KR101478072B1 - Method for Detecting Vehicle - Google Patents

Abstract

Disclosed is a vehicle detection method for detecting a vehicle traveling in a forward or backward direction from an image photographed with a momocular camera mounted in a running ego vehicle. A vehicle detection method includes: receiving a photographed image from a camera; Setting an area of interest for vehicle detection in the input image; Generating a bottom candidate group of the vehicle using the ground surface histogram calculated within the set region of interest; Generating a side face candidate group of the vehicle using the generated bottom face candidate group of the vehicle; Generating a vehicle candidate group by combining the base candidate group and the side face candidate group of the generated vehicle; By applying a weak classifier and a strong classifier to the generated candidate vehicle group, it is possible to improve the accuracy of the vehicle detection performance by detecting whether the vehicle is an actual vehicle, and to simplify the system by reducing the number of sensors in comparison with the existing front collision warning system .

Description

[0001] The present invention relates to a method for detecting a vehicle,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vehicle detection method for detecting a vehicle running in front or rear from an image photographed by a momocular camera mounted in a running ego vehicle.

There is a frequent occurrence of a collision between a vehicle that is running in front of the vehicle due to a driver's inattention or a clock failure during operation of the vehicle. That is, a traffic accident caused by a lane departure due to long-distance driving, a bad weather in the case of rainy weather or night driving, a lack of concentration of the driver, and a collision with the driving vehicle are constantly occurring. In order to prevent such a collision, a driver assistance system is provided which can provide information such as collision warning, collision avoidance, and cruise control by detecting vehicle position and speed through vehicle detection.

Generally, a method of detecting a vehicle ahead is a method using a radar or a camera, and a method using an ultrasonic sensor is a method of detecting an obstacle or a vehicle in a dead zone or a rear parking.

Among them, the radar method is a method of detecting vehicle distance information in various directions by attaching a plurality of radar devices to the front of the vehicle, and it is applied to a system that requires active control since the accurate position of the vehicle is known. However, the method using the radar can be used regardless of the distance, but it is very expensive to commercialize it in a general vehicle, and there is a frequency interference problem. In addition, the existing Forward Collision Warning (FWC) system uses a sensor fusion structure in which the camera again judges an object sensed by the radar, thereby increasing the complexity of the system.

The method using the ultrasonic sensor is a method of detecting the vehicle distance information by emitting the ultrasonic wave backward from the ultrasonic sensor attached to the vehicle and detecting the rear obstacle or the reflected wave reflected from the vehicle. I have. Therefore, when this method is used, there is a problem that detection of the vehicle is performed efficiently only in the near region within 10 m according to the performance of the ultrasonic sensor.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and it is an object of the present invention to provide a vehicle detection apparatus and a vehicle detection method, which can detect a forward or a rearward vehicle using an efficient camera, Method.

The present invention also provides a vehicle detection method that can simplify the system by reducing the number of sensors in comparison with the existing frontal collision warning system by processing the recognition and judgment of the obstacle or the front / There is another purpose to provide.

According to an aspect of the present invention, there is provided a vehicle detection method comprising: receiving a photographed image from a camera; Setting an area of interest for vehicle detection in the input image; Generating a bottom candidate group of the vehicle using the ground surface histogram calculated within the set region of interest; Generating a side face candidate group of the vehicle using the generated bottom face candidate group of the vehicle; Generating a vehicle candidate group by combining the base candidate group and the side face candidate group of the generated vehicle; A weak classifier and a strong classifier are applied to the generated candidate vehicle group to detect whether the vehicle is actually a vehicle.

Also, the setting of the ROI sets the ROI by projecting the points at the maximum distance for recognizing the vehicle in the 3D real coordinate system onto the input image.

Also, the generation of the bottom candidate group of the vehicle considers a specific region within the set region of interest as a ground; Calculating a ground level histogram representing the brightness of the ground; Calculating a threshold value of a shadow region using an average value (μ) and a variance value (?) Of the calculated ground surface histogram; Obtaining a binarized image by comparing brightness values of all points within the region of interest with the threshold value; Extracting a black and white borderline of the obtained binarized image to obtain an outline image; And detects a base candidate group of the vehicle from the obtained outline image.

In order to detect the candidate base group of the vehicle from the obtained outline image, clustering points having a size of 1 in each row of the outline image are clustered, and a cluster having a length of at least the minimum vehicle length and less than the maximum vehicle length, ≪ / RTI > Among the calculated clusters, clusters forming the bottom of the same vehicle are integrated to detect the base candidate group.

