LU503001B1

LU503001B1 - Method for predicting pedestrian crossing without signal lamp control

Info

Publication number: LU503001B1
Application number: LU503001A
Authority: LU
Inventors: shuai Ling
Original assignee: Univ Tianjin
Priority date: 2022-11-04
Filing date: 2022-11-04
Publication date: 2023-05-04

Abstract

The invention relates to a method for predicting pedestrian crossing without signal lamp control, which is technically characterized in that: Step 1, collecting the track data of pedestrians and motor vehicle at intersections by using a data acquisition subsystem; Step 2, building a GBDT prediction model by using the model prediction subsystem, and preliminarily predicting the data collected by the data acquisition subsystem by using the GBDT prediction model; Step 3: optimizing the GBDT prediction model by using the model tuning subsystem and select the optimal parameters. By adopting GBDT technology, the invention can realize the prediction function of pedestrian crossing behavior according to the track data collected in real time. The invention can be applied to the vehicle-road cooperation scene in automatic driving, realizes the real-time prediction function of pedestrian crossing behavior at signalless road intersections, controls the vehicle speed, and is helpful for automatic -driving vehicles to make better real-time decisions according to road conditions when running.

Description

METHOD FOR PREDICTING PEDESTRIAN CROSSING

WITHOUT SIGNAL LAMP CONTROL

TECHNICAL FIELD

The invention belongs to the technical field of automation, and relates to a behavior modeling and predicting system of traffic participants, and in particular to a method for predicting pedestrian crossing without signal lamp control.

BACKGROUND

For a long time, the pedestrian crossing behavior at signalless intersections is an important issue related to the road safety. According to the World Health

Organization, more than half of all deaths related to traffic accidents occurred among pedestrians and cyclists.

The problem of road traffic fatalities is becoming more and more serious in low and middle income countries, where, the road traffic fatality rate in low-income countries is more than three times that in high-income countries; although the situation of middle-income countries is better than that of low-income countries, it is still not optimistic. The reason for this phenomenon is not only the higher proportion of vulnerable groups among road users in low and middle income countries, but also their lack of road safety awareness and related policies. Although various countries have introduced the relevant policy of "letting pedestrians go first", the yield of giving way is still unknown. Therefore, it is essential to understand whether and how pedestrians make safe crossing behaviors, and understanding the decision-making mechanism of pedestrians crossing streets is the most important thing to promote road safety. At present, the decision-making mechanism of pedestrian crossing behavior is mainly studied from two aspects: (1) Using the investigation based on psychological theory, this paper explores the formation mechanism of pedestrian crossing behavior from the perspective of human cognition. The Theory of planned behavior (TPB) is one of the widely used theories,

its main idea is to explore the influence of many subjective factors on the dangers HUS03001 caused by improper crossing behavior, which mainly comprise behavior attitude, subjective norms and perceived behavior control. (2) Using historical objective data and based on the traditional statistical learning model to study the technical methods of pedestrian crossing behavior, such methods mainly use non-track data to explore the waiting and non-waiting behaviors of pedestrians crossing the street.

The above-mentioned decision-making mechanisms can't realize the real-time prediction of pedestrian crossing behavior. They mainly consider subjective factors, but doesn't fully consider objective factors such as environment and vehicle characteristics. Therefore, the following problems exist: (1) Traditional statistical learning models mostly use non-track data to study the decision-making mechanism of pedestrians’ crossing behavior, which can't reflect the influence of the moving characteristics of vehicles and pedestrians on the decision-making of pedestrians’ crossing behavior, and can't dynamically depict pedestrians’ crossing behavior. (2) The existing related technologies of pedestrian crossing behavior prediction can't explain the importance of various predictors, and can't give the importance ranking of different predictors, which will affect the explanatory ability of the prediction model to some extent. (3) The prediction accuracy of the traditional model is not high, and there is the problem of unbalanced prediction results, which will mistakenly predict the 'go' behavior as the 'wait' behavior.

SUMMARY

The purpose of the invention is to overcome the shortcomings of the prior art, and provide a method for predicting pedestrian crossing without signal lamp control, which is fast, accurate, reliable and has a reasonable design.

