KR20100097907A - A scheme of extracting forward vehicle area using the acquired land and road area information - Google Patents

Abstract

PURPOSE: The detection rate of the vehicles is enhanced against the road image having the complex background. The efficiency red phosphorus because of limiting to the extracted vehicles domain, the vehicles is detected. CONSTITUTION: The fetch strategy of the front vehicles domain using lane and road region area information. The exact lane is detected from the complex road image. The step of considering connectivity between pixel it uses the edge chain code. The lane according to the direction of the running vehicle is detected(12).

Description

A Scheme of Extracting Forward Vehicle Area Using the Acquired Land and Road Area Information}

Communication

Researches for intelligent cars are actively conducted at home and abroad for the purpose of ensuring safe driving as well as driverless cars. Recently, researches for improving automobile safety and accident prevention have been actively conducted in automobile manufacturers, government and university research institutes. Existing methods have focused on the prevention of traffic accidents that alert drivers to the situation when they are expected to collide with other cars while driving or to leave the lane. In addition, the method distinguishes obstacles including vehicles and people on the road and provides distance information to the target.

Recently, many image processing techniques have been proposed to separate lanes, vehicles, and obstacles from road images obtained in complex environments such as downtown roads. In other words, techniques related to vehicle recognition, tracking, and collision avoidance have been proposed. It uses a combination of sensors such as laser, radar, ultrasound, and infrared to detect nearby terrain information and obstacles. In addition, methods for distinguishing vehicles and obstacles from road images acquired using CCD cameras have been studied. Since the techniques using the image are greatly influenced by day and night time, lighting environment and noise, it is difficult to detect obstacles. Therefore, the obstacle detection technique robust to the external lighting environment is required.

As the development of various images and sensor-based driverless vehicles is being actively studied, the DARPA Grand Challenge competition, which is a driverless automobile competition, is being held successfully. The goal of the competition, which began in 2004, is to drive the car on its own on a designated course in the desert. In 2007, the urban environment of the model (DARPA Urban Challenge) was constructed to drive for each situation, and 53 teams participated. Most of the autonomous cars participating in the competition recognize the environment by using a combination of sensors from various fields such as radar, laser, GPS and CCD camera. CCD cameras are used in the field of detecting, tracking, and avoiding vehicles in front of the vehicle.

Vehicle detection is essential information for safe driving in a vehicle driving assistance system. Related research has received much attention and various detection algorithms have been proposed. The most prominent feature that can be detected in the vehicle is the vertical and horizontal components in the edge image. Sun has proposed a technique that uses Gabor filters to detect features of vertical and horizontal components, and uses a classifier such as SVM to verify the existence of a real car in the road area. In order to increase the efficiency of the vehicle detector in the complicated road image, the vehicle area in which the vehicles exist in the road area should be extracted.

The model-based vehicle region extraction technique, which assumes the road shape as a general combination of straight lines and curves, is robust against noise and loss of lane information. However, because it focuses on a particular type of road, it is not suitable for vehicle area detection in any road type. In order to solve the above problem, a method of extracting a vehicle area from lane detection using a B-Snake technique, which is applicable to both straight and curved roads, has been proposed.

Kang also proposed a lane detection technique using feature information to distinguish color and width information of roads and lanes. Lane candidates were selected through template matching, and lanes were selected through the lane model for the highest value among the candidates.

Many studies have proposed a method for extracting a vehicle region from lane detection using Hough Transform, which is a technique for detecting a linear component in an image. Hough transform is known as an efficient method for detecting pixels arranged in a straight line in an image. Various modified Hough transform methods using this feature have been used for the research of finding vehicle area. Hua-jun detects lanes through the study of the vanishing point estimation using the transform into Hough space. Extraction of vehicle area using general Hough transform is well applied when background information is simple in road image. As road information with complex and diverse backgrounds such as downtown roads has complex edge information, accurate lane detection is difficult and it is difficult to extract a vehicle area.

Extraction of vehicle area using general Hough transform is well applied when background information is simple in road image. As road information with complex and diverse backgrounds such as downtown roads has complex edge information, accurate lane detection is difficult and it is difficult to extract a vehicle area. In this complex road image, we want to improve the efficiency of vehicle detection.

In order to increase the efficiency of vehicle detection in complex road images, it is a technique of extracting a vehicle area where a vehicle exists by finding road areas from lane detection using chain codes.