In addition, the generation of the side candidate of the vehicle is performed by obtaining a vertical-edge image using the Sobel operator on the input image; Projecting a base candidate group of the generated vehicle onto a vertical edge image to set a sub region; Calculating a cumulative sum of columnar edges of the vertical edge in the subarea to calculate peak points of the cumulative sum; The distance between two points of the calculated maximum points is found as a pair satisfying the length condition of the vehicle and is set to the left and the right, thereby generating a side candidate of the vehicle.

Horizontal-edge power, vertical-edge power, and symmetry are used as feature points of the weak classifiers.

SVM (Support Vector Machine) is used as the classifier.

As described above in detail, according to one aspect of the present invention, it is possible to improve the accuracy of the vehicle detection performance by detecting a forward or a rearward vehicle by using an efficient camera in detecting a vehicle over a certain distance.

Also, the recognition and judgment of the obstacle or the front / rear vehicle are processed using only the image of the monocular camera, so that the system can be simplified by reducing the number of sensors in comparison with the existing frontal collision warning system.

1 is a flowchart showing a vehicle detection method according to an embodiment of the present invention.
2 is a view for explaining a method of setting a region of interest shown in step 200 of FIG.
3 is a flowchart showing in detail the bottom detection method of the vehicle shown in step 300 of FIG.
FIG. 4 is a diagram for explaining the ground surface histogram calculation shown in step 310 of FIG.
FIG. 5 is a view for explaining the binarized image acquisition and the outline image acquisition shown in steps 330-340 of FIG.
6 is a diagram for explaining generation of a bottom candidate group of the vehicle shown in step 350 of FIG.
7 is a flowchart showing in detail the method of detecting the side surface of the vehicle shown in step 400 of FIG.
FIG. 8 is a view for explaining the vertical edge image acquisition shown in step 410 of FIG.
9 is a diagram for explaining the auxiliary area setting shown in step 420 of FIG.
FIG. 10 is a diagram for explaining generation of a side face candidate group shown in steps 430-440 of FIG. 7 and final generation of a vehicle candidate group.
11 is a view for explaining passage of the weak classifier shown in step 500 of FIG.
12 is a view for explaining passage of the strong classifier shown in step 600 of FIG.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings.

1 is a flowchart showing a vehicle detection method according to an embodiment of the present invention. As shown in FIG. 1, a vehicle detection method according to an embodiment of the present invention includes receiving (100) a forward or backward image by a momocular camera installed in front of or behind a vehicle, (300-400), and a candidate region of the vehicle (300-400) is detected by detecting the bottom face and the side face of the vehicle within the set region of interest, The non-vehicle is firstly removed (500) by applying a weak classifier to the candidate class of the vehicle passing through the weak classifier (600).

Hereinafter, each step (100-600) of the vehicle detecting method shown in FIG. 1 will be described in detail with reference to the drawings.

2 is a view for explaining a method of setting a region of interest shown in step 200 of FIG. As shown in FIG. 2, the points in the maximum distance (for example, 200 m) that can recognize the vehicle in the three-dimensional actual coordinate system are projected onto the image using the H-matrix. Thus, an image coordinate system as shown in Fig. 2 (b) is generated. That is, the line shown in (b) of FIG. 2 corresponds to a line obtained by projecting images of the points at the maximum distance (for example, 200 m) at which the vehicle can be recognized in the three-dimensional actual coordinate system. Here, the H-matrix is a 3 × 4 intrinsic and extrinsic matrix, and is a function for mapping a three-dimensional coordinate system to a two-dimensional image coordinate system. Here, an area below a line (a line shown in (b) of FIG. 2) which is a collection of points projected by an image becomes a fixed interest area.

Hereinafter, the method of detecting the bottom surface of the vehicle shown in step 300 of FIG. 1 will be described in detail with reference to FIGS. 3 to 6. FIG.

3 is a flowchart showing in detail the bottom detection method of the vehicle shown in step 300 of FIG.

Referring to FIG. 4, a specific region within the region of interest is considered as ground (see FIGS. 4 (a) and 4 (b)). Generally, a portion of the input image that is determined as the camera proximity area is regarded as the ground, assuming that the recent room of the monocular camera that photographs the front or rear of the vehicle corresponds to the ground where no obstacle or vehicle exists. The brightness of the ground assumes a Gaussian distribution and a ground Histogram is calculated 310 (see FIG. 4 (c)). The threshold value of the shadow region is calculated using the average value (μ) and the variance value () of the histogram (320).

