TWI480808B - Vision based pedestrian detection system and method - Google Patents

Description

Pedestrian detection system and method

The present invention relates to a pedestrian detection system and method, and more particularly to a technique for increasing the sensitivity and accuracy of pedestrian detection by using an image gradation gradient bar graph alignment technique.

The existing image-based driving safety devices for pedestrian detection have a wide variety of designs and productions, and the results are different.

Traditionally, the prior art has achieved the purpose of detecting pedestrians through template comparison. The main purpose is to construct a humanoid model or module with different angles and postures, and compare the detected images to achieve pedestrian detection. In the prior art, a human silhouette or an edge image is used to represent a human figure and is converted into a DT (distance transform) image. In order to more effectively overcome the displacement, proportional and rotational changes of the object, a human figure of the microwave coefficient is developed. In addition, the HG (histogram of oriented gradients) is also used to represent humanoid features, and through the SVM (Supported Vector Machine) machine learning method to identify and classify effectively distinguish between pedestrians and non-pedestrians. The core of image pedestrian detection technology.

Therefore, HOG can overcome the variation of pedestrian appearance and achieve better detection results. HOG is calculated by dividing the image into complex blocks and counting the pixel gradients in each block in each direction. The sum of the intensities and form a gradient statistical histogram. HOG has a strong ability to describe the edge information of the object, and because of the statistical calculation method, the HOG can adapt to the edge displacement and partial rotation.

However, due to the nature of statistical calculations, HOG lacks information on materials, that is, for a single complete line and fragmented lines, it is impossible to effectively separate. Therefore, people are in a messy environment, and HOG is prone to misjudgment.

Therefore, how to solve the above-mentioned shortcomings is urgently needed to be solved by the industry.

The object of the present invention is to provide a pedestrian detection system for capturing an image of a vehicle around a vehicle by a camera module disposed in a vehicle, and then extracting a direction gradient straight line from the two feature extractors in the processing module. HOG (histogram of oriented gradients) and features of the histogram of gradient of granule feature (HOGG), which integrates these two features to obtain a single feature with both features, through the support vector Machine classifier (SVM) to judge pedestrians/non-pedestrians.

Another object of the present invention is to provide a pedestrian detection method, the method comprising: capturing an image by a camera module; capturing a fine grain direction gradient (HOGG) feature of the image and converting the image into a fine grain direction gradient image; A classifier classifies and judges the fine grain direction gradient image. The method for extracting the gradient characteristic of the fine grain direction of the image comprises: dividing the image into a plurality of fine particles; calculating an average value of each of the fine particles; The average difference in fine grain intensity in the diagonal direction is converted into a complex feature vector; the feature vector in the statistical block unit; and the fine grain direction gradient image is obtained.

Similarly, the image can perform an HOG capture step to convert the image into a directional gradient image, which is combined with the fine grain direction gradient image into a HOG+HOGG feature image, so that the classifier can be further based on the HOG. The +HOGG feature image produces more accurate classification results, which improves the ability of pedestrians to judge.

In order to facilitate the review committee to have a better understanding and understanding of the technical means and operation process of the present invention, the embodiments are combined with the drawings, and the details are as follows.

Referring to FIG. 1a and FIG. 1b, a pedestrian detection system according to a preferred embodiment of the present invention includes a camera module 10 and a processing module 20.

The camera module 10 is disposed in a vehicle 4 and provides an image captured around the vehicle 4.

The processing module 20 has two feature extractors 201, 202 and a classifier 203, and receives the image for image analysis to determine whether there is a pedestrian in the image.

In the preferred embodiment, the camera module 10 can capture the image of the front of the vehicle 4 at the mirror position after the vehicle 4 is disposed as shown in FIG. 1a, and can also be disposed behind the vehicle 4. The square or the roof of the vehicle 4 is used for image capture in all directions, as shown in Figure 1b.

In the preferred embodiment, the processing module 20 can transmit the analysis result of the image, and then generate a display signal to a display module 30, which can display not only the image but also the pedestrian appearing in the image. Mark it.

In the preferred embodiment, the processing module 20 uses HOG (histogram of oriented gradients) and fine grain direction gradient bar graphs (HOGG, respectively) through each of the feature extractors 201 and 202. The feature of the image is taken out, and the classifier 203 is further used to determine whether there is a pedestrian in the image by using a Supported Vector Machine (SVM).

