KR101133222B1 - System and Method for monitoring traffic and accident based on image processing - Google Patents

Abstract

The accident detection system according to the present invention performs a preliminary detection by applying a Gaussian mixed model through the image when a stationary vehicle is present due to an accident, and then determines the accident confirmation by referring to the speed and trajectory change of the individual vehicle. How to carry out accurate accident detection.

Description

System and Method for monitoring traffic and accident based on image processing

The present invention relates to an accident detection system and method, and more particularly to an image-based accident detection system and method.

Increasing use of vehicles is increasing traffic problems such as traffic congestion, accidents and air pollution.

In particular, in the trend that the number of vehicles that the road can afford due to the increase of the vehicle gradually decreases, the traffic congestion caused by the accident vehicle worsens the traffic flow of the entire road, so it is necessary to promptly detect and take measures. In addition, in the case of an accident that occurs on the highway, the traffic flow is often worsened by the second and third accidents caused by the accident vehicle. Normally, 20-50% of accidents occurring on highways are connected to secondary accidents, and the traffic flow in the event of accidents is more affected by non-recurrent congestion than by recurrent congestion. AID) is a reality that needs a system.

However, the current accident detection system is based on the loop detector. The loop detector should determine whether an accident has occurred based on information such as traffic volume, vehicle occupancy, and speed of the vehicle driving on the road. However, when determining whether an accident has occurred in a loop detector-based system, if the congestion of the vehicle is not caused by a traffic accident, vehicle congestion caused by repair work occurring on the road and a cause not related to other accidents If a car accident occurs, the occurrence of a traffic accident cannot be judged properly.

Therefore, at present, a method of determining an accident using a loop detector includes a section in which it is determined that an accident has occurred based on traffic information (average speed of a vehicle driving on a road, a road share of a vehicle, and a traffic volume passing through a road). When present, CCTV must be used to directly determine the road on which the loop detector is located. The accident detection system using the loop detector has a side that is difficult to write a sequential record or determine the cause of the situation before and after the current occurrence of the accident. This takes a lot of time to determine the cause of the accident and makes clear analysis difficult.

The traffic accident detection system and method developed and applied to the road until now are largely using a model-based accident detection algorithm and an image-based accident detection algorithm.

First, a traffic incident detection algorithm based on model is a method of designing a mathematical model according to traffic information measured using a detector (for example, a loop detector) installed on a road, and the model is a pattern recognition. (pattern recognition, hydromechanics, statistical forecast, filtering model).

Intelligent detection algorithm without model is a neural network that uses intelligent detection algorithms using artificial intelligence because a model cannot fully include the characteristics of nonlinear traffic flows. We apply complex methods such as fuzzy-logic algorithms.

The model-based accident detection system and method mainly perform a traffic detection by analyzing continuous traffic flow data for each point using a plurality of loop detectors hypothesized on a road. In the case of the microscopic AID system of the TVS group, which is applied to the model-based accident detection system, each vehicle is tracked from one point where the loop detector is hypothesized to the next point where the vehicle passes one point and then the next point. By comparing the speed between the predicted time and the actual value, the value is judged to be an accident if the value exceeds the significance level. When performing an accident detection using a microscopic AID system, the tracking accuracy of each vehicle must be accompanied, but when the vehicle changes lanes, speed comparison using the vehicle is meaningless.

Meanwhile, the LA Freeway Surveillance and Control Project, which began in 1971, conducted approximately 700 loop detectors on 42 miles of Santa Monica freeway, San Diego freeway and Harbor freeway. . At this time, when the accident occurs by comparing the vehicle occupancy rate (the ratio of the vehicle occupied on the road) between the loop detectors, the road occupancy rate of the loop detector hypothesized at a point earlier than the point where the accident occurred, The accident detection was carried out based on the theory that the road share rose sharply and the road share at a later point dropped sharply.

In the experiments on the Tokyo Expressway Control System, accidents are detected by comparing the states of two consecutive detectors. At this time, a moving average method of 5 minutes or more has been applied. However, the experiment on the okyo Expressway Control System can also infer whether an accident occurred by referring to the change in the total traffic volume, but it is solved in terms of the accuracy of the accident detection because the change in the traffic volume is not always directly related to the accident. I have left a task.

On the other hand, as a method of performing accident detection by analyzing an image, a knowledge-based accident detection algorithm has been proposed. The knowledge-based accident detection algorithm has the advantage of high detection rate, detection in a short time, and direct identification of the accident site.

These knowledge-based algorithms perform accident detection using a one-class classification method, and establish a regional distribution and compare the feature space and the regional distribution of the accident domain. As a method of determining a decision function and judging an accident according to the conditions of this decision function, there is an advantage that a prior statistical hypothesis about the regional distribution is not necessary.