In order to achieve the above object, the present invention comprises the steps of (a) considering the connectivity between the pixels using the edge chain code to detect the correct lane in a complex road image; (b) detecting lanes coinciding with the direction of the traveling vehicle and detecting adjacent lanes by finding the width of the laneway and the vanishing point of the lane from the center lane; (c) extracting a vehicle region in which the vehicle exists using edge information of the vehicle in the road region including the driving lane and the adjacent lane.

The vehicle area extraction method using lane and road area information is highly efficient for detecting a vehicle because it can increase the detection rate of the vehicle in the road image having a complicated background and limit the detection to the extracted vehicle area.

According to an embodiment of the present invention, a method of extracting a front vehicle area using lane and road area information includes: (a) considering connectivity between pixels using an edge chain code to detect an accurate lane in a complex road image; (b) detecting lanes coinciding with the direction of the traveling vehicle and detecting adjacent lanes by finding the width of the laneway and the vanishing point of the lane from the center lane; (c) extracting a vehicle region in which the vehicle exists using edge information of the vehicle in the road region including the driving lane and the adjacent lane.

FIG. 1 is a flowchart illustrating a lane detector method considering directionalities of connected pixels to which a chaincode is applied from an edge road image of step (a). An 8-way chaincode is used to consider the connectivity between pixels using edge chaincode to detect accurate lanes in a complicated road image. Chaincodes are a useful way to determine the shape of an object using the connectivity of its outline components. Since the chaincode contains information on the direction of the connected pixels, it is easy to detect features connected in a certain direction such as a straight line. In addition, it has the advantage that it is possible to remove the spindle direction and other components based on the starting point. Therefore, it is possible to separate straight line components in a specific direction from complex edge components existing in an image.

Considering the set of pixels continuously connected in a certain direction in a straight line in the image, the lane in the road image can be seen as a set of straight lines in a direction toward the vanishing point as shown in FIG. In order to detect the road lane, the linear components having a direction toward the vanishing point are separated from the edge road image.

The lane components appearing on the road image are generally represented by a straight line towards the vanishing point. The vanishing point of the road image captured from the camera mounted at the horizontal center of the vehicle is located at the horizontal center and the road lane faces the vanishing point. In order to easily find a lane in the road image, the road image is divided into left and right directions based on the horizontal center of the image as shown in FIG. 3. The straight lines obtained from the set of connection pixels toward the center in each divided region are candidates for the road lane. FIG. 9 shows the result of the vanishing point using the lane components and the respective center direction components of the road for the actual road image. FIG. 4 (13) shows a detection result for the linear component having the chain code value of the central direction component in both regions of the edge image. FIG. 4 (14) shows that the road lanes remaining after removing the noise component among the lane candidates obtained in FIG. 4 (13) face the vanishing point.

The acquisition of lane information in the road image provides not only information for determining the driving direction of the traveling vehicle but also a warning signal for avoiding collision by quickly detecting vehicles while the vehicle is driving. In order to shorten the time for detecting vehicles adjacent to the driving road without searching the entire road image, the road area including the center roadway and the adjacent roadway is separated, and the vehicle area including the front vehicles is found in the road area. 5 shows a flowchart of a technique for finding a vehicle area for detection of a forward vehicle.

In the present invention, the road is defined as a region including left and right adjacent roadways as well as a center roadway of a traveling vehicle in a road image in which a road has multiple lanes. The center roadway of the driving vehicle is the area connecting the vanishing point obtained in the previous chapter with the lane adjacent to the horizontal center of the road image. The left and right roadways are obtained by correcting the width of the center roadway and extending left and right as shown in FIG. 6 (11). As shown in FIG. 6 (11), the road area defined in the present invention includes a central roadway C and left and right roadways L and R calculated from the area of the center roadway. The information of the detected lanes and the vanishing point provide many clues to find the road area of the driving vehicle. The location of the vanishing point obtained by extending the road lanes and both lane information adjacent to the horizontal center of the image determine the center lane of the traveling vehicle.

In a situation where the vehicle is traveling on the central roadway, the front vehicle must be detected in real time in the road area to avoid collision with the front vehicle. Real-time front vehicle detection is possible by limiting the search range of the vehicle detector to the vehicle area where the vehicle is expected to be present on the road area. The vehicle area is defined as the area occupied by the front vehicle closest to the traveling vehicle in the roadway of the road area. The width of the vehicle area is the width of the roadway calculated from the lane information, and the height is set to a size proportional to the width of the vehicle area because the height varies depending on the type of the vehicle or the driving vehicle and the distance.

