KR101050687B1 - Pedestrian recognition device and method using posterior probability and pedestrian protection device using it - Google Patents

Abstract

The present invention relates to a pedestrian recognition device and method using a posterior probability and a pedestrian protection device using the same.

The present invention provides a device for recognizing a pedestrian in a periphery of a vehicle, the apparatus comprising: a pedestrian classifier configured to classify a pedestrian in an image by extracting features from an image of the periphery of the vehicle; A classifier optimizer that optimizes a pedestrian classifier based on a performance evaluator; And a post-processor that performs post-processing to calculate A Posteriori Probability using previous outputs of the optimized pedestrian classifier and detects pedestrians according to the calculated post-probability. Provide a pedestrian recognition device.

According to the present invention, it is possible to improve the recognition rate for the images corresponding to the pedestrians while maintaining a very low recognition rate reduction for the images corresponding to the obstacles. Can improve.

Pedestrian, awareness, protection, performance, evaluation, optimization, post, probability

Description

Pedestrian Recognition Method and Apparatus Using A Posteriori Probability and Pedestrian Protection Apparatus Using Same}

The present invention relates to a pedestrian recognition device and method using a posterior probability and a pedestrian protection device using the same. More specifically, the present invention relates to a device and a method for improving the accuracy of recognizing a pedestrian in an image of surroundings of a vehicle, and consequently, a pedestrian protection device to better protect the pedestrian.

Pedestrian protection is one of the most interesting topics in the automotive industry. Recently published research on pedestrian protection systems has shown that pedestrian safety is emerging as an important global problem that has long been applied. Among various technologies, vision-based pedestrian recognition is recognized as essential for the Active Pedestaian Protection System (APPS).

Pedestrian awareness is one of the basic elements of an active pedestrian protection system. In particular, there have been various studies to develop a vision-based pedestrian classifier. Recently, the DaimlerChrysler (PX) pedestrian image database has been published, and there have been attempts to objectively compare popular features with classifiers.

In order to solve the above problems, the present invention is to improve the performance of the pedestrian recognition system and further improve the performance of the pedestrian protection system by recognizing the pedestrian using the posterior probability in the image taken around the vehicle. have.

According to an aspect of the present invention, there is provided an apparatus for recognizing a pedestrian in a periphery of a vehicle, the apparatus comprising: a pedestrian classifier configured to classify a pedestrian in an image by extracting a feature from an image of the periphery of the vehicle; A classifier optimizer that optimizes a pedestrian classifier based on a performance evaluator; And a post-processor that performs post-processing to calculate A Posteriori Probability using previous outputs of the optimized pedestrian classifier and detects pedestrians according to the calculated post-probability. Provide a pedestrian recognition device.

In addition, according to another object of the present invention, in the device for recognizing and protecting the pedestrian in the vicinity of the vehicle, the performance evaluator based on the pedestrian classifier for classifying the pedestrian in the image by extracting features from the image taken around the vehicle Post-processing to calculate A Posteriori Probability using the previous outputs of the optimized pedestrian classifier, and generate pedestrian information by detecting pedestrians according to the calculated post-probability. A pedestrian identifier transmitting; A driving trajectory measuring unit configured to receive driving information of the vehicle from a plurality of sensors detecting the behavior of the vehicle, calculate and transmit a driving trajectory of the vehicle; A collision avoidance control unit which receives a driving trajectory from a driving trajectory measuring unit and receives pedestrian information from a pedestrian recognizer, evaluates a risk of collision between the vehicle and a pedestrian, and generates and transmits a braking command; And an active braking unit for braking the vehicle according to the braking command when the braking command is received from the collision avoidance control unit.

According to still another object of the present invention, there is provided a method for recognizing a pedestrian in a periphery of a vehicle, the method comprising: classifying a pedestrian in an image by extracting features from an image of the periphery of the vehicle; The classifier optimizer optimizing the pedestrian classifier based on the performance evaluator; And performing post-processing to calculate A Posteriori Probability using the previous outputs of the optimized pedestrian classifier and detecting the pedestrian according to the calculated post-probability. A pedestrian recognition method is provided.

As described above, according to the present invention, the recognition rate of the images corresponding to the pedestrians can be improved while the recognition rate reduction of the images corresponding to the obstacles is kept very small. As a result, the reliability of the pedestrian recognizer can be improved. Improve the reliability of pedestrian protection systems.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. First of all, in adding reference numerals to the components of each drawing, it should be noted that the same reference numerals have the same reference numerals as much as possible even if displayed on different drawings. In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear.

The present invention summarizes four experimental results for the DCX database. First, the Gabor filter bank (GFB) -based feature is competitive in pedestrian recognition. Second, the optimization of the classifier using the performance estimator greatly improves the performance of the resulting classifier. Third, if the imaging system uses the same capture method as histogram equalization and the database, the actual pedestrian recognition system can be reconstructed into an open database without the need for retraining using a new database acquired with the actual imaging system. have. Fourth, postprocessing based on Aposteriori Probabilty increases the recognition rate of continuous image sequencing in the process of maintaining a positive error rate.