The invention solves the existing technical problems by adopting the following technical scheme:

A method for predicting pedestrian crossing without signal lamp control HUS03001 comprises the following steps:

Step 1, collecting the track data of pedestrians and motor vehicle at intersections by using a data acquisition subsystem;

Step 2, building a GBDT prediction model by using the model prediction subsystem, and preliminarily predicting the data collected by the data acquisition subsystem by using the GBDT prediction model,

Step 3: optimizing the GBDT prediction model by using the model tuning subsystem and select the optimal parameters.

Further, the specific implementation method of Step 1 comprises the following steps:

Step 1.1, dividing the intersection area into an intersection area, a waiting area and a walking area, where the waiting area is an area with a width of 2m on both sides of the boundary line of the motor vehicle lane; the intersection area is the part of the motor vehicle lane except the waiting area; and the walking area is the area outside the waiting area;

Step 1.2, collecting the area by the data acquisition subsystem where the pedestrian location point falls, and judging the pedestrian crossing behavior, where the pedestrian crossing behavior comprises Wait and Go;

Step 1.3, the data acquisition subsystem stores data according to the following three model prediction variables: pedestrian characteristics, vehicle characteristics and environmental characteristics.

Further, the specific implementation method of Step 1.2 is as follows: for a pedestrian track and a specific position point of the pedestrian track, executing the following steps:

Step 1.2.1, if the position point falls in the walking area or the crossing area, entering the next position point;

Step 1.2.2, if the location point falls in the waiting area and there are no vehicles to pass by, entering the next location point;

. . . . _ . LU503001

Step 1.2.3, if the location point falls in the waiting area and there are vehicles to pass by, following the following steps:

Step 1.2.3.1, if the coordinates of pedestrians and passing vehicles at the current . x,,>x > . Xe >X 4 moment ! are satisfied “7! Mv and Pre 7 Vos , and satisfied “7° Y and

Voir 7 Votsi t+i ; ; ;

P. + at the next moment , recording all candidate factors, and generating x . . a sample 'go' mark, where “7 represents the abscissa of pedestrian position at the x . . moment *, “” represents the abscissa of vehicle position at the moment 7, Vp represents the ordinate of pedestrian position at the moment ‘, and Fos represents the ordinate of vehicle position at the moment ‘;

Step 1.2.3.2, otherwise, the data acquisition subsystem records all the candidate factors and generates a sample marked 'wait';

Step 1.2.4, moving to the next position point and repeating the above steps. After the position point of the current track is finished, moving to the next track data and repeating the above steps for the position point of the next track data.

Step 1.3, the data acquisition subsystem stores the track data of pedestrians and motor vehicles according to the pedestrian characteristics, the vehicle characteristics and the environmental characteristics.

Further, the pedestrian characteristics comprise the following variables: current the walking speed of pedestrians, the average speed of pedestrians from track start time (, to current time !, the maximum speed of pedestrians from track start time {, to current time !, the straightness rate of pedestrians' walking track, the waiting time of pedestrians after arriving at the roadside, and the number of pedestrians waiting to cross the street at the same time; the vehicle characteristics comprise the following variables: the current speed of the motor vehicle, the maximum speed of the motor vehicle from the track start time /, to the current time !, the speed variance of the motor vehicle from the track start time f, to the current time ’, and the number of motor vehicles to be passed; the environmental characteristics comprise the following variables: the distance between the passing motor vehicle and pedestrians in 7505001 the lane direction and the distance between the passing motor vehicle and pedestrians in the crosswalk direction.

Further, the specific implementation method of step 2 comprises the following 5 steps:

Step 2.1, the model prediction subsystem receives the data collected in step 1;

Step 2.2, generating GBDT prediction model: First, learning an initial pedestrian crossing behavior prediction model with data, and getting the predicted value of pedestrian crossing behavior and the residual error after prediction. Then, the previous prediction model learns the prediction model based on the residual error, and the next model builds a model on the gradient side where the residual error decreases, so that the residual error decreases in the gradient direction until the residual error between the predicted value and the real value is zero. Finally, the predicted value of the test sample is the accumulationof the predicted value of many previous prediction models.