Due to the lack of universality of existing algorithms, Kun Z has developed a universal incident detection framework that consists of Baysian networks based on expert traffic knowledge that can cover both freeway and arterial roads. Suggested.

In addition, D. Kim proposed a combined model of fuzzy logic and Adaptive Resonance Theory as a fuzzy ARTMAP algorithm that complements the problem of backpropagation theory. This is based on the results of learning the average speed and the traffic volume of a moving vehicle for 30 seconds using neural network input vectors. However, there are many difficulties in detecting an accident that has a hit rate of more than 95% by the average speed and the traffic volume as an accident detection factor that occurs in the actual traffic situation.

However, if the distance between detectors is wide (typically 500m), such knowledge-based algorithms will rapidly reduce the detection efficiency.If the traffic volume is very low, the accident detection rate will be low, and if the traffic volume is very heavy, Frequent occurrences of congestion occur. In addition, even if an accident can be detected, there is a disadvantage in that the cause of the accident is not identified and the real-time situation is not identified.

On the other hand, the VISTRAM system of Tsinghua University in China collects a profile of a part of a video frame corresponding to a part of a road lane and a vertical and horizontal direction, and reconstructs it into the form of an image, and whether the vehicle is present in the reconstructed image. We have proposed a way to detect accidents by extracting information such as vehicle type and vehicle speed. This method is able to produce stable data about changes in lighting and some obstructions, but it is known to be inadequate for extracting various information through the tracking of accident vehicles or individual vehicles.

In addition, Shunsuke Kamijo attempted to detect accidents through image processing and based on the hidden Markov model (HMM) model, using the Markov random field (MRF) algorithm to detect accidents through tracking of intersection vehicles. The vehicle detection rate was 90%. The HMB model is an estimate of the accuracy of accident detection through vehicle extraction and has not been verified in actual road conditions.

Traficon Autoscope is commercialized as a video accident detection system, and most of these products are spot detection, stationary vehicle detection, and roadside detection, such as abnormal road driving detection. It is introduced. However, the method of using the general video frame used by Autoscope has the advantage that it can calculate various information, but stable information can be extracted only when the problems such as obscuration, shadow, and light change are solved.

Most accident detection methods up to now are analyzed by analyzing the change of traffic situation rate. The data provided by the loop detector is mainly analyzed, but recently, the accident detection algorithm through video is considered a lot. Is in. In particular, the accident detection by the image can immediately see the accident situation, it can also grasp the degree of triangularity.

Therefore, the applicant replaces the existing loop detector method that shows the management and operation limitation by using the image-based accident detection system, accurately records the sequential situation before and after the traffic accident, An objective and clear analysis of the accident mechanism of traffic accidents is possible. The present applicant intends to propose a reliable accident detection system that obtains a high detection rate and shows a low judgment error based on the image. In addition, we will propose an image-based accident detection system and method that can accurately interpret the accident location and the accident situation, and the accident detection rate does not drop even in bad weather such as cloudy weather, rain, fog, and snow, which have a high risk of accidents.

An object of the present invention is to provide an accident detection system and method with high accuracy and low error determination by using the captured image to minimize the influence of weather or traffic information.

Another object of the present invention is to provide an image-based accident detection system and method for determining whether an accident occurs by distinguishing a road from a vehicle using a Gaussian Background Mixture Model and tracking the state of the separated vehicle.

In addition, another object of the present invention is to provide an accident detection system and method based on the image to collect the various traffic information generated on the road, as well as the accident detection of the vehicle occurring on the road.

The above object is, by applying a Gaussian mixture model (Gaussian mixture model) to the traffic image obtained by the camera to obtain an image of the road as a reference image, the preliminary detection unit for detecting a vehicle in a stagnant state in the reference image, And a definite detection unit configured to second determine whether an accident has occurred in the vehicle with reference to the traffic information around the vehicle.

The above object is obtained by applying a Gaussian mixture model to a traffic image obtained by a camera to obtain a reference image of a road, and satisfies a preset time interval among frames constituting the reference image. And detecting the vehicle in the stationary state from the reference image with reference to the frames, and finally determining whether an accident has occurred in the vehicle by referring to the surrounding traffic information of the detected vehicle.

The accident detection system and method according to the present invention perform a preliminary detection by applying a Gaussian mixture model through the image when a stationary vehicle due to an accident exists, and then refer to the speed and trajectory change of the individual vehicle, By determining the confirmation, accurate and quick accident detection is possible.

