over the original data sequence until the entire data sequence has been traversed. However, in the trajectory sequences of intersection entrances or exits, when a vehicle stops and passes through, the trajectory sequence will contain a large number of identical trajectory points within a certain period, meaning that the lateral and longitudinal distances remain unchanged during the red light period. If the sliding window is used to fit these trajectory points sequentially, the large number of unchanging trajectory points will affect the fitting of the overall movement trend, disrupting the original stopping state movement trend, and making it difficult to determine whether there are errors in the trajectory points during the stopping time, thus affecting the filtering effect. Therefore, when the window fits to the trajectory points of the stopping state, a series of stopping trajectory points are treated as a single point for fitting. If the stopping point is an outlier, the fitted value is used to replace the series of trajectory points whose positions remain unchanged (or remain essentially unchanged). Figure 4 (a) and (b) represent schematic diagrams of the data in the sliding window in traditional SG filtering and improved SG filtering, respectively. By integrating the points of abnormal parking states into a single point for subsequent processing, the influence of trajectory points of abnormal parking states on the fitting of the overall motion trend can be significantly reduced.

[0084] In the improved SG filtering algorithm, it is necessary to identify the trajectory points in abnormal parking states (i.e., points whose coordinate positions remain unchanged or basically unchanged within consecutive timestamps) for the trajectory points in each sliding window. The trajectory points in abnormal parking states are integrated into one trajectory point. Then, the SG filtering method is applied to the integrated trajectory points in each sliding window to obtain the second filtering result.

[0085] Fitting point x based on the longitudinal displacement of the vehicle trajectory i For example, the sliding window width of its filter is k = n L +n R +1, x in the window i The number of dots on the left and right are n respectively. L and n R The polynomial z(x) fitted within the window i This can be represented as follows:

[0086]

[0087] Where, c0~c p The coefficients of the polynomial to be fitted are represented by , and p represents the set polynomial fitting order.

[0088] To perform polynomial fitting on k data points within a sliding window, the following matrix and vector are defined:

[0089]

[0090]

[0091]

[0092] wherein each element w in W i represents the longitudinal displacement measurement value (i.e., actual measurement data) of the trajectory point within the sliding window.

[0093] The cost function δ can be obtained from the above matrix, which can be expressed by the formula:

[0094]

[0095] It can be obtained that:

[0096] C=(M T M) -1 M T W,

[0097] Further, the estimated value after S-G filtering smoothing can be obtained wherein,

[0098]

[0099] According to the actual measurement data and the estimated value, the absolute difference Δθ between the two can be determined, which can be expressed by the formula:

[0100]

[0101] By traversing the trajectory point data of each vehicle, the trajectory filtering of the vehicle is completed.

[0102] After the two-stage joint filtering is completed, the noise can be identified and processed, and the de-noising processed trajectory data can be obtained after the noise data is eliminated.

[0103] The noise data identification and processing method of Kalman filtering is adopted to eliminate obvious error noise point data, and the improved S-G polynomial filtering algorithm is used for noise identification and processing of the trajectory point according to the trajectory characteristics of the vehicles at the intersection entrance or exit, the trajectory data missing is completed, and the noise point data is cleaned by taking the trajectory curve smoothness and the trajectory integrity as the evaluation indexes, through the overall vehicle motion trend and trajectory fitting. Taking the field radar data of the intersection of Fengtian Road in Chongqing Shapingba District as an example for analysis, the results show that, compared with the prior art, the method is more excellent in the smoothness of trajectory fitting, and the highest improvement of 40% can be achieved on the entrance of the intersection transverse distance, and the motion fluctuation of the fitted trajectory is the smallest in the transverse and longitudinal distances of the entrance.

[0104] In this implementation, the intersection is located in Fengtian Avenue, Shapingba District, Chongqing, near Chongqing Library subway station. The intersection is equipped with radar traffic data collection equipment, and is a cross intersection where the main road intersects with the branch road. The traffic flow of the main road is large, which is conducive to the collection of traffic data. The main road of the intersection is a six-lane two-way road, and the branch road is a four-lane two-way road. The street trees in the overall range of the intersection provide good shading and a wide field of view. The signs and markings are complete, and the existing signal lamp components are at an appropriate height, which can provide suitable installation conditions for the radar equipment. The intersection channelization and equipment layout can refer to Figure 1 .

[0105] The experimental data is a data set after removing repeated sampling, which has 1018025 data points and the collection period is 16:48:00-19:38:00.

[0106] The Kalman filter is used to preliminarily denoise the trajectory data set and repair missing values and repeated target data outliers. The purpose is to ensure the continuity of the trajectory when using S-G filtering and to eliminate the influence of trajectory points with large errors on the fitting effect of S-G polynomial filtering. The prior estimate, posterior estimate and measurement value of a certain trajectory after Kalman filtering are shown in Figure 5 . Figure 5 From Figure 5 , it can be seen that the Kalman posterior estimate value is smoother than the trajectory measurement value.