Introduction

Some researchers have contributed greatly to the vision-based pedestrian recognition research by publishing a large database of pedestrian imagery called the DCX database and reporting the results of key features and classifiers. They compared three types of features: Principle Component Analysis (PCA), Harr Wavelet, and Local Receptive Field (LRF). In addition, they provide four types of classifiers: Quadratic Support Vector Machine (Quadrtic SVM), Radial Basis Function Support Vector Machine (RBF SVM). , Hereinafter referred to as 'RBF-SVM', a k-Nearest Neighbor Classifier and a Feed-Forward Neural Network. They concluded that Local Featuve is superior to global features and adaptive features are superior to non-adaptive features. Of all the combinations of features and classifiers, the LRF-Quadratic SVM combination performed the best.

After the DCX database is published, histogram gradients (HOG) with evaluation of Kernel Principle Component Analysis (KPCA), Fisher's Linear Discriminant, and Boosting Cascacle Classifier (HOG) Experimental results such as Gradients, hereinafter referred to as 'HOG', and comparisons between three features (LRF, HOG, and Region Covariance) have begun to be reported. In particular, several researchers have shown that HOG outperforms LRF using databases from DCX, MIT and Institut National de Recherche en Informatique et en Automatique (INRIA). In addition, the NiSIS Competition 2007, a competition involving the DCX database and the pedestrian recognition system, was held and the results were presented at the NiSIS Annual Symposium.

The present invention reports four experimental results on the DCX database. First, GFB-based features are competitive in pedestrian recognition. The GFB-RBF SVM, which is optimized by a genetic algorithm (GA) using cross-validation criterion (GA), was more practical than LRF-SVM and HOG in view of performance and complexity. Second, the optimization of classifiers using performance evaluators greatly improves the performance of the resulting classifier. Since the optimization by the α-performance evaluator does not need to leave Training Sampies for cross-validation, the Learning Procdure can use a full set of training samples. Third, once the imaging system uses histogram equalization and the same capturing method as its database, the actual pedestrian recognition system may consist of a published database. By applying the pedestrian recognition system trained with the DCX database to the pedestrian image database, it has been found that other features of many database images can be overcome by proper manipulation. Fourth, A Poteriori Probability-Based Post-processing increases the recognition rate of consecutive image sequences while maintaining a false positive. This is derived from the fact that pedestrians continue to change their appearance, but non-pedestrians are fixed.

2. GFB-RBF SVM based pedestrian recognition system

A. Gabor Filter Bank (GFB)

The Gabor Filter is defined as a product of a two-dimensional Gaussian function and a sine wave, and can measure an image power of a specific frequency and direction at a specific position. GFB is a group of heirloom filters with various forms in the frequency domain. The GFB-based features have already shown good performance in various recognition problems such as structure analysis, face recognition and vehicle recognition. In the present invention, Manjunath's methods are followed, such as dividing equality in phase angle and dividing logarithmically in magnitude. All filters are designed such that the half height boundary meets the boundaries of the surrounding filters.

The gabo filter located on the u axis in the frequency domain is defined as shown in Equation (1). W is the frequency of the spatial wave. σ _u and σ _y are the standard deviations of the u and v axes, respectively.

In the u-axis direction, the mean and standard deviation of the s-th Gaussian function located on the u-axis, and G _s are in the exponential manner with the mean and standard deviation of the first Gaussian function, as shown in Equation (2). U _l represents the lowest average frequency, U _h represents the highest average frequency, and σ _l represents the first u-axis standard deviation. Multiplication factor a in algebraic magnitude is defined as in Equation 3, which indicates that the frequency ratio of the Gabor filter is consistent in magnitude.

The average value U _s of the s-th Gaussian function located on the u-axis is defined as in Equation 4. The u-axis standard deviation σ _u and the v-axis standard deviation σ _v are defined as in Equations 5 and 6, respectively.

The direction of a two-dimensional waveform is set by replacing (x, y) with (x ', y'), where (x ', y') is the k of the K Equal Angular Section (Equation 7). Is rotated to the first section. σ _x and σ _y represent the standard deviation of the x and y axes. In particular, the Gabor Wavelot is generated by increasing the rating to a ^-3 to repeatedly divide the power of the High Pass Filter into two Child-Filters.

1 is an exemplary diagram showing a designed GFB in the frequency domain, which is composed of 24 Gabo filters.

In FIG. 1, the number of origin Ks is 6, the number of magnitudes S is 4, the lowest average frequency U _l is 0.05, the highest average frequency U _h is 0.4, and the filter size is 16 × 16. 23 Gabor filters are shown.

B. Feature Vector Construction

In the present invention, the feature extractor has four steps: histogram equalization, partitioning the image into sub-images, convolution of the auxiliary images with gabor filters, and statistics. Calculating the characteristic values.