1 is a system conceptual diagram of an image-based accident detection system according to an embodiment of the present invention.
2 is a reference diagram for an example of a pixel value distribution of a traffic image.
3 illustrates that frames B0 to Bn of reference frames are arranged among the frames constituting the traffic image.
4 shows a flow of detecting an accident point after buffering a reference image at 10-second intervals and comparing the image with the image one minute ago.
Fig. 5 shows a vehicle route in a state where the accident vehicle is stuck on the road.
Fig. 6 shows the trajectories of subsequent vehicles trying to proceed by avoiding an accident vehicle.
7 is a reference diagram for a method of determining whether an accident occurs by the speed determining unit.
8 is a reference diagram for a method of determining whether an accident has occurred by referring to a movement trajectory of a vehicle at an accident occurrence point.
9 is a flowchart illustrating an example of an accident determination method based on a speed change amount and an individual vehicle trajectory.
10 and 11 illustrate experimental results of an image-based accident detection method using an accident simulation image.

In the image-based accident detection system and method according to the present invention, when a vehicle driving on a road causes an accident, the vehicle is first determined as an accident vehicle by referring to a reference image of the road.

By accurately determining whether an accident has occurred with reference to the traffic information caused by the congested vehicle, an accurate accident determination is performed as compared with a conventional accident detection method or system. At this time, in order to distinguish the road from the vehicle, a reference image of the road is required, and the reference image of the road may be obtained by applying a Gaussian mixture model to the traffic image.

1 is a system conceptual diagram of an image-based accident detection system according to an embodiment of the present invention.

Referring to FIG. 1, the image-based accident detection system includes a preliminary detection unit 110 and a definite detection unit 120 that acquire a traffic image through a plurality of cameras 10a to 10n installed on a road.

The preliminary detection unit 110 distinguishes the image of the road and the vehicle from the traffic image, sets the image of the road as the reference image, and refers to the state of the moving vehicle in the reference image to determine whether an accident has occurred in the vehicle. To judge. To this end, the preliminary detection unit 110 performs pixel values (eg, contrast) of frames satisfying a predetermined time interval (for example, 30 seconds to 5 minutes) among a plurality of frames constituting the traffic image acquired through the cameras 10a to 10n. , Digital values for at least one of color and luminance) are compared to obtain a reference image excluding an image of the vehicle. At this time, the preliminary detection unit 110 obtains a standard deviation for a frame that satisfies the Gaussian distribution under the assumption that the pixel values of the images constituting the two frames to be compared have a Gaussian distribution. When having a value within the deviation (eg 2.5), the frame can be estimated as the reference image.

Preferably, the preliminary detection unit 110 includes an object extraction unit 111 and a preliminary determination unit 112.

The object extracting unit 111 extracts a reference image which is a reference for the road from each frame constituting the traffic image, and determines an image of the vehicle moving on the road based on the extracted reference image.

The preliminary determination unit 112 stops the vehicle state, that is, stops for a predetermined time (for example, three minutes) or a significant number of lane changes in the vicinity of the vehicle, by using the divided image of the road and the vehicle in the traffic image. At that time, the vehicle is preliminarily determined to be an accident vehicle. The preliminary determination unit 112 determines the accident vehicle using the traffic images acquired through the cameras 10a to 10n and does not perform the accident determination using traffic information such as traffic volume and road occupancy. The secondary accident determination using the traffic information is performed by the determination detecting unit 120.

The method of distinguishing the road from the vehicle in the preliminary detection unit 111 will be described in detail with reference to the equation.

If the image detection method is largely divided into two methods, it can be divided into a tripwire method and a tracking method. The present invention is based on a tracking method for tracking individual vehicles. Since the tripwire method has a short sensing area, it is relatively accurate in detecting accidents by performing density estimation between loop detectors, while the tracking method has a long sensing area, which is a density estimation method. The advantage is that accident detection is possible directly in the detection area.

In order for the preliminary detection unit 111 to extract the moving object (vehicle) from the sequence of frames constituting the traffic image, it is a priority to distinguish the background (road) and the moving object (vehicle).

To this end, the preliminary detection unit 111 of the present invention uses a difference of the background image, that is, a difference occurring in the reference image of the road. When the pixel value of the image on the road changes rapidly, when the image on the road moves, or when a moving object (vehicle) located in the image is stationary, it is difficult to generate an accurate reference image of the road.

For example, when clouds are passing (sudden illumination change cases), when tree branches shake, or when there are wave changes (motion change cases), when the vehicle is parked, it is difficult to obtain a reference image of the road, and the reference image is accurate. If not, it becomes difficult to extract moving objects (vehicles) moving on the road.

As a solution to this problem, the preliminary detection unit 111 acquires an accurate reference image using a Gaussian Mixture Model (GMM).

