CN103927526A - A vehicle detection method based on Gaussian difference multi-scale edge fusion - Google Patents

Abstract

A vehicle detection method based on Gaussian difference multi-scale edge fusion is characterized in that Gaussian scale transformation is firstly carried out on an image to obtain four Gaussian images with adjacent scales. And then carrying out difference operation between adjacent scale images on the four adjacent scale Gaussian images to obtain three Gaussian difference images with different scales, carrying out edge detection on the Gaussian difference images by adopting a Sobel operator, carrying out edge fusion of scale upward search to obtain vehicle edge information as much as possible, simultaneously removing a large number of background edges, carrying out a series of morphological operations such as expansion, closing operation, hole filling and the like on the fused edge images to obtain a connected domain image representing the vehicle, and finally determining an external rectangle of the position of the vehicle in the original image according to the position information of the connected domain to realize vehicle detection. The invention processes the multi-scale image, reduces the algorithm complexity, reduces the calculated amount, can effectively improve the vehicle detection efficiency and obtains better detection results.

Description

A vehicle detection method based on Gaussian difference multi-scale edge fusion

technical field

The invention belongs to the field of video detection, and in particular relates to a vehicle detection method based on Gaussian difference multi-scale edge fusion.

Background technique

Moving object detection is a key technology of computer vision and image pattern recognition. Vision-based vehicle detection technology is a research hotspot in intelligent transportation image processing, and has a wide range of applications in the field of intelligent transportation, such as vehicle assisted driving systems, traffic parameter statistics systems, etc.

Vehicle detection methods based on computer vision can be roughly divided into three categories: model-based, neural network-based learning, and feature-based methods. The model-based detection method matches the detected candidate vehicle area with the pre-established vehicle model in the computer database to detect the vehicle, but the disadvantage of this method is that it completely relies on geometric modeling of all different types of vehicles, which is difficult. Achieved.

The learning-based detection method uses samples to train the neural network, and uses the trained network for vehicle recognition. This method is often used to verify the detection results of other methods. Feature-based methods localize vehicles by detecting their local features such as symmetrical parts (wheels, headlights, taillights, etc.), edges, and shadows. The advantage of this method is to detect vehicles using features that are distinguishable in most environments, and it is suitable for vehicle detection problems in rainy and snowy days or even at night. The method of detecting vehicles by detecting wheels is easily affected by vehicle driving posture, occlusion and other problems, and the method of detecting vehicle lights is also interfered by street lights and city lights in night scenes, affecting the detection results. The vehicle detection method based on edge detection (including vehicle shadow detection) is inaccurate due to the existence of background edges (such as lane lines, railings, trees, etc.), so how to maximize the detection of vehicle edges while suppressing the background edge , which becomes the key issue to improve the detection accuracy of this method.

Contents of the invention

The purpose of the present invention is to overcome the problems in the prior art and provide a vehicle detection method based on multi-scale edge fusion of Gaussian difference, which reduces the complexity of the algorithm, reduces the amount of calculation, and obtains better detection results, effectively The detection efficiency is improved.

In order to achieve the above object, the present invention takes the following technical solutions to achieve:

Step 1: collect a traffic video of a road section, grayscale an image in the video and perform Gaussian pyramid multi-scale transformation, use the Gaussian kernel of four adjacent scale parameters to perform convolution operation with the image, and obtain four adjacent scale images Gaussian image G _l , where l represents four adjacent scales, l=1,2,3,4;

Step 2: Perform adjacent-scale image difference operation on the four adjacent-scale Gaussian images G _l to obtain three adjacent-scale Gaussian difference images D _l , and the scales of the three Gaussian difference images are: σ, 2×σ ,2×2×σ; where l=1,2,3, σ is the smoothing parameter;

Step 3, use the Sobel operator to perform edge detection on the three Gaussian difference images D _l obtained in step 2, calculate the gradient magnitude of each pixel in the difference image in the horizontal and vertical directions, and set the threshold T ₁ , Keep the pixel point whose gradient amplitude is greater than T ₁ , this pixel point is an edge point and its gray value is set to 255, otherwise it is set to 0, and the edge detection binary image E _l corresponding to three adjacent scales is obtained, where l= 1,2,3;