The brightness value of all points in the ROI of the input image (FIG. 5A) is compared with the threshold value of the shadow area calculated in step 320, and if the brightness value of the point is smaller than the threshold value, the shadow area (= 1) If the brightness value of the point is larger than the threshold value, it is determined as a non-shadow region (= 0) and a binary image as shown in FIG. 5B is obtained (330). Next, the black-and-white boundary of the binarized image obtained in step 330 is extracted to obtain the contour image as shown in FIG. 5C (340).

Then, the bottom candidate group of the vehicle is detected from the outline image acquired in step 340 (350). Generally, the underside of the vehicle is a dark area due to shadows. Clusters having a size of 1 in each row of the outline image (FIG. 6A) are clustered to calculate a cluster whose length is equal to or smaller than the minimum vehicle length and equal to or smaller than the maximum vehicle length (see the line shown in Fig. Among the clusters thus calculated, the clusters forming the bottom surface of the same vehicle are integrated (see the line shown in (c) of FIG. 6). The lines shown in (c) of Fig. 6 become the base candidate group of the final vehicle.

Hereinafter, the method of detecting the side surface of the vehicle shown in step 400 of FIG. 1 will be described in detail with reference to FIGS. 7 to 10. FIG.

7 is a flowchart showing in detail the method of detecting the side surface of the vehicle shown in step 400 of FIG.

Referring to FIG. 8, a vertical-edge image as shown in FIG. 8B is obtained by using a 3 × 3 sobel operator in an input image (FIG. 8A) (410). Next, a base candidate group of the vehicle obtained in the step 350 of FIG. 3 is projected onto a vertical edge image (FIG. 9A) (refer to FIG. 9B) c) is set (420).

Thereafter, a column direction cumulative profile of the vertical edge in the subarea is obtained, and the local maxima of the cumulative sum are calculated (see Fig. 10 (a)). A pair of left and right sides of the calculated maximum distance is searched for a pair satisfying the length condition of the vehicle (430, see FIG. 10 (b)). The bottom candidate of the vehicle and the side candidate are combined to produce the final candidate vehicle (440).

11 is a view for explaining passage of the weak classifier shown in step 500 of FIG. When a final vehicle candidate group is generated through the above-described process, a weak classifier is applied to the generated vehicle candidate group (see FIG. 11A) to primarily remove the non-vehicle 500, 11 (b)). Horizontal-edge power, vertical-edge power, and symmetry are used as feature points of the weak classifiers.

12 is a view for explaining passage of the strong classifier shown in step 600 of FIG. The strong classifier is applied to the vehicle candidate (see Fig. 12 (a)) that has passed through the weak classifier through the above-described process to determine whether the vehicle is finally in the vehicle 600 (see Fig. 12 (b)). SVM (Support Vector Machine) is used as the classifier.

Claims (7)

Receiving a photographed image from a camera;
Setting a region of interest for vehicle detection within the input image;
A specific area within the set area of interest is regarded as a ground, a ground surface histogram showing the brightness of the ground surface is calculated, and a threshold value of the shadow area is calculated using the average value (μ) threshold value of the boundary of the contour image, calculating a binarized image by comparing brightness values of all points in the area of interest with the threshold value, extracting a black and white boundary line of the obtained binarized image to obtain an outline image, Clustering points having a size of 1 in a row to calculate clusters whose length is equal to or less than the minimum vehicle length and equal to or less than the maximum vehicle length; Combining the clusters forming the bottom surface of the same vehicle among the calculated clusters to generate a bottom candidate group of the vehicle;
Generating a side face candidate group of the vehicle using the generated bottom face candidate group of the vehicle;
Generating a vehicle candidate group by combining the base candidate group and the side face candidate group of the generated vehicle;
And a weak classifier and a strong classifier are applied to the generated vehicle candidate group to detect whether the vehicle is an actual vehicle.
2. The method of claim 1,
And projecting the points at a maximum distance at which the vehicle can be recognized in the three-dimensional real coordinate system onto the input image to set the ROI. delete delete 2. The method according to claim 1, wherein the side-
Acquiring a vertical-edge image using the Sobel operator on the input image;
Projecting a base candidate group of the generated vehicle onto the vertical edge image to set a sub region;
Calculating a cumulative sum of columnar edges of the vertical edge in the subarea to calculate peak points of the cumulative sum;
Calculating a distance between two points of the calculated maximum points to find a pair satisfying a length condition of the vehicle, and setting the left and right sides to generate a side face candidate group of the vehicle. The method according to claim 1,
Wherein the feature point of the weak classifier is a horizontal-edge power, a vertical-edge power, and a symmetry. The method according to claim 1,
Wherein the SVM is a SVM (Support Vector Machine).

Images