In the Gaussian mixture model, in the sequence of each frame constituting the traffic image, the pixel value of each frame uses a Gaussian distribution having a constant mean value and standard deviation of the pixel values of the previous frames. The state of pixel values varies according to the sequence, but the overall mean or standard deviation is a constant.

Therefore, among the frames constituting the traffic image, a frame satisfying an average value (different depending on the road) and a standard deviation may be selected as a reference image for the road, and the rest may be regarded as an image including a moving object. have.

로 정의한다.The frame sequence of traffic video varies according to the pixel value of each frame.

.

라고 정의하고, 확률밀도함수가 아래의 수학식 1의 가우시안 분포를 따른다고 가정한다. In the present specification, the pixel value x may be defined for an image value such as contrast, color, and luminance. The pixel value x changes according to the image frame sequence, which is defined as a pixel process. This pixel process has a constant distribution, which is a probability density function

It is assumed that the probability density function follows the Gaussian distribution of Equation 1 below.

는 각각 k 상태 분포에서 평균과 분산을 의미한다.here

Are mean and variance in k state distribution, respectively.

고 했을 때,

이 되며, 영상에 적용하여 픽셀 값의 분포 예는 도 2 와 같다.At this time, the prior probability of the state

When you say

2 is an example of distribution of pixel values applied to an image.

When k states in FIG. 2 are independent of each other, X becomes a sum-of-Gaussian mixture model of Equation 2.

이고, 모든 파라미터

는 픽셀값 x로부터 추정(estimation)이 가능하다.here,

, All parameters

Can be estimated from the pixel value x.

If the pixel process is stable (stationary process), the maximum likelihood function can be found using the Expectation Maximum algorithm. Stauffer and Grimson estimated parameters by K-means approximation.

First, the posterior probability for estimating the current state is expressed by Equation 3.

here,

이고,

ego,

가 최대가 되는 k(이를 Stauffer와 Grimson는 match라 하였음)를 MAP(maximum a posterior)라고 하였고, 이것을 현재상태로 결정한다.

The maximum k is called Stauffer and Grimson's match, and is called MAP (maximum a posterior).

은 현재까지의 k 가우시안 분포에서 match가 될 때까지(일반적으로 match는 가우시안분포의 표준편차가 2.5이내일 경우로 정의한다) 조정을 해주는 데, 우선

는 수학식 4의 방법으로 학습시킨다.New video

Is adjusted until k is matched in the current Gaussian distribution (usually match is defined as the standard deviation of the Gaussian distribution is within 2.5).

Learn by the method of equation (4).

는 학습률(learning rate),

는 match일 경우 1, 아닐 경우 0이 된다.here,

Is the learning rate,

Is 1 if match, 0 otherwise.

,

는 수학식 5와 수학식 6에 의하여 갱신된다.Also when a new video is input

,

Is updated by equations (5) and (6).

here,

, 또는

을 계산하여 K별 랭킹(ranking)을 계산한 다음 큰 값을 배경영상으로 간주하면 된다. 즉,

가 큰 값이나

작은 값을 가질 때 배경 영상(기준 영상)일 확률이 높다고 보고, 수학식 7과 같이

의 누적으로 전체 사전확률 T와 비교하여 클 때 이를 배경(기준 영상)으로 판단한다. A method of finally estimating a background image as a reference image using the above method is as follows. In the distribution of Equation 2, a moving object (vehicle) and a background (road) exist together, and in order to distinguish a background and a foreground,

, or

Calculate the K-by-K ranking and then consider the large value as the background image. In other words,

Is a large value or

When it has a small value, it is considered that the background image (reference image) is likely to be high.

When it is larger than the total prior probability T due to the accumulation of, it is determined as a background (reference image).

이다.Where T is the overall prior probability

to be.

2 is a reference diagram for a process of estimating a reference image using a Gaussian mixture model. Referring to FIG. 2, when a preliminary judgment on an accident is performed using an image (reference image) of a background (road), an accident vehicle stands at a point where an accident occurs in both a minor accident or a large accident. Using the mixed model, the accident vehicle is also included in the reference image of the background and stored.

In the present invention, the preliminary detection unit 111 may make a sequence of only the reference image (X, Y) at intervals of n seconds and store it in the buffer, and then obtain a difference between predetermined time periods and detect the accident vehicle in a region exceeding a threshold value. .

The reason for a certain time difference is that when an accident occurs in a vehicle, the accident vehicle becomes a background, and is intended to block an accident detection error caused by being included in a reference image. This will be described with reference to FIG. 3 together.

3 illustrates that frames B0 to Bn of reference frames are arranged among the frames constituting the traffic image. Frames for these reference images are buffered, and frames satisfying a predetermined time interval, such as BO and B6, are selected and used to determine whether an accident has occurred in the vehicle. When the time interval is n = 100 according to Equation 8 below, assuming that the interval with respect to the reference image is 10 seconds, it is possible to detect whether or not an accident occurs in the vehicle around the time interval of approximately 1 minute.

here,

Here, if n = 100 seconds, the background image is buffered in units of 10 seconds.