[0107] After preliminary denoising by Kalman filtering, S-G polynomial filtering is used for secondary denoising of the longitudinal distance and lateral distance of the trajectory data. Improved S-G polynomial filtering is performed on the trajectory data in the entrance and exit areas of the intersection and the physical area. After multiple experiments, the fitting window is selected as w=16 and the polynomial fitting order is selected as p=3. As shown in Figure 6 , the S-G polynomial filtering and improved S-G polynomial filtering are used to filter the longitudinal distance and lateral distance. Among them, Figure 6 (a) is the fitting of the longitudinal distance of the vehicle trajectory. It can be seen that the fitted trajectory of S-G polynomial filtering and improved S-G filtering has a high similarity to the original trajectory; Figure 6 (b) is the fitting of the lateral distance of the vehicle trajectory. The measurement value has a large fluctuation in this direction. Because the improved S-G polynomial filtering considers the parking condition (the longitudinal distance and lateral distance of the vehicle are fixed within a certain time range), the fitted trajectory points of the parking state are more matched with the overall vehicle trajectory motion state, i.e. the improved S-G polynomial filtering fitted trajectory is smoother. Figure 7 Figure 7 ​It can also be seen from the figure that the fitting trajectory of the improved S-G polynomial filtering method is more consistent with the kinematic characteristics of the vehicle.

[0108] By calculating the distance Δθ between the original trajectory point and the corresponding fitting trajectory point, the points greater than the threshold value are determined as noise trajectory points, as shown in the following formula: Figure 8 The noise trajectory points on the measured trajectory are marked and replaced by the fitting values. The black circles represent noise trajectory points, which are determined based on overall noise offset, so some noise points are close to the measured values. From the figure, it can be seen that the noise trajectory curve is smoother after the improved S-G filtering. Figure 8

[0109] The fitting comparison between the prior art and the method of the present application is performed, and the fitting trajectory evaluation indexes are shown in Table 1, and the fluctuation indexes of the vehicle trajectory motion characteristics are shown in Table 2. The prior art includes traditional S-G filtering, Kalman-traditional SG filtering (K-SG), and the method of the present application is Kalman-improved SG filtering (K-ISG). The physical area refers to the area where the vehicle turns or straightens, as shown in the following figure: Figure 9 .

[0110] Table 1 S-G filtering fitting trajectory evaluation index

[0111]

[0112] Table 2 original trajectory and S-G filtering fitting trajectory motion characteristic index

[0113]

[0114] It can be seen from Table 1 that in the longitudinal distance and transverse distance data of the intersection entrance, exit and physical area, the S-G filtering is better than the K-SG and K-ISG filtering in terms of curve fitting error indexes SSE and RMSE, but the smoothness is the worst, indicating that there is overfitting phenomenon. In the longitudinal distance and transverse distance of the entrance, compared with the K-SG filtering, the trajectory fitting error value of the K-ISG filtering is slightly larger, but the smoothness R performs better, and the fitting trajectory is more consistent with the vehicle motion trend; in the exit and physical area data, there is almost no parking phenomenon, and the effects of the two are basically the same.

[0115] It can be seen from Table 2 that in the longitudinal distance and transverse distance of the intersection entrance, exit and physical area, the motion fluctuation of the S-G fitting trajectory, the K-SG fitting trajectory and the K-ISG fitting trajectory is smaller than that of the original trajectory, and the K-SG fitting trajectory and the K-ISG fitting trajectory perform better, and in the transverse and longitudinal distance of the entrance, the motion fluctuation of the K-ISG fitting trajectory is the smallest, and in other data, it is not much different from the K-SG. ​

[0116] Finally, it is to be explained that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit the present application. Although the present application is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or equivalently replaced without departing from the purpose and scope of the technical solutions, which should be covered in the scope of claims of the present application.

Claims

1. A two-stage joint denoising method for intersection vehicle trajectory point data based on driving characteristics, characterized in that, The application relates to a method for processing vehicle trajectory point data of a road intersection, comprising the following steps: Step 1: collecting original trajectory point data of vehicles at a road intersection by a data acquisition device; Step 2: preprocessing trajectory point data in the original trajectory point data; Step 3: grouping the preprocessed trajectory point data according to vehicle IDs, each vehicle ID being associated with a group of data, so as to determine trajectory point data of each vehicle; Step 4: for the trajectory point data of each vehicle, adopting a two-stage joint denoising method for denoising, wherein the two-stage joint denoising comprises: filtering the trajectory point data of each vehicle by using a Kalman filtering method to generate a first filtering result; filtering the first filtering result by using an improved S-G filtering method to generate a second filtering result, and according to the second filtering result, performing noise trajectory point identification and processing to obtain a denoising result, wherein the improved S-G filtering method comprises: identifying trajectory points in an abnormal parking state for each trajectory point in a sliding window, and integrating the trajectory points in the abnormal parking state into one trajectory point; adopting an S-G filtering method for filtering the integrated trajectory points in each sliding window.

2. The two-stage joint denoising method for intersection vehicle trajectory point data based on driving features according to claim 1, characterized in that, The preprocessing comprises data cleaning, wherein the data cleaning comprises processing missing data, abnormal data and repeated data in the trajectory point data.

3. The two-stage joint denoising method for intersection vehicle trajectory point data based on driving features according to claim 1, characterized in that, The application further comprises sorting the preprocessed trajectory point data according to sampling time; grouping the sorted trajectory point data according to sampling time; for each group of data after grouping, calculating an IoU value of any two data in each group of data and removing repeated target data according to the IoU value.

Images