In order to find the vehicle area, horizontal edge projection is applied to the area of each roadway constituting the road area. Using the vertical and horizontal components that are most prominent in the vehicle, the portion of the edge component appearing in each region with a strong accumulation of the projection change component is determined as the vehicle region in which the vehicle is placed. Equation 1 shows the projection relationship of the edge components in the horizontal and vertical directions.

It is difficult to determine the height of the vehicle because the vehicle travels in contact with the road and the ratio of the width and height of the vehicle area is determined by the type and distance of the vehicle ahead. In this paper, the height of the vehicle area is set to 2/3 of the width of the vehicle area so as not to miss the detection of the vehicle ahead.

The predetermined vehicle area represents an approximate area including the front vehicle lying on the roadway. In this paper, a vehicle detector using Bayesian Discriminating Features (BDF) is used to verify the effectiveness of the extracted vehicle area for the detection of the vehicle ahead.

The vehicle detector using the BDF technique consists of a learning process and a test process. In the learning process, the feature matrix is first constructed by extracting predefined features for all input images (∈ R ^mxn ) of the vehicle and non-vehicle training sets. In this paper, the vehicle detector using BDF is normalized gray scale image (∈ R ^mn ), one-dimensional Harr Wavelet representation (∈ R ^{(m-1) n} , R ^{m (n-1)} ), The features of horizontal / vertical projection (∈ R ^m , R ⁿ ) were used. Equation 2 shows a representation of a feature vector composed by normalizing each feature.

는 각각 특징들을 정규화한 그레이 영상, 수평/수직 1-D Harr Wavelet 표현, 수평/수직 프로젝션 특징이다. 입력 영상에 대한 특징벡터 는 3mn의 차원을 갖는다. Where t is the transposition operator

Are gray image normalized features, horizontal / vertical 1-D Harr Wavelet representation, and horizontal / vertical projection features. The feature vector for the input image has a dimension of 3mn.

Principal component analysis (PCA) is performed by constructing a covariance matrix from feature vectors. Each input image is projected into the eigenspace to obtain a weight vector, and finally, each of the vehicle and non-vehicle classes is modeled with a normal distribution. Equation 3 is a trailing probability distribution of the vehicle class and the non-vehicle class based on the Bayesian theory.

According to the Bayesian Decision Rule, when the trailing probability that the input feature vector belongs to the vehicle class is greater than the trailing probability that belongs to the non-vehicle class, the vehicle is judged to be non-vehicle when the reverse characteristic is opposite (Equation 4).

In the test process, when the test image is input, the feature extraction step is performed in the same manner as in the training process. The extracted feature vector is projected into the eigenspace through the transformation matrix obtained through principal component analysis in the training process to obtain the weight vector. Finally, stochastic statistical values for each class are calculated and Bayesian Decision Rule is used to determine whether the input image is car or non-car.

7 shows a process of obtaining a road area and a vehicle from a driving image. 7 (12) shows the center roadway of the driving vehicle, and FIG. 7 (11, 13) shows the road area including the center roadway and the left and right roadway for the road image of the driving vehicle. 7 (14) shows the result of detecting the vehicle in the road area.

1 is a flowchart of a lane detection technique using a chaincode of the present invention.

2 is a diagram showing a linear component in the vanishing point direction of the present invention.

FIG. 3 is a diagram of dividing a main axis direction component bisecting based on the center of an image of the present invention and pointing toward the center in each region.

4 is an image showing a process of obtaining a road lane and a vanishing point from a road image of the present invention.

5 shows a flow chart of extraction of a vehicle region for front vehicle detection of the present invention.

6 is an image of an embodiment and classification of a road area in a road image.

7 is an image of a vehicle region detection result of a front vehicle in a road region.

Claims (3)

A technique for detecting a lane using a chain code from a camera input image for the purpose of detecting a road area. A technique for detecting left and right lanes by measuring the width of the center lane in order to find left and right lanes of a vehicle from a road area. A method using Bayesian Discriminating Features (BDF) to verify the effectiveness of the vehicle area from the detection of forward vehicles located in the center and left and right lanes of the vehicle.

Images