LU101647B1

LU101647B1 - Road pedestrian classification method and top-view pedestrian risk quantitative method in two-dimensional world coordinate system

Info

Publication number: LU101647B1
Application number: LU101647A
Authority: LU
Inventors: Lin Mao; Dawei Yang; Rubo Zhang; Yehao Xu; Junda Huang; Siyu Chen
Original assignee: Univ Dalian Minzu
Priority date: 2018-06-29
Filing date: 2019-06-24
Publication date: 2020-08-04
Also published as: WO2020001395A1; LU101647A1

Abstract

A method for sorting on-highway pedestrians and top-view pedestrian risk quantification method based on planar world coordinate system belong to the field of moving object tracking processing. In order to solve the problem of the current classification of road pedestrians, the in-vehicle camera acquire road pedestrian sequence. And classify road pedestrians based on the magnetic relationship between road pedestrians, which is characterized by the moving speed of the pedestrian, the relative distance between pedestrians in the sequence image, and the relative distance between the pedestrian and the camera. Classification of the moving speed, the relative distance between pedestrians in the image, and the relative distance between pedestrians and cameras, and the determination of the magnetic relationship can be used as an important reference for autonomous vehicles or assisted driving systems in obstacle avoidance and path planning.

Description

BL-5126 1 ROAD PEDESTRIAN CLASSIFICATION METHOD AND TOP-VIEW PEDESTRIAN 0 4’ RISK QUANTITATIVE METHOD IN TWO-DIMENSIONAL WORLD COORDINATE

SYSTEM Technical Field The invention belongs to the field of moving object tracking processing. Specifically, it is a classification method that uses magnetic model to distinguish the potential danger levels of pedestrian on the road and top-view pedestrian risk quantification method on planar world coordinate system. Background Technology Moving object tracking processing technology is an important research topic in the field of machine vision. With the application of autonomous vehicles and assisted driving systems, how to properly use object tracking processing technology to protect the safety of pedestrians and vehicles is also a hot research direction now. At present, when only using in-vehicle cameras, the classification of pedestrians by analyzing their historical trajectory and speed is a major way, which protect pedestrians and vehicles with using object tracking processing technology. First analyze the pedestrian trajectory and speed to calculate the probability of pedestrians colliding with vehicle, and then use the collision probability and establish corresponding classification rules to classify pedestrians into different types. Most of the present pedestrian classification methods in object tracking processing technologies calculate the probability that a pedestrian will directly collide with a vehicle, or only detect special pedestrians with specific characteristics. The patent application number is CN201610048194.1, and the name is "Blind Person Detection and Recognition Method and System based on Combined Features and Vehicle Camera". First use HOG feature to detect pedestrians, and then use the three classification features of guide dogs, blind sticks and blind glasses to detect the presence of blind people near the pedestrian detection results. The patent application number is CN201610048233.8, and the name is "Traffic

BL-5126 2 Police Detection Method and System based on Coat Feature and Posture Detection" LU101647 Traffic police caps and fluorescent vests are used as classification features to determine whether traffic police exist in pedestrian detection results.

Although it is possible to identify the object by detecting the specific features of the specific object, this method has a lot of limitations because it can only detect the specific object. in the article "Analysis of Pedestrian Collision Risk using Fuzzy Inference Model", Hariyono et al.

Analyzed moving speed and direction of pedestrians and vehicles, and relative distance between pedestrians and vehicles, calculated the collision probability between pedestrians and vehicles, and classified pedestrians by the size of the collision probability.

However, in order to accurately collect the speed and position information of pedestrians, this method uses a camera mounted on the side of the road to photograph the vehicle road conditions, which causes the method to be greatly affected by environment.

In the article "Estimation of Collision Risk for Improving Driver ’s Safety", Hariyono and others established a dangerous area in front of the car, and classified pedestrians who entered the dangerous area as dangerous pedestrians.

However, there are many unexpected situations on the road that pedestrians instantly change from ordinary pedestrians to dangerous pedestrians.

For example, while crossing a road the pedestrian holding a mobile phone suddenly decelerates, causing the pedestrian who was originally in a safe state to suddenly become the pedestrian collide with a vehicle.

The road pedestrian classification method used in our present invention is to establish a magnetic model, which analyze the physical relationship between pedestrians to reach judgment of potential danger of pedestrians, instead of estimating the danger of a direct collision between pedestrians and vehicles.

It establishes new road pedestrian classification rules to take those pedestrians with higher potential dangers as the main analysis object, thereby enriches the classification results of the current road pedestrian classification methods and better protects safety of vehicles and pedestrians.

In many areas of China, there is a long-term danger situation of mixed traffic with

BL-5126 3 people and vehicles. As a vulnerable group in road traffic, pedestrians account for a LU101647 large proportion of the fatality rates of accidents all year round, and should be protected by obstacles of vehicles. Therefore, it is self-evident that the improvement of cars' ability to evade pedestrian safety is of great importance.

Pedestrian risk analysis method based on automotive onboard system mainly uses sensors to sense vehicle environmental information, and combines the pedestrian object movement status to judge the pedestrian object danger and adjust the driving decision accordingly to achieve early protection of the dangerous pedestrian object. Pedestrian risk analysis methods based on on-board images are current mainstream research directions. Many researchers identify pedestrian poses and analyze pedestrian movement trends to classify dangerous pedestrians. Among them, Joko Hariyono and others used the optical flow method to segment the pedestrian outline, and used pedestrian posture comparison method to identify the horizontal movement trend of pedestrians, and it is determined that pedestrians moving towards the vehicle area are dangerous pedestrians. In addition, Keller and Gavrila and others used Gaussian dynamic process models and trajectory probabilistic hierarchical matching to identify the standing or horizontal motion state of pedestrian targets in the image.

Pedestrian risk analysis methods based on in-vehicle images mostly analyze the pedestrian risk directly from the image perspective. However, due to the distorted imaging of in-vehicle images “near large, far small’, researchers often can only recognize the posture of pedestrians, not the exact state of pedestrians. Accordingly, the existing pedestrian risk analysis methods usually can only give a qualitative two-category judgment of pedestrian danger or not. Therefore, its main purpose is to provide real-time warning for the driver, but not to provide precise data support for the decision of the car dealer.

In order to achieve accurate driver assistance and improve the intelligent autonomous cruise performance, Chinese Patent Application No. CN107240167A discloses a driving recorder pedestrian monitoring system, and a quantitative pedestrian risk analysis method is given. The system uses sensing equipment

BL-5126 4 including body-sensing controllers, infrared sensors, and detectors to obtain LU107647 pedestrian information in the vehicle environment, and calculates the pedestrian collision coefficient by matching the pedestrian depth image stream with the pedestrian object model, thereby realizing pedestrian danger warning. Although the invention gives a quantitative risk analysis result, the quantified risk factor is derived from the pedestrian's posture. In fact, the judgment result is intention of pedestrian collide with the vehicle. Therefore, the quantized coefficient does not possess the objective nature of kinematics, and it is not enough to reflect the degree of pedestrian's real sports risk.