Step 4, perform multi-scale edge fusion on three binary edge maps E _l corresponding to three different scales, where l=1, 2, 3, the specific steps are:

(1) Search for each edge pixel in the edge image E l of three adjacent scale Gaussian difference images D _l at four adjacent scales _l . Since the edge displacement between adjacent scales does not exceed 1, the scale is The differential Gaussian image D of _{l-1 searches for a corresponding area of 3×3 in the edge image of l-1} , and all edge points appearing in this area are marked as edge points to obtain candidate edge images;

(2) l=l-1; if l>1, jump to step (1), otherwise execute step (3);

(3) When l=1, the edge image E _l is the edge image after fusion;

Step 5: Perform morphological processing on the fused edge image obtained in step 4 using an expansion template, set a threshold T ₂ , connect edge points or lines whose pixel spacing is smaller than the threshold T ₂ , and obtain continuous edges; then perform morphological closing operation , to bridge the holes and cracks in the edge image to obtain a further closed edge image; finally, fill the internal holes in the closed area through image filling to form a complete connected area;

Step 6: Mark the connected regions, calculate the area of each connected region, set the area threshold T ₃ , and eliminate the connected regions whose area is smaller than the area threshold T ₃ ; determine the coordinates of the smallest circumscribed rectangle of each connected region according to the coordinates of the connected regions, and finally It is displayed in the original grayscale image to complete the detection of the vehicle.

In said step one, the scale is jointly determined by the smoothing parameter σ and the smoothing parameter k, and k is determined by the parameter s, and k=2^(1/S), S+3=N, wherein N is the number in each layer of the Gaussian pyramid For the number of Gaussian pictures, N=4, S=1, σ=0.5, k=2, then the four adjacent scales are σ, kσ, 2kσ, 3kσ respectively.

The template in the horizontal direction of the Sobel operator in the step 3 is (-1,0,1;-2,0,2;-1,0,1), and the template in the vertical direction is (-1,-2,-1; 0,0,0; 1,2,1), the threshold T ₁ is obtained by the method of maximum between-class variance.

In the fifth step, the size of the expansion template is 3×3, the size of the morphological closing operation template is 8×8, and the value range of T ₂ is 1-8, and the unit is pixel.

The setting process of the area threshold in the step 6 is: sort the areas of the marked n connected domains, and use 1/4 of the area of the largest connected domain as the area threshold.

Compared with the prior art, the present invention has beneficial effects: the present invention discloses a vehicle detection method based on multi-scale edge fusion of Gaussian difference, firstly, the Gaussian kernel of four adjacent scale parameters is used to perform convolution operation with the image, Gaussian images of four adjacent scales are obtained. Then, the Gaussian images of four adjacent scales are differentiated to obtain Gaussian difference images of three adjacent scales, and the Sobel operator is used for edge detection. Then, the edge fusion of the three detected edge images with different scales is performed on the scale-up search to obtain as much vehicle edge information as possible while removing a large number of background edges. Then, a series of morphological operations such as expansion, closing operation, and hole filling are performed on the fused edge image to obtain a connected domain image representing the vehicle. Finally, according to the location information of the connected domain, the outer rectangle of the vehicle location is determined in the original image to realize vehicle detection. The invention processes non-sampled multi-scale images, avoiding the lack of edge information or false edges caused by image interpolation operations; the multi-scale edge image upward search and fusion method is used to obtain more edge information while suppressing background edges. This method reduces the complexity of the algorithm, reduces the amount of calculation, can effectively improve the efficiency of vehicle detection, and obtain better detection results.

Description of drawings

Fig. 1 is the grayscale image to be detected;

Figures 2 to 4 are Gaussian difference images of three adjacent scales;

Figure 5 is the Sobel operator template;

Figure 6 is the edge detection result under scale 1;

Fig. 7 is the edge image after fusion;

Fig. 8 is the connected domain image obtained by the morphological closing operation;

Figure 9 is a complete connected domain image obtained after hole filling;

Figure 10 is a connected domain image representing a vehicle;

Figure 11 shows the vehicle detection results in rainy days.