Step 2.3, determining the minimum loss function: determining the previous integrated pedestrian crossing behavior prediction model as 7 ,, each newly added prediction model h, minimizes the loss function value through model fitting;

Step 2.4, adding the regular term to the GBDT prediction model.

Further, the specific implementation method of Step 3 comprises the following steps:

Step 3.1, the model tuning subsystem selects data sets to generate training sets: extracting N samples from the data sets collected by the data collection subsystem in bootstrap mode, using as training sets of each base classifier of the GBDT prediction model to generate base classifiers;

Step 3.2, selecting features and determining the best splitting point: setting M as the number of features of the input data, when splitting each node of the base classifier, firstly selecting m features from the M feature, and then selecting the best splitting point from the m features for splitting;

Step 3.3, calculating the basic model error: for the basic classifier, using OOB HUS03001 data as a test set to calculate its prediction error;

Step 3.4, calculating the overall model error: repeating steps 3.1 to 3.3, calculating the prediction errors of all base classifiers, and getting the prediction error of GBDT prediction model after averaging;

Step 3.5, selecting different parameter combination modes: calculating the prediction error of GBDT prediction model according to steps 3.1 to 3.4, and selecting the model with the smallest prediction error as the final training model,

Step 3.6, getting the model classification result by voting: using each base classifier to classify the input samples, and selecting the one with the most predicted samples as the classification result.

The invention has the advantages and positive effects that: 1. The method adopts the gradient boosting decision tree (GBDT) technology, collects the trajectory data of pedestrians crossing the street, and establishes a prediction model of pedestrian crossing behavior based on the collected data. Finally, the optimized prediction model can realize the prediction function of pedestrian crossing behavior according to the trajectory data collected in real time, and can give the importance ranking of the predicted variables, which is helpful for the automatic vehicle driving system to understand the pedestrian crossing behavior more deeply.

The method can be applied to the vehicle-road cooperation scene in automatic driving, realizes the real-time prediction function of pedestrian crossing behavior at signalless road intersections, controls the vehicle speed, and contributes to better real-time decision-making of self-driving vehicles according to road conditions when running. 2. By using the trajectory data instead of static data, the prediction result of the prediction system is more accurate than that of the previous model, and the influence of vehicle and pedestrian characteristics on pedestrian crossing decision is further explained. Moreover, the data acquisition subsystem can collect pedestrian data in real time, realize the dynamic prediction of pedestrian crossing behavior, and expand the three kinds of decision variables.

3. By using GBDT integrated learning model and based on the calculation 750500) method of variable importance, the invention can additionally identify the importance degree of different decision variables, give the importance ranking of different variables, and provide decision support for vehicle automatic driving system.

BRIEF DESCRIPTION OF FIGURES

FIG. 1 is an overall flow chart of prediction model generation of the present invention;

FIG. 2 is a schematic diagram of the division of crossing area, waiting area and walking area of the present invention;

FIG. 3 is a schematic diagram of a data acquisition area;

FIG. 4 is a position diagram of pedestrian data collection points in video images;

FIG. 5 is a schematic diagram of the walking tracks of motor vehicles and pedestrians in video images.

DESCRIPTION OF THE INVENTION

The embodiments of the present invention will be described in further detail below with reference to the drawings.

A method for predicting pedestrian crossing without signal lamp control, as shown in FIG. 1, realizes the pedestrian crossing prediction function through three operated systems: a data acquisition subsystem, a model prediction subsystem and a model tuning subsystem, and specifically comprises the following steps:

The specific implementation method of this step comprises the following steps:

Step 1.1, dividing the intersection area into an intersection area, a waiting area and a walking area.

As shown in FIG. 2, the waiting area is an area with a width of 2m on both sides of the boundary line of the motor vehicle lane, it is worth noting that some pedestrians don't wait outside the lane when crossing the street, but walk into the lane at a short distance to wait for the passing vehicles. Therefore, the waiting area is defined as the HUS03001 two-sided area 2m away from the lane boundary.

The intersection area is the remaining area of the motor vehicle lane, that is, the part of the motor vehicle lane except the waiting area.