No. CN104239741A Chinese patent application is based on the automobile driving safety assistance method of car risk field. From comprehensive perspective of people, cars, and roads, it analyzes the kinetic energy field, potential energy field, and behavior field of the driving environment to build a vehicle, fusion builds a driving risk field model of the risk of vehicles driving to obstacles, quantifies the driving risk of vehicles to road obstacles, and evaluates different degrees. The invention introduces a potential field theory to give the driving risk field a reasonable kinematic principle, so that the risk quantification result can be used objectively and effectively for driving decisions.

Contents of Invention To solve the problem of classifying pedestrian on the highway, the invention proposes an on-highway pedestrians classification method. The vehicular camera captures the image of on-highway pedestrians and these pedestrians are classified according to the magnetic relationship among pedestrians. This magnetic relationship is represented by the speed of the pedestrians, the relative distance between the pedestrians, and the relative distance between pedestrians and camera. Calculate the probability of each specific model in the magnetic relation displayed in the current image, judge that the probability of the specific model exceeds the probability threshold, then use the current model to reflect the magnetic relation of on-highway pedestrians in the current image. The definition of repulsion model is two pedestrians

BL-5126 5 moving in opposite direction have collision course among several travel routes, and LU101647 their speed slow down or even stop when pedestrians encounter, after the encounter the speed would recover. The definition of suction model is among a number of pedestrians walking in the same direction, two or more pedestrians walk in an overlapping or adjacent state with similar speed and direction. The definition of non-magnetic model among several pedestrians, there are at least three pedestrians, one of them is faster or slower than at least two other pedestrians.

The on-highway pedestrians sorting technique analyzes common road conditions and finds out pedestrians who are prone to sudden accidents but not necessarily collide with vehicles. According to the moving speed of pedestrians, the relative distance between pedestrians in the image and the relative distance between pedestrians and the camera, pedestrians are classified by magnetic force. The determination of the magnetic relationship can be used as an important reference for obstacle avoidance and path planning of autonomous vehicles or auxiliary driving systems.

In order to solve the problem of determining pedestrian target risk from the image perspective, the invention also proposes a pedestrian risk quantification method of overlooking the planar world coordinate system. The detailed process is as follow: Step 1. Calculate the track points of all pedestrian targets for the pedestrian images captured by vehicle-mounted cameras. The number of pedestrian targets is N, obtain and update all pedestrian target trajectory point vectors in real-time.

Step 2. All the vectors of the pedestrian trajectory are mapped to the top-view planar world coordinate system and obtain the corresponding N pedestrian trajectory vectors in the overlooking planar world coordinate system.

Step 3. Set the observation range of pedestrian movement and construct the pedestrian trajectory matrix corresponding to N pedestrian trajectory vectors in the top-view planar world coordinate system M,. From the top-view planar world coordinate system independently corresponds to the pedestrian path matrix

BL-5126 6 . LU101647 overlooking the planar world coordinate system M,. Step 4. Construct vehicular risk matrix of top-view planar world coordinate system M, and duplicate the same copy matrix Mj, of itself.

Step 5. For pedestrian target ie[l,N], calculate the risk coefficient of adjacent pedestrian R.

R =(M},,M})/k, k, is the number of pedestrian trajectory points.

Pedestrian risk quantification method is to quantify pedestrian target risk in vehicular video images.

Its function is to quantify the risk of driving to the pedestrian into a normalized risk index so as to provide an important decision-making data basis for the advanced assisted driving of intelligent vehicles and the pedestrian target obstacle avoidance function of autonomous cruise.

The beneficial effects of the algorithm include: (1) Pedestrian risk analysis uses the top-view planar world coordinate system with the advantage of intuitive perspective, which allows the driver to observe the movement trend of various pedestrians from a more accurate perspective. (2) The driving risk matrix in top-view planar world coordinate system describes a kind of static risk distribution in front of the vehicle.

Its risk distribution is related to the urban speed limit, and is not affected by the road environment and vehicle speed which reduces the complexity of practical applications. (3) Considering the motion of different pedestrian targets and vehicles in the top-view planar world coordinate system alone, the motion of pedestrians does not interfere with each other, so specific pedestrian targets can be given corresponding attention according to the attention needs of drivers or autonomous driving system. (4) The normalized adjacent pedestrian risk coefficient of pedestrian target is quantified, from 0 to 1, refiecting the different risk degrees of pedestrian target, which can be used to classify dangerous pedestrians and determine the priority of vehicle driving avoidance.

BL-5126 7 LU101647 Brief Description of the Drawings Figure 1 is logic diagram of on-highway pedestrians sorting technique based on magnetic model.

Figure 2 is schematic diagram of repulsion pedestrian classification results.

Figure 3 is schematic diagram of suction pedestrian classification results.

Figure 4 is schematic diagram of non-magnetic pedestrian classification results.

Figure 5 is schematic diagram of pedestrian classification results with different magnetic models.

Figure 6 is schematic diagram of this invention.

Figure 7 is image coordinate system.

Figure 8 is world coordinate system.

Figure 9 is top-view planar world coordinate system.

Figure 10 is parameter diagram 1. Figure 11 is parameter diagram 2. Figure 12 is trajectory point diagram in head-up view.

Figure 13 is pedestrian trajectory matrix diagram in top-view planar world coordinate system.

Figure 14 is driving risk matrix diagram in top-view planar world coordinate system.

Figure 15 is computation diagram of approaching pedestrians risk coefficients.

Figure 16 is computation result of approaching pedestrians risk coefficients in implementation plan 1. Figure 17 is computation result of approaching pedestrians risk coefficients in implementation plan 2. Figure 18 is computation result of approaching pedestrians risk coefficients in

BL-5126 8 implementation plan 3. LU101647 Detailed Descriptions The following is a further description of the invention in combination with the attached figures and the specific classification process.

The 1° implementation plan A Pedestrian of interest determining method based on magnetic model.

The logic block diagram is shown in Figure 1. In order to distinguish the pre-existing direct collision probability calculations for pedestrians and vehicles, the present invention estimates the potential danger of pedestrians by establishing a magnetic model.

It is a road pedestrian classification method based on magnetic model.

This method can be implemented by software. in the case of using only a vehicle-mounted camera, if the information of the pedestrian's moving speed, moving direction, distance from the camera, and relative position between pedestrians is known, the invention can be combined with a magnetic model to analyze and calculate the degree of compliance between pedestrians and classification conditions set by the magnetic model.

It divides pedestrians into ordinary pedestrians and magnetic pedestrians.

The classification results can further enrich the road pedestrian information obtained by the existing target tracking processing technology, and provide more comprehensive road condition reference information for autonomous vehicles and assisted driving systems.

To achieve the above objectives by the following technical solutions.