Figure 12 shows the vehicle detection results on sunny days.

Fig. 13 is a schematic diagram of the vehicle detection algorithm process of Gaussian difference multi-scale edge fusion.

Detailed ways

The present invention provides a vehicle detection method based on Gaussian difference multi-scale edge fusion, which performs edge detection on a vehicle image with a size of W×H in multiple scales, and determines the vehicle position through the edge position information after multi-scale fusion. In order to realize vehicle detection.

The present invention is described in detail below in conjunction with accompanying drawing, referring to Fig. 13, the method of the present invention adopts following several steps to realize specifically:

Step 1: collect a traffic video of a road section, collect a vehicle image I with a size of 517×363 in the video, grayscale it to obtain a grayscale image, and then perform Gaussian pyramid multi-scale transformation. Using four Gaussian kernels G of adjacent scale parameters to perform convolution operation with image I, four unsampled Gaussian images G _l (l=1,2,3,4) of adjacent scales are obtained, where l represents four adjacent scale.

The scale is jointly determined by the smoothing parameter σ and the smoothing parameter k, and the smoothing parameter k is determined by the parameter s. k=2^(1/S), S+3=N, where N is the number of Gaussian pictures in each layer of the Gaussian pyramid (according to the Lowe paper), and N=4 in the present invention, S=1, σ=0.5 , k=2, take σ=0.5, S=1, k=2, the four adjacent scales are σ, kσ, 2kσ, 3kσ, namely σ, 2×σ, 2×2×σ, 3×2× σ.

Step 2: Perform adjacent-scale image difference operations on the four Gaussian images G _l (l=1,2,3,4) of the same size and different scales that have not been down-sampled to obtain three adjacent-scale Gaussian difference images D _l , where l=1, 2, 3 and the scales of the three Gaussian difference images are: σ, 2×σ, 2×2×σ, respectively.

Step 3, use the Sobel operator to perform edge detection on the three Gaussian difference images D _l obtained in step 2 respectively. The Sobel operator calculates the gradient magnitude of each pixel in the difference image in the horizontal and vertical directions, and uses the maximum between-class variance (OTSU) method to set the threshold T ₁ , and retains the pixels whose gradient magnitude is greater than T ₁ as The edge points are set to 1, otherwise they are set to 0, and three corresponding edge detection maps E _l (l=1,2,3) of different scales are obtained.

The horizontal template of the Sobel operator is (-1,0,1;-2,0,2;-1,0,1), and the vertical template is (-1,-2,-1;0,0,0 ;1,2,1).

Step 4, perform multi-scale edge fusion on the three edge detection images E _l (l=1, 2, 3), the steps are:

(1) Search for each edge pixel in the edge image E _l of the differential Gaussian image D _l at scale l. Since the edge displacement between adjacent scales does not exceed 1, in the differential Gaussian image D _l- In the edge image of ₁ , the corresponding area of 3×3 is searched, and all edge points appearing in this area are marked as edge points, and the candidate edge image is obtained;

(2) l=l-1; if l>1, jump to step (1), otherwise execute step (3);

(3) When l=1, the edge image E _l is the edge image obtained after fusion.

In step five, the edge image obtained in step four is subjected to morphological processing using an expansion template, which is called line expansion processing, and the edge points or lines whose pixel spacing is smaller than the threshold T ₂ are connected to obtain continuous edges. Then, the morphological closing operation is performed to bridge the holes and cracks in the edge image to obtain a further closed edge image. Finally, the internal cavity of the closed area is filled by image filling to form a complete connected area.

Wherein, the size of the expansion template is 3×3, the size of the morphological closing operation template is 8×8, and T ₂ =1˜8 (in pixels).

Step 6: mark the connected regions and calculate the area of each connected region, set the area threshold T ₃ , and the connected regions smaller than the area threshold T ₃ are regarded as background regions and eliminated. Determine the coordinates of the smallest circumscribed rectangle of each connected domain according to the coordinates of the connected domain, and finally display it in the original grayscale image to complete the detection of the vehicle.