Referring to Figures 2 to 4, the pedestrian detection method provided by the preferred embodiment of the present invention has the following steps.

First, in step S11, an image is captured by a camera module 10, and then step S12 is performed.

In step S12, the processing module 20 captures the fine grain direction gradient feature of the image, and converts the image into a fine grain direction gradient (HOGG) image, and then proceeds to step S13.

Step S13, classifying the fine grain direction gradient (HOGG) image through a classifier in the processing module 20, and determining whether there is a pedestrian feature in the fine grain direction gradient (HOGG) image.

In the preferred embodiment, as described in step S12, when the processing module 20 When the fine grain direction gradient characteristic of the image signal is captured, a step S14 can be performed at the same time (as shown in the flow chart of FIG. 3). As shown in FIG. 3, in step S14, the direction gradient feature of the image signal can be captured by the processing module 20, and the image is converted into a direction gradient (HOG) image, and then the merged image feature becomes the direction. The sum of the features of the gradient bar graph (HOG) and the fine grain direction gradient bar graph (HOGG) (step S15).

In the case of a fine grain direction gradient bar graph (HOGG), the image is divided into a plurality of cells 5 ' (cell), and each of the cells 5 ' includes a plurality of fine particles 5 " (granule), as shown in Fig. 5a. A unit 5 ' composed of 2 × 2 fine particles 5 " is shown (step S121). Let G _i represent the region of each of the fine particles 5 " , such as G ₁ , G ₂ , G ₃ and G ₄ shown in Fig. 5a, and assume that the intensity of the image on the coordinates (u, v) can be expressed as I (u, v), then the average particle strength average f (G _i ) can be obtained by applying the formula (step S122):

Where | G _i | is after. "Contact area by area, each of the fine particles to obtain 5 'of each of the fine particles 5 of average strength, can be further transmitted through each of the fine particles of the average of the diagonal direction 5' strength The difference, that is, through f ( G ₁ )- f ( G ₄ ) and f ( G ₂ )- f ( G ₃ ), obtains the feature vector of each unit 5 ' (step S123), which is determined by a magnitude (magnitude) And an orientation (orientation), as shown below:

Orientation: θ _Cell = atan2( f ( G ₁ )- f ( G ₄ ), f ( G ₂ )- f ( G ₃ )) (3)

Suppose a block 5 consists of 4 x 4 cells, as shown in Figure 5b. Then, the block 5 can scan in units of 5 ' of the unit, and then select 9 units 5 ' to perform operations, and obtain 9 feature vectors. If the block 5 is divided into 9 equal parts every 20 degrees in the range of 0 to 180 degrees, as 9 bins, and vote with the orientation value of the feature vector of each unit 5 ' , The size value of the feature vector represents the number of votes, and statistics can be performed in the block 5. If illustrated by another embodiment, a 128×64 pixel image, which is divided into 32×16 cells 5 ′ and equivalent to 16×8 blocks 5, can generate 15×7 voting results. That is, there will be 105 feature vectors as features of the 128×64 pixel image.

Therefore, each unit 5 ′ included in the image can perform the above operation to obtain a representative feature quantity of the image, and the regional statistic is used to convert the image into the fine grain direction gradient (HOGG). Image (step S124).

In addition, in the case of a Directional Gradient Bar Graph (HOG), the image is also divided into the cells 5 ' (cell), and each of the cells 5 ' is also composed into blocks 5 . To find the feature vector. The difference is that the direction gradient bar graph (HOG) is obtained by using the difference between the average intensity of a single unit 5 ' and the neighboring unit 5 ' , and the feature vector is obtained, and the direction is also obtained by voting statistics. Gradient bar graph (HOG) image, so the method steps are not repeated here.

Furthermore, as shown in FIG. 3, the fine grain direction gradient (HOGG) image and the direction gradient (HOG) image converted by the image are combined by an image combining module (step S15) to become a HOG+. HOGG feature image.