The first step is to enter the moving speed of the pedestrianv, the relative distance between the pedestrians in the images, and the relative distance between the pedestrian and the camera: . The second step is to build a magnetic model.

The magnetic model is divided into three sub-models: the repulsive model, the suction model and the non-magnetic model.

These three sub-models are independent of each other and do not affect each

BL-5126 9 other, and each corresponds to a magnetic probability: B (repulsive probability), P, LUT01647 (suction probability), Æ, (non-magnetic probability). The collection of the three magnetic probabilities is the total magnetic probability of the pedestrian Purr ={B, Bs, Pn} (4) Let magnetic probability § = 0.7, 6 can be a constant coefficient.

When the probability of P,P,or B, of a pedestrian's magnetic force exceeds § (in the case of a non-magnetic force probability of B,, Pn=1),the pedestrian is judged as a magnetic force pedestrian.

The magnetic model in the present invention sets the pedestrian to move from the left side to the right side of the image as a positive direction, and move the pedestrian from the right side to the left side as a negative direction. or the value of 5, take repulsive pedestrian as an example.

When two pedestrians who are facing each other and whose routes collide with each other are closer, the two pedestrians will be judged as repulsive.

This application is based on the general distance (with others 360cm or more) in the distance of the human social space as the basis for the determination, converted to a probability threshold of about 0.7. Similarly, when two in the same direction and close, can also be based on the human social space distance in the general distance (with others 360cm or more) as a basis for determination, converted to probability threshold is also about 0.7. Therefore, when pedestrians are relatively close to each other, this is used as the threshold for determining the probability of magnetic pedestrians.

Of course, for the social space distance can be in the implementation of the scheme with the current scheme as a reference, adaptive adjustment of the general distance, so that the adaptation of modification, or can set their own probability threshold. (1) Definition of the Repulsion model: among several pedestrians who collide in the opposite direction, the moving speed of the pedestrians will slow down or even stop when they meet, until the original speed is restored after the encounter.

The opposite direction of movement is similar to the magnetic same-sex repulsion effect, a

BL-5126 10 condition known as the repulsive model.

LU101647 The number of pedestrians in the repulsion model is at least two, and any pedestrian who travels in the direction of a collision of the route will construct a repulsion model.

The determination of the repulsion model is based on the distance between the pedestrians in direction and the distance between the pedestrians and the vehicle: First, compare the distance between two pedestrians facing each other and the vehicle to determine whether the two pedestrians’ routes will collide.

If the traveling paths of the pedestrians facing each other collide, the distance between the two pedestrians facing each other is compared to determine whether the two pedestrians are about to collide.

Equation (2) is the formula for calculating Repulsion probability Si di; . - . PB, = (m I +n” Ja, if vv; <0 (2) Type PR, is the repulsion probat between pedestrian i and the camera. z; is the distance between pedestrian j and the camera m and n are constant coefficients, and m,ne(0,1).v; and v; are the moving speeds of pedestrians i and j , respectively. vv; <0 means pedestrians i and ; are facing each other.

Repulsion probability only exists between pedestrians who are facing each other.

Repulsion probability will gradually increase when the pedestrians approach each other.

After the pedestrians staggered from each other, Repulsion probability will gradually decrease.

In formula (2), mand n are constant coefficients, and m,ne(0,1). In one embodiment of the present invention, m is 0.4 and n is 0.4. In response to pedestrian encounters on the road, pedestrians i and j move ata speed of 1.5 m/s in a 640 * 480 image, where the position of pedestrian i is (30, 90) and keeps moving forward, pedestrian ; Position is(600, 100) and keep moving negatively.

When pedestrians i and ; are approaching each other, the Repulsion probability P. will gradually increase.

When the two approach to a certain distance,

BL-5126 11 their Repulsion probability will exceed 5 to be judged as a magnetic pedestrian. LU101647 Pedestrian encounters are common on the road. At the same time, there is a possibility of a pedestrian collisions and an accident due to differences in pedestrian conditions. For example, there are pedestrians with mobile phones or pedestrians who rush to the road among pedestrians of encounter. The former is easily to distracted with others due to distraction, while the latter is equally easily to distracted with others due to faster movement speeds. So, by using the repulsion model to distinguish the encounter pedestrians in the pedestrians. (2) Definition of the Suction model: Among several pedestrians walking in the same direction, two or more pedestrians maintain overlapping or adjacent state of walking. The direction and size of their moving speeds are similar to each other, which is similar to the magnetic attraction effect. This situation is called Suction model.

The number of pedestrians in the suction model is at least two, and all pedestrians walking in the same direction who overlap or walk close to each other will constitute a suction model. Whether the pedestrian is walking in pairs is the main basis for determining the suction model. Whether the pedestrian is walking in pairs is compared by comparing the distance between pedestrians in the same direction and the distance between the pedestrian and the car. Equation (3) is the formula for calculating suction probability.

PB, = (m +n 415 J, if vv; >0 (3) where Æ, is the suction probability of pedestrian i.s; is the distance between pedestrians i and /.dj is the difference between the pedestrians i and jis the distance from the camera and dj =|z; —z;| Where z; is the distance between pedestrian i and the camera, and z; is the distance between pedestrian ; and the camera. v; and v; are the moving speeds of pedestrians i and /, and viv; >0 means that pedestrians i and ; are traveling in the same direction. k is

BL-5126 12 the suction parameter and k; =-[min(T|v;|.7}v;[)/s;], where T is the elapsed time. LU107647 min(T}v;.T|v;|) represents the smaller displacement of pedestrians i and j over time. The longer the pedestrians staying in the same direction stay relatively close and fixed relative distance, the greater the pedestrian's suction probability. In formula (3), m,n, and | are constant coefficients, and m,n,! € (0,1), in one embodiment of the present invention, m is 0.4, n is 0.4, and 7 is 0.2. For pedestrian companions on the road, pedestrians i and j; move at a speed of

1.5 m/s in a 640 * 480 image, where the position of pedestrian ; is (30, 90) and the position of pedestrian ; is (45 , 100), and they all keep moving forward. When pedestrians i and j are moving forward and their relative distances have not changed significantly, their suction probability will gradually increase. When the two move forward a certain distance, their suction probability will exceed 5 and be judged as a magnetic pedestrian.

On roads, it is common for pedestrians to walk in pairs, and walking in pairs will block the sight of each other, weakening pedestrians’ attention and increasing their danger. At the same time, the elderly and children often walk in groups on the street. Among them, children are very likely to cause traffic accidents on the road, and the elderly are one of the people who need to be looked after and courteous. Therefore, the Suction model is used to distinguish companion pedestrians.

(3) Non-magnetic model definition: Among several pedestrians, the speed of a certain pedestrian is significantly different from other pedestrians but does not necessarily satisfy the repulsion model or the suction model. It does not necessarily correspond to the case of repulsion or suction similar to the non-magnetic effect, this situation is called a non-magnetic model.