Among them, the setting of the area threshold T ₃ : sort the areas of the marked N connected domains, and take 1/4 of the area of the largest connected domain as the area threshold.

Specific embodiments of the present invention are given below, and it should be noted that the present invention is not limited to the following specific embodiments, and all equivalent transformations done on the basis of the scheme of the present application all fall within the scope of protection of the present invention.

Example:

Collect a rainy day traffic image with a size of 517×363 to grayscale to obtain a grayscale image, as shown in Figure 1, select the scale parameters σ=0.5, k=2, and use the scales as σ, 2×σ, 2×2 The four Gaussian kernels of ×σ, 3×2×σ are respectively convolved with the grayscale image to obtain non-sampled Gaussian blurred images of four adjacent scales.

The adjacent scale difference operation is performed on the Gaussian blur images of the same size but different scales that have not been downsampled in four adjacent scales to obtain three Gaussian difference images of adjacent scales, that is, the difference between the Gaussian images of scale 1 and scale 2 is obtained. The Gaussian difference map of scale 2 and scale 3 is the Gaussian difference map of scale , and the Gaussian image difference of scale 3 and scale 4 is the difference map of scale . The difference of Gaussian diagram is shown in Figures 2-4.

Using the horizontal direction template (-1,0,1;-2,0,2;-1,0,1) as shown in Figure 5, the vertical direction template (-1,-2,-1;0,0, 0; 1, 2, 1) Sobel operator performs edge detection on three Gaussian difference maps, and obtains three edge detection maps of adjacent scales. The edge map of scale 1 is shown in Figure 6. Take each edge point in the scale 3 edge map as the center, search its 3×3 neighborhood in the scale 2 edge map, keep the edge points in the neighborhood, and get the edge after the first fusion until the end of the search picture. For each edge point in the edge map, continue to perform the same operation above in the edge map with a scale of 1 to obtain the final edge fusion image, as shown in FIG. 7 .

Take T ₂ =3, perform morphological expansion on the edge fusion image with a 3×3 template, and then perform morphological closing operation with an 8×8 template to obtain a connected domain with closed edges, as shown in Figure 8 . Carry out image filling on Figure 8 to bridge the holes in the connected domain to obtain a complete connected domain image (as shown in Figure 9).

Mark the connected domains in Figure 9 (the marking result is 16), calculate the area of each connected domain and sort them in descending order, take 1/4 of the largest connected domain area as the threshold (2300 in this example), and remove For connected domains whose area is smaller than the threshold, the obtained connected domain image (as shown in Figure 10) is regarded as the location of the vehicle. Draw the circumscribed rectangle of the connected domain at the corresponding position in Figure 1 to obtain the vehicle detection results in rainy weather, as shown in Figure 11. Fig. 12 is the detection result obtained by applying the present invention to another scene (sunny day).

It can be seen from Figures 11 and 12 that the vehicle detection is carried out according to the above method, and a better detection result is achieved. This example shows that the scheme algorithm of the present invention is simple, reduces the amount of calculation, and realizes better vehicle detection at the same time.

The invention discloses a vehicle detection method based on multi-scale edge fusion of Gaussian difference. Firstly, Gaussian kernels of four adjacent scale parameters are used to perform convolution operation with images to obtain four Gaussian images of adjacent scales. Then, the Gaussian images of four adjacent scales are differentiated to obtain Gaussian difference images of three adjacent scales, and the Sobel operator is used for edge detection. Then, the detected edge images of three adjacent scales are merged by scale-up search to obtain as much vehicle edge information as possible while removing a large number of background edges. Then, a series of morphological operations such as expansion, closing operation, and hole filling are performed on the fused edge image to obtain a connected domain image representing the vehicle. Finally, according to the location information of the connected domain, the outer rectangle of the vehicle location is determined in the original image to realize vehicle detection. The invention processes non-sampled multi-scale images, avoiding the lack of edge information or false edges caused by image interpolation operations; the multi-scale edge image upward search and fusion method is used to obtain more edge information while suppressing background edges. This method reduces the complexity of the algorithm, reduces the amount of calculation, can effectively improve the efficiency of vehicle detection, and obtain better detection results.