The walking area is the area outside the waiting area, that is, pedestrians must pass through the walking area before reaching the waiting area.

Step 1.2, collecting the area by the data acquisition subsystem where the pedestrian location point falls, and judging the pedestrian crossing behavior (Wait or

Go, WOG).

As shown in FIG. 3 - FIG. 5, the data acquisition subsystem automatically records the pedestrian's walking track, and then judges the pedestrian's crossing behavior (Wait or Go,WOG) by checking the location points of the pedestrian's track at different times according to the following criterions. Where:

For a pedestrian track and a specific position point of the pedestrian track, perform the following steps:

Step 1.2.1, if the location point falls in the walking area or the crossing area, the pedestrian's street crossing behavior (WOG) at the location point cannot be determined, so entering the next location point.

Step 1.2.2, if the location falls in the waiting area and there are no vehicles to pass by, the pedestrian's street crossing behavior (WOG) at this location cannot be determined, so entering the next location.

Step 1.2.3, 1f the location point falls in the waiting area and there are vehicles that will pass by, then executing:

Step 1.2.3.1, if the coordinates of pedestrians and passing vehicles at the current moment ! are satisfied “7! 7% and Jr” Mv , and satisfied “pra 7 Eu and

Fra 7 Vois at the next moment ‘+! , the data acquisition subsystem records all candidate factors (pedestrian characteristics, vehicle characteristics and environmental characteristics) and generates a sample of 'go' mark. Where Xp. represents the

. . _ x . abscissa of pedestrian position at the moment ‘, 7“ represents the abscissa of vehicle position at the moment 7, Fos represents the ordinate of pedestrian position at the moment /, and Fos represents the ordinate of vehicle position at the moment f .

Step 1.2.3.2, otherwise, the data acquisition subsystem records all the candidate factors and generates a sample marked 'wait'.

Step 1.3, the data acquisition subsystem stores data according to the following three model prediction variables: pedestrian characteristics, vehicle characteristics and environmental characteristics. See the following table for specific variables:

Table 1-1. Predicting model variable

Variable

Variable classificati Variable explanations name on 0 (negative sample): rejecting the gap and waiting in

Decision place. flag a . variable 1 (positive sample): receiving the gap and crossing the street pvt Current walking speed of pedestrians

Average speed of pedestrians from track start time % to pv a

Pedestrian current time characteris Maximum speed of pedestrians from track start time f, . pv max tics = . to current time 7,

Straightness rate of pedestrians' walking track (the ratio ps ratio of the distance from the starting point of walking track to

. LU503001 the current point to the track length) pw Waiting time of pedestrians after arriving at the roadside

The number of pedestrians waiting to cross the street at p num the same time vvt Current speed of motor vehicle

Maximum speed of motor vehicle from track start time ; vv max

Vehicle {, to current time? characteris

Speed variance of the motor vehicle from the track start tics vv var time 7, to the current time / v num Number of motor vehicles to be passed

Environm Distance between the passing motor vehicle and dis v ental pedestrians in the lane direction characteris Distance between the passing motor vehicle and dis p tics pedestrians in the crosswalk direction

Step 2, building a GBDT prediction model by using the model prediction subsystem, and preliminarily predicting the data collected by the data acquisition subsystem by using the GBDT prediction model.

The method adopts the gradient boosting decision tree (GBDT) technology to realize the prediction function. GBDT is a Boosting method, its main idea is that each time the model is established in the gradient descent direction of the loss function of the previously established model. Loss function is to evaluate the performance of the model (generally, fitting degree + regularization term), it is considered that the smaller the loss function, the better the performance, if the loss function keeps decreasing, the model can be continuously modified to improve its performance. So the best way is to decrease the loss function along the gradient direction.

The concrete implementation method of this step comprises the following steps:

Step 2.1, data input. Inputting the data collected by the data acquisition HUS03001 subsystem, and selecting the appropriate loss function to measure the fitting effect of

GBDT model.

Assuming that x, is a set of decision variables, the expression of target predicted value y; is as follows: =F) =X, hx) where, h represents the base classifier, and the learning ability of the base classifier is weak.