There are at least three pedestrians in the non-magnetic model. When one of the pedestrians has a speed higher or lower than at least two other pedestrians, the non-magnetic model is formed. The determination of non-magnetic model is mainly

BL-5126 13 based on the value of pedestrian moving speed.

Formula (4) is non-magnetic LU101647 probability calculation formula, P, Je (m) -v})>& (a) "10, others where F, is the non-magnetic probability of pedestriani,v;. v; and v, are the moving speeds of pedestrians i. j and h respectively.

Only when speed of pedestrians is higher or lower than at least two other pedestrians, and speeds of these three pedestrians satisfy the formula (1? _v7)+ (u? —v2)> €? (5) Pa; is equal to 1, otherwise itis 0. ¢ is a constant coefficient and ¢ = 1.5. For special groups of pedestrians on the road, they generally move slowly, in a 640*480 image, pedestrian i moves at a speed of 0.5 m/s, and pedestrians ; and h move at a speed of 1.5 m/s.

The position of pedestrian 7 is (30, 90), the position of pedestrian ; is (170, 240), the position of pedestrian } is (310, 420), and they all keep moving forward.

When the speed of these three pedestrians remains the same, pedestrian i 's non-magnetic probability P, will exceed the set § and thus be judged as a magnetic pedestrian.

For pedestrians moving faster on the road, in a 640*480 image, pedestrian i moves at a speed of 2 m/s, and pedestrians j and h move at a speed of 1.5 m/s.

The position of pedestrian i is (310, 420), the position of pedestrian j is (170, 240), the position of pedestrian » is (30, 90), and they all keep moving forward.

When the speed of these three pedestrians remains unchanged, the pedestrian i 's non-magnetism probability 7, willexceed & and will be judged as a magnetic pedestrian.

On the road, special groups and pedestrians on the road are also common.

Among them, special groups are one of the pedestrians who need attention and care.

They

BL-5126 14 often tend to move slowly, however, those pedestrians who move faster can cause LU101647 traffic accidents because of their faster speed, which probability of a traffic accident is much greater than that of a pedestrian walking normally. Therefore, the two types of pedestrians are distinguished by a non-magnetic model.

The magnetic probability of the same pedestrian may be different due to calculations with different pedestrians. At this time, the larger magnetic probability shall prevail.

Thirdly, Specific classification of pedestrians. After the calculation of the magnetic model, each pedestrian will have its own magnetic probability. According to the magnetic probability, pedestrians are divided into four categories: ordinary pedestrians, repulsion pedestrians, suction pedestrians, and non-magnetic pedestrians. Among them, repulsion pedestrians, suction pedestrians and non-magnetic pedestrians are three types of magnetic pedestrians. Setting (1) When P. exceeds 6, the pedestrian is a repulsion pedestrian; Setting (2) When P, exceeds 6, the pedestrian is a suction pedestrian; Setting 3) When P, exceeds 6, the pedestrian is a non-magnetic pedestrian. (4) Among the three magnetic probabilities, P, has the highest priority, P has the middle priority, and PF. has the lowest priority. It is known that pedestrian i 's magnetic probability is Pua, ={P,, Ps,» Bn, } - When non-magnetic probability exceeds 6, pedestrian i is an ordinary pedestrian; when only one magnetic probability exceeds 6, pedestrian i is judged as the corresponding type of magnetic pedestrian according to the above setting; when two or three magnetic probabilities exceed 5, pedestrian i is judged as a higher-priority magnetic pedestrian based on the priority of magnetic probabilities.

Through the above technical solution, the magnetic pedestrian model-based road pedestrian classification method provided by the present invention has the beneficial effects that:

BL-5126 15 The current classification method of road pedestrians mainly analyzes and LV101647calculates the direct collision probability between pedestrians and vehicles.

It ignores the complexity and variability of actual roads, cannot analyze road conditions well, and is greatly affected by the environment.

The invention adopts a magnetic model to classify road pedestrians, considers three common pedestrian situations that are prone to danger, analyzes some potential dangers of pedestrians, and realizes attention to some special groups on the road.

The present invention finds pedestrians prone to accidents by analyzing common road conditions and environments, and such pedestrians do not necessarily collide with vehicles, so they cannot be classified by existing methods.

Compared with the existing methods, the pedestrian classification result of the present invention can further enrich the pedestrian information obtained by the current road pedestrian classification method, and provide more sufficient road condition information for autonomous vehicles and driver assistance systems.

First use the existing road pedestrian classification method to find pedestrians that will directly collide with the vehicle, and then use the method of the present invention to find other pedestrians that are prone to danger.

In this way, when the autonomous vehicle or assisted driving system performs obstacle avoidance and path planning, which can get a more secure path plan.

In the path selection of autonomous cars, different types of magnetic pedestrians have different basic path choices.

Repulsion pedestrians and suction pedestrians will only affect the vehicle's path selection when they are on the vehicle's driving path.

For repulsion pedestrians, the vehicle will preferentially choose to pass between pedestrians when repulsion pedestrians intersect with each other; for suction pedestrians, the vehicle will preferentially choose suction pedestrian moves in the opposite direction.

When there is a non-magnetic pedestrian in front of the vehicle, regardless of whether the pedestrian is on the vehicle's driving route, the vehicle will choose to stop driving first, wait for the non-magnetic pedestrian to leave the front view of the vehicle, and then continue driving.

Because a non-magnetic pedestrian has two states of fast or slow movement, for a pedestrian who moves slowly, the pedestrian may be a special crowd and needs to be courteous; for a pedestrian who moves fast, even if the pedestrian is not in the

BL-5126 16 vehicle On the driving route, he may also move near the front of the car at a later time LV101647 and collide with the vehicle, so when a non-magnetic pedestrian appears, the vehicle will choose to stop.

The above is the problem of vehicle path selection when there is a single magnetic pedestrian.

When there are multiple magnetic pedestrians and the basic paths of these magnetic pedestrians conflict with each other, the vehicle will choose to stop going, otherwise it will choose to travel on the basic path.

For example, there are non-magnetic pedestrians and repulsion pedestrians in front of the vehicle, and the vehicle will choose to stop at this time; there are repulsion pedestrians and suction pedestrians in front of the vehicle, the suction pedestrian walks from the right to the left from the front view, and the suction pedestrian is on the right side of the repulsion pedestrian. the two path choices conflict, at this time the vehicle will choose to stop.

The above is the influence of magnetic pedestrians on vehicle path selection.

When dangerous pedestrians and magnetic pedestrians that collide with the vehicle appear at the same time, the vehicle will first consider a route that can avoid dangerous pedestrians, and then determine whether the route conforms to the magnetic pedestrian's path selection.

Stop if you meet.

Take the dangerous pedestrian in the middle of the front view as an example, and the pedestrian walks to the right of the front view.

In order to avoid the pedestrian, the vehicle can choose to pass from the left side of the pedestrian.