으로부터 사고차량이 포함된 기준 영상

까지 버퍼에 저장이 되고,

을 계산하면 사고차량의 전경이 추출되어 사고 지점 및 사고차량을 검지할 수 있다.After a certain period of time after the accident, all the stopped vehicles are included in the reference video, including the accident vehicle except for the moving vehicle.

Video with accident vehicle from

Is stored in the buffer until

By calculating, the foreground of the accident vehicle is extracted and the accident point and the accident vehicle can be detected.

를 기준으로 10초 후 , 20초 후

로 기준 영상을 버퍼링한 시퀀스는

로 표시할 수 있으며, 여기서

연산을 하면, 사고로 인한 정차 차량을 추출할 수 있다. 4 shows a flow of detecting an accident point after buffering a reference image at 10 second intervals and comparing the image with the image one minute ago. In other words,

After 10 seconds, after 20 seconds

The sequence that buffered the reference image with

, Where

The operation can extract the vehicle stopped by the accident.

However, even in the case of extreme congestion, since there is no movement of the vehicle, it is determined as in the case of an accident, which is determined by the determination detecting unit 120.

When the determination detector 120 determines that an accident has occurred in the preliminary detection unit 110, the determination detector 120 finally determines whether the accident has occurred by referring to traffic information of the area where the accident is detected and the area before and after the area where the accident is detected. The final detection unit 120 determines whether an accident occurs by determining whether there are any characteristics when an accident occurs by combining traffic flow change, road occupancy change, lane change amount, and the like in an area where the accident is expected to occur. Preferably, the determination detecting unit 120 includes a speed determining unit 121, a occupancy determining unit 122, and a lane determining unit 123.

When the accident occurs on the road, the speed determiner 121 has a characteristic that the vehicle speed increases in front of the accident occurrence point and the vehicle speed decreases in the rear, based on the direction in which the vehicle proceeds. This is because the front of the accident occurrence point is not affected by the accident, and the rear is directly affected by the accident.

The speed determining unit 121 determines whether an accident has occurred by comparing the average speeds of the vehicles with respect to the front and the rear of the vehicle at the point where the occurrence of the accident is predicted by the preliminary detection unit 110 based on the traveling direction of the vehicle.

The occupancy determining unit 122 determines a road occupancy rate of the vehicle for the front and the rear of the vehicle based on the traveling direction of the vehicle at the point where the occurrence of the accident is predicted by the preliminary detection unit 110. In general, when an accident occurs on the road, the rear of the accident spot occupies the road due to the vehicle being accumulated, but in the front, the road occupancy decreases because there is no subsequent vehicle. The occupancy determining unit 122 determines whether an accident occurs by comparing the occupancy rates of the roads in front of and behind the point where the accident occurrence is predicted by the preliminary detection unit 110.

The lane determination unit 123 determines whether an accident has occurred based on the increase in the number of lane changes of the vehicle at the point where the accident occurrence is predicted by the preliminary detection unit 110. On the road consisting of multiple lanes, there is an attribute of driving to avoid a vehicle having an accident, so if an accident occurs, a large number of lane changes occur in an accident occurrence area, and a lane change that avoids a specific lane occurs.

The lane determination unit 123 tracks lane changes occurring in the front and rear directions based on the point where the accident occurrence is predicted by the preliminary detection unit 110, and finally compares the lane change amount at the point where the accident occurrence is predicted. The occurrence can be determined. The way in which the lane change occurs in the vicinity of the accident vehicle will be described with reference to FIGS. 5 and 6 together.

Fig. 5 shows a vehicle route in a state where the accident vehicle is stuck on the road. Due to the congestion of the vehicle, road signs no longer occur and stop.

Fig. 6 shows the trajectories of subsequent vehicles trying to proceed by avoiding an accident vehicle. As shown in FIG. 6, subsequent vehicles following the accident vehicle perform a lane change to use a lane different from the lane in which the accident vehicle travels, resulting in a large number of routes of the lane change.

On the other hand, the speed determination unit 121, the occupancy determination unit 12, and the lane determination unit 123 outputs the determination result that all the accident occurred on the road when an accident occurred. However, when one of the three does not output the same result, the determination detector 120 when the two or more results of the speed determination unit 121, occupancy determination unit 12, and lane determination unit 123 is the same, According to the result or by assigning weights to each of the speed determining unit 121, the occupancy determining unit 12, and the lane determining unit 123 differently, the sum of the weights is compared with a predetermined reference value to determine whether an accident occurs. May be determined.