GBDT model is a greedy prediction model, and its expression is as follows:

Fx)=F _,(x)+h, (x)

Step 2.3, determining the minimum loss function: determining the previous integrated pedestrian crossing behavior prediction model as / ,, each newly added prediction model 4, minimizes the loss function value ZL, through model fitting; where h, is satisfied the following formula: h, =argmin L, =arg min)" /( y, FE (x)+h (x); through the first-order Taylor formula expansion, / is approximately expressed as the following formula:

U Fa ER CD EU By (00) +, (0) [5652]

for a binary classification problem, because the predicted values of regression HUS03001 tree are continuous, for the value of a predicted value (0 or 1), F5) = Zen (x) is not exactly the same. The mapping from the value Fy (x) to a certain classification depends on the loss function. The probability that the predicted value of a given target value Yi is positive classification is PO, =1lx)=0F (x, ), where

T represents the sigmoid function.

Step 2.4, to solve the problem of over-fitting, adding a regular term to the GBDT prediction model. The learning rate (contraction factor) can be used as a regular term to reduce the contribution of each basic prediction model. Determining model learning rate as S(0<5< D3 the model expression after adding the regular term is as follows:

F(x) =F, (x)+&%h, (x)

As the learning rate (contraction factor) decreases, the optimal loss function of the training model will also decrease, and the fitting effect of the model will be better.

However, this will require the predicting system to add more basic predicting models.

Therefore, it is necessary to weigh the number of basic models and the learning rate.

Another important parameter for GBDT model is the complexity of the prediction model (decision tree), and the complexity represents the depth of the prediction model.

In order to capture more complex relationships among explanatory variables, it is necessary to increase the depth of the model. In a word, the optimal performance of the prediction model depends on the overall selection of three parameters: the number of basic models, the learning rate and the complexity of the models.

GBDT prediction model can build a forest by integrating many decision trees, and there are many decision trees in the forest. GBDT belongs to bagging integration algorithm, and adopts bootstrap method to randomly extract training data. Assuming that the size of a data set is N, performing N times random sampling with a return and as the training set of a tree. According to the knowledge of probability theory, about HUS03001 1/3 of the data in the data set is not selected, which is called Out of bag (OOB). Then 1/3 of the OOB data can be used as the test set of the base classifier, so there is no need to divide the training set and the test set for GBDT algorithm.

The specific implementation method of this step is as follows:

Step 3.1, selecting data sets to generate training sets: extracting N samples from the data sets collected by the data collection subsystem in bootstrap mode, using as training sets of each base classifier of the GBDT prediction model to generate base classifiers;

Step 3.3, calculating the basic model error: for the basic classifier in (1), using

OOB data as a test set to calculate its prediction error;

It should be emphasized that the embodiments described in the present invention are illustrative rather than restrictive, so the present invention comprises but is not limited to the embodiments described in the specific embodiments, and other embodiments obtained by those skilled in the art according to the technical scheme of the present invention also belong to the scope of the present invention.

Claims

CLAIMS LU503001

1. A method for predicting pedestrian crossing without signal lamp control, characterized by comprising the following steps: Step 1, collecting the track data of pedestrians and motor vehicle at intersections by using a data acquisition subsystem; Step 2, building a GBDT prediction model by using the model prediction subsystem, and preliminarily predicting the data collected by the data acquisition subsystem by using the GBDT prediction model, Step 3: optimizing the GBDT prediction model by using the model tuning subsystem and select the optimal parameters; the specific implementation method of Step 1 comprises the following steps: Step 1.1, dividing the intersection area into an intersection area, a waiting area and a walking area, where the waiting area is an area with a width of 2m on both sides of the boundary line of the motor vehicle lane; the intersection area is the part of the motor vehicle lane except the waiting area; and the walking area is the area outside the waiting area; Step 1.2, collecting the area by the data acquisition subsystem where the pedestrian location point falls, and judging the pedestrian crossing behavior, where the pedestrian crossing behavior comprises Wait and Go; Step 1.3, the data acquisition subsystem stores data according to the following three model prediction variables: pedestrian characteristics, vehicle characteristics and environmental characteristics; the specific implementation method of Step 1.2 is as follows: for a pedestrian track and a specific position point of the pedestrian track, executing the following steps: Step 1.2.1, if the position point falls in the walking area or the crossing area, entering the next position point; Step 1.2.2, if the location point falls in the waiting area and there are no vehicles to pass by, entering the next location point;