If there is a repulsion pedestrian on the right side of the vehicle at this time, it will not affect the route selection of the vehicle, because the repulsion pedestrian is not on the driving route.

If the repulsion pedestrian is on the left side of the front of the vehicle, the vehicle will choose to wait for the repulsion pedestrian to stagger and pass between them.

If there is a suction pedestrian on the right side of the car, no matter what direction he moves, it has no effect on the path selection of the vehicle.

If a suction pedestrian is walking on the left side of the vehicle and to the right of the front view, the vehicle will choose to drive from the left side of the suction pedestrian.

If a suction pedestrian is walking on the left side of the vehicle and to the left of the front view, the vehicle will choose to pass between the suction pedestrian and the dangerous pedestrian.

At the same time, it can be imagined that the suction pedestrian and the dangerous pedestrian may meet

BL-5126 17 the judgment requirements of the repulsion pedestrian at the previous moment, so the LU101647 driving route also meets the basic path selection of repulsion pedestrians.

If there is a non-magnetic pedestrian in front of the vehicle, the vehicle will stop moving regardless of whether the non-magnetic pedestrian is on the left or right side of the vehicle.

Implementation example 1 Repulsion pedestrian classification situation This implementation is for the classification of repulsion pedestrian, the simulation results are shown in Fig 2. Fig 2 lists three images in the continuous video frames and the pedestrian classification results of that frame.

The magnetic probability of pedestrian only conforms to the requirement of the magnetic force repulsion pedestrian.

In the video, the three pedestrian targets move at a speed of about 1.2m /s, two of which are moving in a positive direction and one is moving in a negative direction, and all of them keep moving in a straight line without changing the moving speed.

From frame 8 to frame 33, pedestrians B and C are approaching.

Until frame 33, the repulsion probability of pedestrians B and C exceeds, and they are determined as repulsion pedestrians.

In frame 72, the repulsion probability of pedestrians Aand C exceeds, and they are determined as repulsion pedestrians.

At this point, pedestrian B and C have finished the encounter process, the probability of repulsion decreases, and pedestrian B is judged as an ordinary pedestrian.

Implementation example 2 Suction pedestrian classification situation This implementation is for the classification of suction pedestrian, the simulation results are shown in Fig 3. Fig 3 lists three images in the continuous video frames and the pedestrian classification results.

The magnetic probability of pedestrian only conforms to the requirement of suction pedestrian.

In the video, the three pedestrian targets move in a forward direction at a speed of about 1.2m/s, and keep moving in a straight line without changing the moving speed.

From frame 11 to frame 39,

BL-5126 18 pedestrians B and C walk together. In frame 39, the suction probability of pedestrian B LU101647 and pedestrian C exceeds, and they are determined as suction pedestrians. At the next 75 frames, pedestrian B and C maintained the results for the suction pedestrian. Implementation example 3 Non-magnetic pedestrian classification situation This implementation is for the classification of non-magnetic pedestrian, the simulation results are shown in Fig 4. Fig 4 lists three images in the continuous video frames and the pedestrian classification results. The magnetic probability of pedestrian only conforms to the requirement of non-magnetic pedestrian. In the video, the three pedestrian targets move in the negative direction at different speeds, and keep moving in a straight line without changing the moving speed. The speed of pedestrian A is about 0.5 m/s. The speed of pedestrian B and pedestrian C is about

1.3 m/s. In frame 9, the non-magnetic probability of pedestrian A equals to 1, and he is determined as non-magnetic pedestrian. In the following frames 42 and 103, the results of pedestrian A remain the same. Implementation example 4 Blended magnetic pedestrian classification This implementation is for the classification of blended magnetic pedestrian, the simulation results are shown in Fig 5. Fig 5 lists three images in the continuous video frames and the pedestrian classification results. The magnetic probability of pedestrian conforms to the requirement of repulsion pedestrian and non-magnetic pedestrian. In the video, pedestrian A walks on the positive direction with the speed of

1.19 m/s. Pedestrian B walks on the positive direction with the speed of 1.83 m/s and the pedestrian C walks on the negative direction with the speed of 1.21 m/s. All of them are moving in a straight line without changing their speed. In frame 6, the non-magnetic probability of pedestrian B equals to 1, and he is determined as non-magnetic pedestrian. In frame 54, the repuision probability of pedestrian B and pedestrians C exceed. Because the non-magnetic probability of pedestrian B also

BL-5126 19 exceeds, pedestrian B is determined as non-magnetic pedestrian and pedestrians C LU101647 is determined as repulsion pedestrian. In frame 109, pedestrian B and pedestrian C have finished the encounter process, the probability of repulsion of pedestrian B is reduced and lower than 6, but his nonmagnetic probability hasn't changed and he's still classified as nonmagnetic. The probability of repulsion of pedestrian C and pedestrian A increases and exceeds J due to their constant proximity, and they are judged as repulsion pedestrians. The second implementation plan As shown in figure 6, the present invention discloses a top-view pedestrian risk quantification method based on planar world coordinate system. This method can classify and quantify road pedestrians by the road pedestrian classification method described in the first embodiment Pedestrian risk. This method can be implemented using software. By transforming the video of the on-board camera, the quantified risk level of the pedestrian target in front of the vehicle in the condition of overhead is solved.

The main implementation steps of this method are as follows: : Step 1: For images with size 1920x1080 (Unit: Pixel) calculate the pedestrian trajectory points of all N pedestrian targets on a frame-by-frame basis, and obtain and update all the pedestrian object head-up pedestrian trajectory point vectors FFF, at real time.

Step 2: Map all head-up pedestrian trajectory points to a planar world coordinate system, and based on the origin of the planar world coordinate system o,, use 10m horizontal distance and 0-20m vertical distance as the pedestrian motion analysis range to obtain corresponding N two-dimensional world coordinate systems pedestrian trajectory matrix Mr-M3. My Step 3, Copy N copies of the front risk matrix copy m! m2... mY of the top-view planar world coordinate system.

BL-5126 20 Step 4, For the pedestrian target i[1,N], use the formula R, =(M},M;})/k, to -4101647 calculate the neighboring pedestrian risk coefficient R. The present disclosure provides a detailed introduction to the above method. This method is aimed at the problem that it is difficult to accurately determine the pedestrian target risk by directly using the image angle of view. The principle is shown in figure 6. And calculate the risk before the vehicle in the planar world coordinate system. Furthermore, the pedestrian trajectory matrix and vehicular risk matrix are generated through quantitative mapping. Each pedestrian target has an independent pedestrian trajectory matrix and shares the same vehicular risk matrix to realize quantitative risk calculation and obtain normalized neighboring pedestrian risks for different pedestrian targets. The neighboring pedestrian risk coefficient is used as the output of a method for quantifying the risk of overhead pedestrians based on a two-dimensional world coordinate system, and can be used to support the driving decision module work of assisted driving and autonomous vehicles.