On the other hand, the AID system needs to have a high detection rate (Detection probability) and at the same time a false alarm probability must be low. In other words, if it is determined that the accident is not an accident and frequently warns (wrong call), the reliability of the AID system is lowered and the efficiency of accident detection is reduced.

Reijmers suggested the following as a criterion for judging an accident.

(1) Measurement of the changes in detector occupancy

(2) analysis of time headway

(3) analysis of the changes in velocity

(4) measuring the average travel time

When the image-based accident detection system according to the present invention becomes a primary preliminary judgment on whether an accident occurs by using a reference image calculated by applying a Gaussian mixture model, Reijmers suggests whether the accident is due to congestion or accident. Further determination may be performed by referring to traffic information of a traffic volume, a speed change amount, and an occupancy change amount. In addition, since the image-based accident detection system and method of the present invention are processed based on individual vehicle tracking in the detection area, the accident detection can be performed in consideration of the case where the movement trajectory of the individual vehicle is normal or abnormal.

7 is a reference view for a method of determining whether an accident has occurred by the speed determining unit 122.

Referring to FIG. 7, a speed change amount of a subsequent vehicle when an accident occurs is shown in FIG. 4. For example, a vehicle running at a speed of 80km / h decelerates to less than 10km / h in 3 to 4 seconds, and if the next lane is empty before the accident, it changes immediately, otherwise it is near the accident. You will stop at. In particular, when tracking individual vehicles, a certain amount of speed decrease is inevitable at any particular point.

In order to measure the speed change of the individual vehicle, the speed differential value of the individual vehicle according to Equation 9 is obtained, and this value always has a negative number as shown in Equation 9 in the case of deceleration.

값이 임계값을 초과하면 사고로 인한 속도 감속이라 판단하여 사고로 간주할 수 있다.The speed change amount is also calculated by the equation (10)

If the value exceeds the threshold value, it can be regarded as an accident by judging that it is a speed reduction due to an accident.

8 is a reference diagram for a method of determining whether an accident has occurred by referring to a movement trajectory of a vehicle at an accident occurrence point.

The pattern of trajectory flow of a vehicle after an accident can be divided into two types.

First, when a minor accident occurs, the phenomena that occur at this time are changed to a car to avoid the accident spot by the subsequent vehicles behind the accident spot. In normal traffic conditions on a continuous road, the number of lane changes is not high, but if an accident occurs, the number of times is increased.

On the other hand, as shown in Fig. 78 is a reference diagram showing the change amount of the trajectory and the speed in the normal communication, it can be seen that each vehicle travels the road according to the driving direction while substantially maintaining its own lane.

Second, the tramway is blocked due to a major accident, which prevents the trajectory from appearing in the detection area.

The lane determination unit 123 of the present invention may perform an accident detection by measuring the number of lane changes of an individual vehicle.

The vehicle's trajectory uses a region based tracking algorithm. It separates individual vehicles from the road and stores them in the DB in succession for each ID that gives the vehicle the center coordinates of the rectangle from which the vehicle area is extracted.

In order to calculate the trajectory of each vehicle, the vehicle position (coordinate value on the screen) of each frame constituting the traffic image is converted into actual travel distance information, and the trajectory of each vehicle is calculated. The pattern for the trajectory is divided, and when an accident occurs in one lane, the number of lane changes is increased so that subsequent vehicles may leave the accident lane.

Referring to FIG. 7 again, FIG. 7 illustrates a vehicle trajectory and a lane change trajectory on a road in normal communication. In FIG. 7, the case where the trajectory of the vehicle deviates by more than the standard deviation from the mean of the center point of the lane is regarded as a lane change, and the number of times is counted according to Equation (11).

는 차량 ID k 차량이 차로의 x좌표, y좌표는 차량 진행 방향이므로 비교에서 제외하고, 차로변경은 x 좌표값으로 비교하여, 제외하고, 차로변경은 x 좌표값으로 비교하여, 아래의 수학식 12에 따라 차로변경의 누적 횟수를 산출한다.
Where j is the car,

Since the vehicle ID k vehicle is the x coordinate of the lane and the y coordinate is the vehicle traveling direction, except for the comparison, the lane change is compared with the x coordinate value, except that the lane change is compared with the x coordinate value. The cumulative number of lane changes is calculated according to 12.

) 이상이면 사고로 인한 차로변경이라 판단하여 사고로 간주할 수 있다.Next, the cumulative number of lane changes calculated according to Equation 12 is a reference amount (

), It can be regarded as an accident by deciding that the lane is changed by the accident.

Therefore, the accident determination method based on the speed change amount and the trajectory of the individual vehicle follows the flowchart of FIG. 9.