Claims (5)

1. the vehicle checking method merging based on difference of Gaussian multi-scale edge, is characterized in that, comprises the following steps:

Step 1, gather certain road section traffic volume video, the piece image gray processing in video is obtained to original-gray image, then carry out gaussian pyramid multi-scale transform, utilize gaussian kernel and the image of four adjacent scale parameters to carry out convolution algorithm, obtain the Gaussian image G of four adjacent yardsticks _l, wherein l represents four adjacent yardsticks, l=1,2,3,4;

Step 2, to the Gaussian image G of these four adjacent yardsticks _lcarry out adjacent scalogram as calculus of differences, obtain the difference of Gaussian image D of three adjacent yardsticks _l, the yardstick of three these difference images of panel height is respectively: σ, 2 × σ, 2 × 2 × σ; Wherein l=1,2,3, σ is smoothing parameter;

Step 3, this difference image of three panel heights D that step 2 is obtained _ladopt Sobel operator to carry out rim detection, calculate the gradient magnitude of each pixel on level, vertical both direction in difference image, and threshold value T is set ₁, retain gradient magnitude and be greater than threshold value T ₁pixel, this pixel is marginal point to establish its gray-scale value be 255, otherwise is made as 0, obtains the rim detection binary map E of corresponding three adjacent yardsticks _l, wherein l=1,2,3;

Step 4, to three width two-value outline map E of corresponding three different scales _lcarry out multi-scale edge fusion, wherein, l=1,2,3, concrete steps are:

(1) under four adjacent yardstick l, search for the difference of Gaussian image D of three adjacent yardsticks _ledge image E _lin each edge pixel, because the edge dislocation between adjacent yardstick is no more than 1, the difference Gaussian image D that is l-1 at yardstick _l-1edge image in to search for corresponding area be 3 × 3 region, all marginal points that occur in this region are all labeled as marginal point, obtain candidate's edge image;

(2) l=l-1; If l>1 jumps to step (1), otherwise execution step (3);

(3) when l=1, edge image E _lit is the edge image after merging;

Step 5, the edge image after the fusion that step 4 is obtained adopts expansion template to carry out morphology processing, setting threshold T ₂, connect pel spacing and be less than threshold value T ₂marginal point or line, obtain continuous boundary; Carry out again closing operation of mathematical morphology, make hole and the crack of edge image up, the edge image that obtains being further closed; Fill the interior void filling of enclosed region finally by crossing image, form complete connected region;

Step 6, carries out mark to connected region, calculates the area of each connected region, and area threshold T is set ₃, reject area and be less than area threshold T ₃connected domain; The coordinate of determining the minimum boundary rectangle of each connected domain according to connected domain coordinate finally shows in original-gray image, completes the detection to vehicle.

2. a kind of vehicle checking method merging based on difference of Gaussian multi-scale edge according to claim 1, is characterized in that, described step 1 mesoscale is determined jointly by smoothing parameter σ and smoothing parameter k, k is determined by parameter s, and k=2^ (1/S), S+3=N, wherein N is the Gauss's picture number in the every one deck of gaussian pyramid, gets N=4, S=1, σ=0.5, k=2, four adjacent yardsticks are respectively σ, k σ, 2k σ, 3k σ.

3. a kind of vehicle checking method merging based on difference of Gaussian multi-scale edge according to claim 1, is characterized in that, in described step 3, the template of Sobel operator horizontal direction is (1,0,1;-2,0,2;-1,0,1), vertical direction template is (1 ,-2 ,-1; 0,0,0; 1,2,1), threshold value T ₁obtained by maximum variance between clusters.

4. a kind of vehicle checking method merging based on difference of Gaussian multi-scale edge according to claim 1, is characterized in that, in described step 5, bulging die board size is 3 × 3, and form closed operation template size is 8 × 8, T ₂span is 1～8, and taking pixel as unit.

5. a kind of vehicle checking method merging based on difference of Gaussian multi-scale edge according to claim 1, it is characterized in that, in described step 6, the setting up procedure of area threshold is: the area to the n of a mark connected domain sorts, using largest connected territory area 1/4 as area threshold.