The technical solution of the present invention relates to the definition of related image coordinate systems, world coordinate systems, and camera parameters. For details, see figure 6, figure 7, figure 8, and figure 9 of the schematic diagrams.

Image coordinate system definition (see Figure 6): the first left corner of the image is the origin o, horizontal to the right is the u axis, vertical down is the v axis, defined as the image coordinate system.

Definition of the world coordinate system (see Figure 7): The point of origin of the on-board camera optical center as the origin O,, , the driving direction is the positive Y,, direction, the plane is coplanar with the vehicle driving plane and perpendicular to Y,, is the positive direction of the X, axis, and the direction of the camera is the positive direction of the Z,, axis , Defined as the world coordinate system.

Definition of two-dimensional world coordinate system (see Figure 8): The world coordinate system that ignores the world coordinate system Z axis (height axis) is

BL-5126 21 defined as the planar world coordinate system. LU101647 The present invention requires that the assembling method of the on-board camera is shown in fig. 7, which is mounted on the roof and faces the driving direction. The on-board camera requires dynamic shooting, so the camera's intrinsic parameters and assembly parameters are relatively fixed. The intrinsic parameters include the focal length f, the length of the photosensitive element d,, the width of the photosensitive element d,, the image length M , and the image width N. The assembly parameters include the height H, Heading angle y, elevation angle 0, horizontal aperture angle Alpha U, and vertical aperture angle Alpha V.

Internal parameter adaptation values of the present invention: The focal length / is 16mm-23mm; there is no special requirement for the size of the photosensitive element; the image length M is generally selected to be 1920 pixels and should not be less than 1080 pixels; the image width N is generally selected to be 1080 pixels and not less than 640 pixels. Assembly parameter adaptation value of the present invention: The adaptive range of the ground height H is between 1.2m and 1.6m; the ideal assembly angle of the yaw angle is 0°, and the acceptable range of the assembly error is +1"; The ideal assembly angle of the pitch angle is 0°, and the acceptable range of the assembly error is +3°. The calculation method of Alpha U for horizontal holes and AlphaV for vertical holes is: d AlphaU = arctanG +7) ST (1) AlphaV = arctan(—=-) 2x f First, the pedestrian trajectory points in the input image are transformed into the world coordinate system through inverse perspective mapping, and a planar world coordinate system pedestrian trajectory matrix is constructed. . Let p,(u,v,) be the pedestrian trajectory point of the t frame image of the input

BL-5126 22 ; ; ; . LU101647 video, where u and v, represent the column and row coordinates in the image; pr‘ (x.y) is the mapping coordinate of the pedestrian trajectory point of the t frame image of the video in the planar world coordinate system, where x and y represents horizontal and vertical coordinates in a planar world coordinate system.

According to this, F={pk e[r-k,t}} is the input video head-up pedestrian trajectory point vector, and Q= {pilt € [e = kl} is the vector F in the planar world coordinate system.

The mapping transformation steps of the head-up pedestrian track point vector F to the top-view pedestrian track point vector Q are:: Step 1, Calculate mapping factors rFactor and cFactor (see equation (2)), where wu and v are input values representing inverse perspective mapping points in the image, and M and N are fixed values representing image width and height; Qu rFactor = (1—-——) x tan(AlphaV) Mol (2) cFactor = (1-——)xtan(AlphaU) N-1 Step 2, Calculate the initial mapping point (x',y') of the planar world coordinates (see equation (3)), where C,, C,, and C. are fixed values representing the coordinates of the camera in the world coordinate system.

Usually, C, =C,=0 and C.=H areset; 0 is the pitch angle between the camera and the ground. r=Cx 1 + rFactor x tan(6) +C, tan(6) — rFactor (3) , cFactor + cos(0) y = C.

X ——oooue ete + C, tan(0) — rFactor Step 3, Correcting the initial point of a planar map of the world map coordinate system (x,y) coordinate point (see equation (4)), where y is the deflection angle of the camera setting represents.

BL-5126 23 x=x'cosy + y'sin y (4) LU101647 y=-x'siny + y'cosy Step 4, Use a matrix mapping function (as shown in equation (5)) to generate a pedestrian trajectory matrix M, - (n,m) = from (x,y) (5) In formula (5), (x,y) represents the coordinate points of the planar world coordinate system, and (n,m) represents the position of the elements in the operation matrix.

The purpose of constructing the pedestrian trajectory matrix M, is to represent the pedestrian trajectory point information within a limited distance in front of the vehicle in a planar world coordinate system by a matrix method.

Therefore, for the inverse perspective mapping effect, the mapping range of the planar world coordinate system to the operation matrix is set as O, Horizontal +10m and vertical 0-20m.

Based on this, a planar world coordinate system pedestrian trajectory matrix M, can be constructed as shown in fig. 11. Then, a planar world coordinate system vehicular risk matrix M, corresponding to the planar world coordinate system pedestrian trajectory matrix is constructed.

The risk equipotential lines of the planar world coordinate system are composed of six second-order curves about Y, , and satisfy: y=y(x)=ax’ +x +a, (6) In Equation (6), a,, a,, and a,are second-order polynomial coefficient vectors, and satisfy: a, = {-0.16,-0.42,—0.67,—1 .33,-2.33.—3.33} a, ={-0.61,-1.54,-2.47,-4.93,-8.63,—12.33}" (7) a, = {1,2.5,4,8,14,20}"

BL-5126 24 co of) LU101647 2 W, = 0p (0) =C +—— |e dy (8) RW 1 no J Given the vehicular risk weight calculation function affected by the vehicle front distance is shown in equation (8), The vehicular risk weight calculation function itself is a Gaussian distribution function. Among them, C, and C, are normalized parameters, and their values are set to C, =0.05 and C,=47.7; u and co are function expectations and variances, and their physical meaning is the risk distribution parameter affected by the braking capacity of the vehicle, and its value is set to y =0 and oc =8. In formula (8), w, is the normalized risk intensity. The closer the w, value to 1 in a certain area is, the more dangerous the area is. On the contrary, it tends to 0 to indicate a safer area.

Each coordinate in the planar world coordinate system is equipotential to the Yy -axis by formula (6), and the corresponding risk weight is obtained according to formula (8). The vehicular risk matrix is mainly used to match the pedestrian trajectory matrix to quantify the pedestrian risk coefficient. Therefore, the same matrix mapping function is used to construct the vehicular risk matrix. According to this, for the coordinated vehicular risk weights in the planar world coordinate system, further use Equation (5) to map to generate a vehicular risk matrix M, as shown in FIG. 12. Use a matrix mapping function to generate a front risk matrix M, : (n, m) = Sm, V,W,) (x,y,w,) represents the coordinate points of the planar world coordinate system and the corresponding risk intensity, and (n,m) represents the position of the elements in the operation matrix.