)을 초과하는가를 판단한다(S204). 만일, 속도 변화량이 기준량(

)을 초과하지 않을 경우, 정상 소통으로 간주되고(S205), 반대로, 속도 변화량이 기준량(

)을 초과하는 경우, 이상 값으로 판단 후(S206), 수학식 11에 따라 차량의 차선 변경이 발생하는가를 계산하고(S207), 차선 변경일 경우 수학식 12에 따라 차선 변경의 누적 횟수가 기준량(

)를 초과하는 가를 판단하며(S208), 판단결과 초과되는 경우 사고로 판정한다(S209).
9, the speed determining unit 121 of the determination detecting unit 120 calculates a speed of each individual vehicle at a point where an accident is expected to occur (S201), and calculates a speed differential value for the individual vehicle (S202). ), The number of speed changes for the individual vehicle is counted (S203), and the speed change amount of the vehicle is a reference amount (

It is determined whether to exceed (S204). If the speed change amount

If not exceeding, it is regarded as normal communication (S205), on the contrary, the speed change amount is a reference amount (

Is exceeded, and after determining as an abnormal value (S206), it is calculated whether a lane change of the vehicle occurs according to Equation 11 (S207), and if the lane change, the cumulative number of lane changes according to Equation 12 is the reference amount. (

) Is determined to be exceeded (S208), and if it is determined that the result is exceeded, it is determined as an accident (S209).

On the other hand, after the accident occurs, as shown in FIG. 7, the traffic volume and the occupancy of the rear is high, and the traffic volume and the occupancy of the front fall.

After the second accident detection routine is performed in the definite detection unit 120, the detection area is divided into n blocks, and the occupancy rate of each block is compared.

The traffic information for each block is determined by the correlation between the within block and the between block, and in case of an accident, the occupancy of the within block is high or low, while the occupancy of the between block is large.

(1) Within block occupancy divides the detection area into blocks, and then calculates the occupancy of the within block by the following equation (13).

는 검지영역 내 블록별 면적이고,

는 블록 내 차량 ID k대 면적의 합으로 모든 블록의 점유율이 90%이상인 경우는 검지영역 전체가 점유율이 높기 때문에 정체라고 판단하고, 어느 특정 불록 이전부터 점유율이 높으면 사고로 인한 정체로 판단할 수 있다.here,

Is the area of each block in the detection area,

Is the sum of the area of the car IDs in the block, the occupancy rate of all blocks is over 90%, so the entire detection area is congested because of the high occupancy rate. have.

In addition, between block occupancy deviation is a high probability of accident if the difference in occupancy between blocks is high, while if the difference in occupancy between blocks is small, the probability of congestion is high. Can be calculated.

The Between block calculates the difference from the previous block. If the deviation of the block j-1 in the block j-1 is less than 10%, the block is congested.

Applicant's test results of the image-based accident detection system and method described above are as follows.

Actual accident images or simulation data are needed to verify the thinking algorithm. However, this study is an image-based accident detection algorithm, which cannot obtain data through a general traffic simulation program, and requires image data such as CCTV.

Therefore, this study used two methods for algorithm verification. This method utilizes the virtual accident experiment simulation method and the actual CCTV images collected on the real continuous road for more objectivity.

The virtual experiment was collected from 200,000 vehicles from 7.7km on the Chubu Inland Expressway from 10:30 am to 4 pm on February 20, 2009. , Types of stop vehicles due to accidents in the primary or secondary lanes (longitudinal, single collision, multiple collisions, etc.), stop vehicles occurring in the first and second lanes due to accident collision (lateral direction) Twelve scenarios were constructed in consideration of the complex situation such as the occurrence of stop vehicles (or road stops) in the first and second lanes.

In addition, the actual accident images used for verification through the actual accident images were utilized three images collected by Janghang IC CCTV on Route 77 (freeway).

This is a video collected from a 15m high CCTV. First, the type of accident is a video of a night collision that occurred on January 17, 2009 at 18:48. Occurred at the end. Second, the weekly crash video that occurred at 09:50 am on January 29, 2010 is also a collision with a vehicle coming in the fourth lane while changing to a ramp to enter the ramp from the third lane. Third, February 1, 2010 At 09:40 am, the car stopped for about four minutes. This was not an actual accident, but was considered to be an accident because it was an unauthorized stop on the road.

Simulation results on the experimental road near the central inland highway were detected in all 12 simulation accidents, indicating that the detection rate was 12/12 = 100.00%, the false detection rate = 00.00%, and the undetection rate = 00.00%.

In the simulation tests for each accident type on the experimental roads near the central inland highway, 12 cases were detected in all cases with detection time of 25 to 35 seconds, and the results of the accident detection method using the accident simulation image are shown in the following figure. 10 and FIG.