In the preferred embodiment, the classifier 203 uses a training sample for training before performing image classification. The training sample includes a plurality of human images (positive images) and a plurality of unmanned images (reverse images). After training, the human images cause the classifier 203 to output a positive signal when it is determined that the image has a pedestrian; otherwise, the unmanned images cause the classifier 203 to determine that the invisible person appears in the image. , output a negative signal.

In the preferred embodiment, the classifier 203 employs a supported vector machine (SVM), which performs pre-training when offline, establishes a multi-dimensional space through the training samples, and A hyper plane is established between the positive example image and the counterexample image as a basis for the image pedestrian judgment. Since the SVM is a commonly used tool for identification and classification, its operation method will not be described here.

After the classifier 203 outputs the positive signal and the negative signal, the processing module 20 marks the position of the pedestrian included in the image according to the positive signal and the negative signal. The display module 30 performs display.

It can be seen that the pedestrian detection system and method of the present invention captures the image through the camera module 10, and the processing module 20 extracts the feature value and the comparison classification to determine whether the image contains pedestrians. feature. Since the ability of the direction gradient bar graph (HOG) is difficult to judge for complex lines, a technique of introducing a fine grain direction gradient bar graph (HOGG), which is diagonal in each of the units 5 ' divided by the image, is also introduced. The average intensity of each of the fine particles in the direction is subtracted to obtain a feature vector of each of the cells 5 ' , and then converted into the fine grain direction gradient (HOGG) image by the statistics of each of the blocks 5, and the gradient with the direction The (HOG) image is merged into the HOG+HOGG feature image to achieve an improvement in pedestrian judgment.

The detailed description of the preferred embodiments of the present invention is not intended to limit the scope of the present invention, and the equivalent implementations or modifications of the present invention should be included in the present invention. In the scope of patents.

S11~S15‧‧‧Steps

S121~S124‧‧‧Steps

10‧‧‧ camera module

20‧‧‧Processing module

201‧‧‧Character Extractor

202‧‧‧Character Extractor

203‧‧‧ classifier

30‧‧‧Display module

4‧‧‧ Vehicles

5‧‧‧ Block

5 ' ‧ ‧ unit

5 " ‧‧‧ fine grain

1 is a schematic diagram of a pedestrian detection system of the present invention; FIG. 2 is a flow chart of steps of a pedestrian detection method (HOGG) according to the present invention; and FIG. 3 is a flow chart of steps of a pedestrian detection method (HOGG+HOG) according to the present invention; Fig. 4 is a diagram showing the operation steps of the fine particle direction gradient (HOGG) of the present invention; Fig. 5a is a unit composed of a plurality of fine particles; and Fig. 5b is a block composed of a plurality of units.

S11~S15‧‧‧Steps

Claims (5)

A pedestrian detection system includes: a camera module disposed in a vehicle and providing images for capturing around the vehicle; and a processing module having two feature extractors and a classifier, and receiving the The image is subjected to image analysis to determine whether there is a pedestrian in the image, and each of the feature extractors adopts a direction gradient straight bar graph and a fine grain direction gradient straight bar graph to extract the features of the image; wherein the classifier A classification is provided to determine whether there is a pedestrian in the image. The processing module displays the result of the image analysis through a display module. A pedestrian detection method includes: capturing an image by a camera module; capturing a fine grain direction gradient (HOGG) feature of the image, and converting the image into a fine grain direction gradient image; and a classifier The fine grain direction gradient images are classified to determine whether there are pedestrians. The pedestrian detection method according to claim 2, wherein the step of extracting the fine grain direction gradient (HOGG) feature of the image and converting the image into a fine grain direction gradient image comprises dividing the image into a plurality of fine particles; an average value of each of the fine particles; The complex feature vector is obtained from the average difference of the fine grain strength in the diagonal direction of the complex unit; and the feature vector in each block is counted in units of blocks to obtain the grain direction gradient image. The pedestrian detection method according to claim 3, wherein each of the fine particles has an area smaller than each of the units, and each of the unit areas is smaller than each of the blocks. For example, in the pedestrian detection method described in claim 2, the step of extracting the fine grain direction gradient (HOGG) of the image and converting the image into a fine grain direction gradient image can simultaneously perform a direction A gradient (HOG) capture step converts the image into a directional gradient image that is merged with the fine grain direction gradient image by an image combining module.