Step 1.2.3, if the location point falls in the waiting area and there are vehicles to pass by, following the following steps: Step 1.2.3.1, if the coordinates of pedestrians and passing vehicles at the current

. x,,>x > . Xe >X 4 moment ! are satisfied “7! Mv and Pre 7 Vos , and satisfied “7° Y and > ; ; . . . Vern 7 Yes at the next moment ‘+! , recording all candidate factors, and generating x . . a sample 'go' mark, where “7 represents the abscissa of pedestrian position at the x . . moment *, “” represents the abscissa of vehicle position at the moment 7, Vp represents the ordinate of pedestrian position at the moment ‘, and Fos represents the ordinate of vehicle position at the moment ‘; Step 1.2.3.2, otherwise, the data acquisition subsystem records all the candidate factors and generates a sample marked 'wait'; Step 1.2.4, moving to the next position point and repeating the above steps; after the position point of the current track is finished, moving to the next track data and repeating steps 1.2.1 to 1.2.4 for the position point of the next track data.

2. The method for predicting pedestrian crossing without signal lamp control according to claim 1, characterized in that the pedestrian characteristics comprise the following variables: the current walking speed of pedestrians, the average speed of pedestrians from track start time 7, to current time ‘, the maximum speed of pedestrians from track start time / to current time ’, the straightness rate of pedestrians’ walking track, the waiting time of pedestrians after arriving at the roadside, and the number of pedestrians waiting to cross the street at the same time; the vehicle characteristics comprise the following variables: the current speed of the motor vehicle, the maximum speed of the motor vehicle from the track start time f, to the current time , the speed variance of the motor vehicle from the track start time K, to the current time ’, and the number of motor vehicles to be passed; the environmental characteristics comprise the following variables: the distance between the passing motor vehicle and pedestrians in the lane direction and the distance HUS03001 between the passing motor vehicle and pedestrians in the crosswalk direction.

3. The method for predicting pedestrian crossing without signal lamp control according to claim 1, characterized in that the specific implementation method of step 2 comprises the following steps: Step 2.1, the model prediction subsystem receives the data collected in step 1; Step 2.2, generating GBDT prediction model: first, learning an initial pedestrian crossing behavior prediction model with data, and getting the predicted value of pedestrian crossing behavior and the residual error after prediction; then, the previous prediction model learns the prediction model based on the residual error, and the next model builds a model on the gradient side where the residual error decreases, so that the residual error decreases in the gradient direction until the residual error between the predicted value and the real value is zero; finally, the predicted value of the test sample is the accumulation of the predicted value of many previous prediction models; Step 2.3, determining the minimum loss function: determining the previous integrated pedestrian crossing behavior prediction model as F |, each newly added prediction model h, minimizes the loss function value Z, through model fitting; Step 2.4, adding the regular term to the GBDT prediction model.

4. The pedestrian crossing prediction method under the control of no signal light according to claim 1, characterized in that the concrete implementation method of step 3 comprises the following steps: Step 3.1, the model tuning subsystem selects data sets to generate training sets: extracting N samples from the data sets collected by the data collection subsystem in bootstrap mode, using as training sets of each base classifier of the GBDT prediction model to generate base classifiers; Step 3.2, selecting features and determining the best splitting point: setting M as the number of features of the input data, when splitting each node of the base classifier, firstly selecting m features from the M feature, and then selecting the best HUS03001 splitting point from the m features for splitting; Step 3.3, calculating the basic model error: for the basic classifier, using OOB data as a test set to calculate its prediction error;

Step 3.5, selecting different parameter combination modes: calculating the prediction error of GBDT prediction model according to steps 3.1 to 3.4, and selecting the model with the smallest prediction error as the final training model; Step 3.6, getting the model classification result by voting: using each base classifier to classify the input samples, and selecting the one with the most predicted samples as the classification result.