Finally, the planar world coordinate system pedestrian trajectory matrix M, and

BL-5126 25 the planar world coordinate system vehicular risk matrix M, are combining to LU101647 calculate the approaching pedestrians risk coefficient R.

Suppose that there are N different pedestrian targets in a continuous image, and there is a unique head-up pedestrian trajectory point vector F, corresponding to any pedestrian target ie[l,N]. Further, the vector F can obtain the pedestrian trajectory point vector Q, from above in step 2, and can independently correspond to the pedestrian trajectory matrix M} of the two-dimensional world coordinate system from the world coordinate system.

As shown in FIG. 13, the vehicular risk matrix of the top-view planar world coordinate system is copied by copying the same copy M}, as itself, and the pedestrian trajectory matrix M, of the top-view planar world coordinate system is used to quantify the Approaching pedestrians risk coefficient R;. The formula is: R = (Mj, M})/k, (9) Equation (9) is a quantified formula for the risk factor of approaching pedestrians in the present invention, where k, is the number of pedestrian trajectory points, and the output result R; is the approaching pedestrian risk coefficient of the pedestrian target i the closer R; is to 1, the more dangerous the pedestrian target is, and the closer it is to 0, the safer it is.

The physical meaning of the calculation method of formula (9) is to use the pedestrian trajectory matrix to screen the vehicular risk matrix, so as to obtain the pedestrian vehicular risk level corresponding to the position of the pedestrian track point.

The implementation example 1 In this embodiment, the vehicle pedestrian video of a measured road scene with a pixel size of 1920x1080 is used to quantify the neighboring pedestrian risk coefficients of two pedestrian objects in the image.

The calculation results of the

BL-5126 26 neighboring pedestrian risk coefficients can be seen in (a), (b), (c) and (d) in Figure 16. LU101647 Aiming at two pedestrian targets crossing the front of the vehicle in the image, a reasonable quantified result of pedestrian risk is given.

The implementation example 2 This implementation example targets two pedestrian objects in a vehicle video of a measured road scene of size1920x1080. The calculation results of the neighboring pedestrian risk coefficient are shown in (a), (b), (c), and (d) of Figure 17. It can be | seen that the present invention provides an accurate quantified result of pedestrian risk for pedestrian object traveling independently of the vehicle.

The implementation example 3 This implementation example is an in-vehicle video of a measured road scene image with a pixel size of 1920x1080, quantified two pedestrian objects in continuous images.

The calculation results of neighboring pedestrian risk coefficients are shown in Figure 18 (a), (b), (c), and (d). It can be seen that the present invention provides accurate quantification results of pedestrian risk for pedestrians crossing the area in front of the vehicle in the video image.

The above description is only to create a better specific embodiment of the present invention.

However, the scope of protection of the present invention is not limited to this.

Any person familiar with the technical field within the technical scope disclosed in the present invention, equivalent replacement or change according to the technical solution and the inventive concept of the present invention shall be covered by the protection scope of the present invention.

Claims

BL-5126 27 Claims LU101647

1. A method for sorting on-highway pedestrians, characterized in that on-highway pedestrian images are taken by vehicular camera, and on-highway pedestrians are classified by the magnetic relationship between on-highway pedestrians displayed in the image, the described magnetic relationship is represented by the moving speed of pedestrians, the relative distance between pedestrians in the image and the relative distance between pedestrians and camera; the described magnetic relationship is represented by the magnetic model which includes the repulsion model, the suction model and the non-magnetic model, calculating the probability of each specific model in the magnetic relation displayed in the current image, and judging the probability of the specific model exceeds the probability threshold, then the current model is used to reflect the magnetic relationship of pedestrians in the current image; the described repulsion model is defined as: a pedestrian facing each other has a collision path in a number of travel routes, and the pedestrian's movement speed slows down or even stops when pedestrians encounter, after the encounter the speed would recover, the described suction model is defined as: among a number of pedestrians walking in the same direction, two or more pedestrians are walking in an overlapping or adjacent state, and the direction and size of their moving speed are similar; the described non-magnetic model is defined as: in a number of pedestrians, there are at least three pedestrians, one of whom has a higher or lower speed than at least two other pedestrians.

2. The method for sorting on-highway pedestrians according to claim 1, characterized in that the specific method of pedestrian classification is as follows: calculating the magnetic probability of a pedestrian in the current image, Pure ={P-, Pa, Pn}

BL-5126 28 when no model probability exceeds the probability threshold 5, the pedestrian i is LU101647 an ordinary pedestrian; when only one model probability exceeds the probability threshold & , the pedestrian i is judged as the corresponding kind of magnetic pedestrian; when the probability of two or more models exceeds the probability threshold 6, the pedestrian i is judged as the magnetic pedestrian with higher priority according to the priority of the model probability; Rules: when P. exceeds the §, the pedestrian is repulsion pedestrian; when P, exceeds the 6, the pedestrian is suction pedestrian; when P, exceeds the 6, the pedestrian is non-magnetic pedestrian; among the three magnetic probabilities, PR, has the highest priority, FB. has the lowest priority, and the priority of PR, is the middle.

3. The method for sorting on-highway pedestrians according to claim 1, characterized in that the probability that the magnetic relationship of the pedestrian displayed in the current image is the repulsion model is calculated by the following formula: P. =(m" +n )/2, vy, <0 during the approaching of pedestrian i and pedestrian/ who move on the negative directions, the repulsion probability Æ, will gradually increase, when the repulsion | probability FP, exceeds the 6, pedestrian i and pedestrian j were determined as magnetic pedestrians, wherein: P, is the repulsion probability of pedestrian i; |

BL-5126 29 sy is the distance between pedestrian i and pedestrian j; LU101647 dj; is the difference in distance between pedestrian i and pedestrian j from the camera, and dj =|zi —z jl .z; is the distance between pedestrian i and camera, and z; is the distance between pedestrian j and camera; m and n are constant coefficients, and m,n € (0,1); v; is the moving speed of pedestrian i , and v; is the moving speed of pedestrian Js viv; <0 describes the moving directions of pedestrian i and pedestrian ; are opposite.

4 The method for sorting on-highway pedestrians according to claim 1, characterized in that the probability that the magnetic relationship of the pedestrian displayed in the current image is the suction model is calculated by the following formula: PF, =(m" +n +15 )/3. vy; >0 when pedestrian i and pedestrian / who have the same moving directions and keep walking in an overlapping or adjacent state, the suction probability 7, will gradually increase until the P, exceeds 6, pedestrian i and pedestrian j are determined as suction pedestrians; wherein: pis the suction probability of pedestrian i; s is the distance between pedestrian i and pedestrian J; d is the difference in distance between pedestrian i and pedestrian / from the

BL-5126 30 camera, and di =|zi —Z il.z is the distance between pedestrian i and camera, and LU101647 z; is the distance between pedestrian j and camera; m, n,and ! are constant coefficients, and m,n, € (0,1); v; is the moving speed of pedestrian i, and v; is the moving speed of pedestrian J; viv; >0 describes the moving directions of pedestrian i and pedestrian j are identical; ki; is the suction parameter, and ki = —[min(7}v;[. 7]v;|)/ si], T is the past time, and min(T}v;|, Tv |) is the smaller displacement between pedestrian i and pedestrian J after time 7.