In addition, in case of automatic detection of real-time video accidents of test bed near Changhang IC, all three cases were detected, and there were no false or undetected.

In the case of the night accident image, the incident detection alarm was displayed 34 seconds after the actual accident (stopped immediately after the collision), and the time of 33 seconds and 27 seconds was consumed in the daytime accident. In particular, the February 1 image was verified by the image collected in the foggy weather, the accident detection results in the three cases are shown in Table 1.

The detection result for this is shown in FIG. 10.

Through this, 15 accidents were detected in all 15 experiments, although the number of samples was not enough, resulting in a high detection rate of 100.00%, a false detection of 00.00%, and an undetected 00.00%. Since all the actual accidents were possible, the test results under continuous road conditions showed very high reliability.

The present invention preliminary detection of the first accident using a method of storing a background image with a Gaussian Mixture Model for a certain period of time under the premise that there is a stationary vehicle due to an accident, and that there is a vehicle moving in a slow way. We propose an algorithm to determine the accident by comparing the speed variation and trajectory variation of individual vehicles using the stored traffic information.

In addition, since the automatic accident detection algorithm of the present invention is based on individual vehicle tracking in the detection area, a method that is differentiated from the method based on the speed and occupancy difference of the loop detector between points is proposed.

In order to verify the accuracy of this AID system, it is necessary to acquire the accident images of various cases occurring in the actual field and apply the algorithms and measure the result. Was performed. First, 12 simulations by type of accidents assuming accidents on the Jungbu Expressway Expressway test road and second, 3 real accident images collected from Changhang IC CCTV installed for the field test of this study.

In particular, the three actual accident images were collected according to various time zones and weather conditions such as night, daytime and fog conditions, which greatly helped to prove the objectivity of the AID algorithm.

As a result of the test, accident detection was detected within 30 seconds in the simulation image and 27 to 34 seconds in the actual accident image.

The characteristic of this algorithm is Gaussian Mixture Model, which stores background images continuously and detects all objects stopped due to accidents in the detection area. The accident can be detected in the situation, and in addition to the accident vehicle, falling obstacles falling on the road can be detected by the same algorithm.

For the verification of the present invention, accident images were acquired through three actual accident images and a total of 12 experiments, but it was difficult to reflect various accident situations, and as a future task, more types of accident images were collected on the actual road and verified and verified. It is necessary to improve the completeness with the accident detection system by using it in the evaluation, and also to study how to reduce the number of wrong calls for the wrong call, which is wrongly judged as an accident and gives a warning.

10a to 10n: camera 110: preliminary detection unit
111: object extraction unit 112: preliminary determination unit
120: confirmation detection unit 121: speed determination unit
122: share determination unit 123: car determination unit

Claims (11)

A preliminary detection unit which acquires an image of a road as a reference image by applying a Gaussian mixture model to the traffic image obtained by the camera, and detects a vehicle in a stagnant state from the reference image; And
And a determination detector for secondly determining whether an accident with respect to the vehicle has occurred by referring to traffic information around the vehicle.
The preliminary detection unit,
An object extractor configured to distinguish a vehicle from a road by referring to a change in a pixel value of the reference image; And a preliminary determination unit determining whether an accident has occurred in the vehicle by referring to the change in the pixel value of the vehicle.
The object extraction unit,
Obtaining a reference image of the road by applying a Gaussian mixture model to pixel values of each frame constituting the traffic image,
The pixel value is
A value for at least one of contrast, color, and luminance of each frame constituting the traffic image;
The confirmation detection unit,
A speed judging unit which determines whether an accident has occurred by comparing the average speeds of the vehicles with respect to the vehicle forward and rearward based on the traveling direction of the vehicle at the point where the accident occurrence is predicted by the preliminary detection unit;
A share determination unit for determining whether an accident has occurred by comparing road shares of vehicles for the front and the rear of the vehicle based on the direction in which the vehicle is predicted by the preliminary detection unit; And
And a lane determination unit determining whether an accident has occurred based on how much the number of lane changes of the vehicle has increased at the point where the occurrence of the accident is predicted by the preliminary detection unit. delete delete delete The method of claim 1,
The preliminary detection unit,
And a reference image of the road with reference to a Gaussian distribution of pixel values for each frame of the road image. The method of claim 1,
The preliminary detection unit,
The image-based accident detection system of claim 1, wherein the frame of the road image comprises comparing the selected first frame and the second frame according to a predetermined time interval to determine whether an accident has occurred in the vehicle. The method of claim 6,
The first frame and the second frame,
Image-based accident detection system, characterized in that the frame is selected with a time interval of 30 seconds to 5 minutes. delete delete delete delete

Images