5. The method for sorting on-highway pedestrians according to claim 1, characterized in that the probability that the magnetic relationship of the pedestrian displayed in the current image is the non-magnetic model is calculated by the following formula: 2 2 2 2 2 P = 1, (»; -V; )+(, -v, )>e - (0 others pedestrians i, j,h walk on the same direction, the speed of one pedestrian is significantly faster or slower than the other two pedestrians, and the rest of two pedestrians have the same moving speed, all of them keep their current speed ; If non-magnetic probability Fn exceeds the 6, pedestrian i is determined as non-magnetic pedestrians; wherein: Pn, is the non-magnetic probability of pedestrian 7; v; is the moving speed of pedestrian i, and v; is the moving speed of pedestrian J,

BL-5126 31 ; ; . a ie hi . LU101647 and y, is the moving speed of pedestrian 4; vw is higher or lower than v; and vy; if the vw, vj, and v, conform to the following equation, the Py, equals to 1, otherwise the value of Pn, is 0; (7 —vP)+ (v2 _y?)> e € is a constant coefficient.

6. Atop-view pedestrian risk quantification method based on planar worid coordinate system, which is characterized by the use of the road pedestrian classification method mentioned in any one of claims 1-5 to classify the road pedestrian and quantify the pedestrian risk, the risk quantification method includes the following steps: S1: for the pedestrian image taken by the vehicle, calculate the pedestrian tracking points of all pedestrian targets; obtain and update all pedestrian target head-up view pedestrian trajectory point vectors in real time; S2: all the pedestrian trajectory vectors are mapped to the planar world coordinate system, and the corresponding to N number pedestrian trajectory vectors are obtained; S3: set the pedestrian movement observation range, and construct a pedestrian trajectory matrix M, in planar world coordinate system corresponding to the N number of pedestrian trajectory coordinates vector in planar world coordinate system; from the world coordinate system, the pedestrian trajectory matrix M} in the top-view planar world coordinate system is corresponded independently; S4: construct vehicular risk matrix M, matching the pedestrian trajectory matrix, and duplicate the same copy matrix M}, of itself, S5: for any pedestrian i satisfying ;e[1,N], calculate the approaching pedestrians risk coefficients R;: _ a

BL-5126 32 R, = (My, ML) k, LU101647 wherein: k, is the number of pedestrian trajectory points.

7. The method for top-view pedestrian risk quantification based on planar world coordinate system according to claim 6, characterized in that the said step S2 is the mapping transformation of the vector F of the head-up view pedestrian trajectory point to the vector Q of the top-view pedestrian trajectory point; the specific steps are as follows: Step 1, calculate mapping factors rFactor and cFactor : rFactor =(1——**_)xtan( AlphaV) M1 P cFactor =(1- 2 x tan( AlphalU) N-1 wherein u and v are input values representing the inverse perspective mapping points, M and N are constant which represent image height and image width respectively; Alphal denote horizontal direction camera angular aperture, and AlphaV denote vertical direction camera angular aperture; Step 2, calculate preliminary mapping coordinate (x',y') in planar world coordinate system: x=Cx 1+ rFactor xtan(6) +C, tan(6)— rFactor y'=Cx cFactor + cos(6) +C * tan(0)-rFactor 7 | wherein CC, and C. are constant which represent the camera position in world coordinate system; C,=C,=0 and C.=H, H is the distance between camera and road surfaces, 6 is the angle between camera optical axis and road surface; Step 3 correct preliminary mapping coordinate to get mapping coordinate (x,y) in

BL-5126 33 planar world coordinate system: LU101647 x=x'cosy+y'siny y=-x'siny + y'cosy wherein y is used as a constant to indicate yaw angle; the said horizontal direction camera angular aperture Alphal and vertical direction camera angular aperture AlphaV can be calculate by follow equation: d AlphaU = arctan(—*-) 2x f AlphaV tan( a ) ay = arc — ? 2x f wherein /, d,, d, denote focal length, height length of light-sensitive component and width length of light-sensitive component respectively; the method of constructing a pedestrian trajectory matrix M, in planar world coordinate system corresponding to the N number of pedestrian trajectory coordinates vector in planar world coordinate system is by combining matrix transformation function that generate a pedestrian trajectory matrix M, ;the formula of function is shown below: (n,m) = fun (x,y) where (x,y) is mapping coordinate in planar coordinate system, and (n,m) denotes the position of the elements in the operation matrix.

8. The method for top-view pedestrian risk quantification based on planar world coordinate system according to claim 7, characterized in that pedestrian movement | observation coverage, in particular, is horizontal distance of +10 meters from Oy and vertical distance of 0~20 meters from O, in planar world coordinate system.

9. The method for top-view pedestrian risk quantification based on planar world

BL-5126 34 coordinate system according to claim 6, characterized in that the method for LU101647 constructing vehicular risk matrix M, matching the pedestrian trajectory matrix is: the said top-view coordinate in planar coordinate system is equalized to the Y, axis, and the risk weight of the corresponding coordinate points is calculated, and using the matrix transformation function generate vehicular risk matrix; the risk situation in the planar world coordinate system can be constituted by 6 second-order curve fitting with respect to Y axis symmetry; the fitting curve conform to the formula which is described below: y=r)=ax +ax+a, wherein a,,a,, and a, are second-order polynomial coefficient vectors, which satisfied the following equation: a, ={-0.16,-0.42,-0.67,-1.33, —2.33.—3.33} a, = {-0.61,—1.54,—2.47, 4.93, -8.63,-12.33} a, = {1,2.5,4,8,14,20}" the calculate function of vehicular risk matrix is described as below: c os) pa Ww, = 0m = C +—— |e > ‘d RW 1 Jo J y wherein C, and C, are normalized parameters; u and o are the expectation and variance of the function; w, is normalized risk weight; when the risk value closer to 1, the corresponding position of the value is more dangerous; otherwise, it is more secure; combining matrix transformation function generate a vehicular risk matrix M, : (n,m) = Jom (x, Y,W,)

BL-5126 35 wherein (x, y,w,) represents the coordinate points of the planar world coordinate LU101647 system and the corresponding risk intensity, and (n,m) denotes the position of the elements in the operation matrix.

10. The method for top-view pedestrian risk quantification based on planar world coordinate system according to claim 6, characterized in that the vehicular camera is fixed on the top of vehicle and is oriented towards the vehicle driving direction; the ideal range of height from road surface to camera H is 1.2 to 1.6 meters, the stable configuration angle of yaw angle is 0°, error acceptable range is +1°; the ideal configuration angle of pitch angle is 0°, error acceptable range is +3°.

| |