2 illustrates an example in which an input image is divided into nine overlapping auxiliary images.

An input image having a size of 18 × 32 is divided into nine overlapping auxiliary images of size 9 × 16, as shown in FIG. To overcome the Alignment Variation, only the statistical feature values of the Gabor filtering results of the respective auxiliary images are used in the recognition process. Three statistical features (eg, mean, standard deviation, distortion of the convolution result, etc.) are calculated for each secondary image. As a result, the dimension of the feature vector is 648 (9 × 24 × 3 = 648).

C. Support Vector Machines

The support vector machine (SVM) is a sorting method with a hyperplane that maximizes the margin between the hyperplane and the two classes. By introducing the Lagrange Multiplier, maximizing the gain is transformed into a Constrained Optimization Problem. Unseparated data requires regularization parameter C, which is the upper limit of Lagrangian coefficients and means penalty error for classification error. Using RBF as a kernel function allows SVM to solve nonlinear problems. The operation of RBF is controlled by kernel parameters. If the class label y _i ∈ {-1,1} for the data set {x ₁ , ..., x _n } and x _i is known and α represents the Lagrangian coefficient, then SVM learning It can be expressed as 8.

When the Lagrange coefficients and the support vectors are determined, the classification f (x) of the input x may be expressed as in Equation (9). N _s represents the number of support vectors and s _i represents the j th support vector.

SVM will be the best classifier for pedestrian recognition using still image and shows good performance by combining image and motion characteristics. In addition, the GFB-SVM combination may be used for car and pedestrian recognition.

Although SVM performs well in a variety of problems, it is sensitive to learning parameters: Regularization Parameter C and Kernel Parameter. These two parameters greatly affect the performance of the pedestrian classifier. These two parameters can be optimized by a genetic algorithm with a Cross Validation Fitness Function.

3. Performance Evaluator-Based Classifier Optimization

는 학습 오류들의 개수의 상한 경계이다. 어떠한 결정 규칙

의 오류율의 ξα-평가기

는 수학식 10에서와 같이 정의된다. ρ는 2와 같고 R_Δ ²은 모든 학습 샘플들의 커널 함 수의 상한 경계이다. ξα-평가기에 대한 주요 아이디어는 부등식

가 유지될 때에 대한 학습 샘플들과 리브원아웃(Leave-one-out) 테스팅에서 오류를 만들 수 있는 학습 샘플들 간의 관계이다. 이는 d가 리브원아웃 오류들의 개수의 상한 경계임을 암시한다.Joachims proposed an efficient and effective approach for evaluating the general performance of SVMs. Without the need for computationally intensive resampling, the evaluator is expected to be computationally more efficient than cross-validation or booststrapping. The evaluator can only be computed once and then computed immediately without any additional cost. Since the evaluator is obtained from the Lagrangian coefficient α and the Slack Variable ξ, the evaluator is called ξα-evaluator. If the training sample is located on the wrong side of the hyperplane, the residual variable ξ _i corresponding to the training sample x _i is greater than or equal to one. therefore,

Is the upper bound of the number of learning errors. Any decision rules

Ξα-evaluator of error rate of

Is defined as in Equation 10. ρ is equal to 2 and R _Δ ² is the upper bound of the kernel function of all learning samples. The main idea for ξα-evaluator is inequality

Is the relationship between the learning samples for when is maintained and the learning samples that can make an error in leave-one-out testing. This implies that d is the upper boundary of the number of leave one out errors.

D. Chen et al. Proposed that the pedestrian classifier can be optimized by genetic algorithms that use the ξα-evaluator as a fitness function.

Although they recognized that the performance evaluator was suitable as a fit function, they did not take advantage of one important advantage of the performance evaluator, namely, the increase in the size of the training samples. If the optimization method chooses cross-validation as a measure of performance, many of the training samples for validation should be excluded. However, if a performance evaluator is used for the custom function, there is no need to leave training samples for verification. Thus, the learning process may use a full set of training samples and increasing the learning sample will improve the classifier performance. This idea will be demonstrated by experimental results in the next section.

4. Post-Processing: A Posteriori Probability

If the pedestrian recognition system uses only one still image as input, the outputs are independent from each other. However, in practical applications, targets are temporarily tracked by range data or visual data. Thus, images containing one target tracked temporarily can be associated and the likelihood that one target is a pedestrian can be calculated using the outputs of the classifier up to that point.

로 표현될 수 있다. P (Ped) and P (Non) represent the possibility of being a pedestrian and not the pedestrian, respectively. P (Ped) and P (Non) represent the respective possibilities when the classifier outputs output positive and negative results, respectively. If X is a series of classifier outputs and its length is N, then X is temporarily a series of classifier outputs.

It can be expressed as.

Each classifier output x _i has T or F. According to the Bayesian rule, the posterior probability that the target corresponding to X will be a pedestrian may be calculated as in Equation 11.

Since the classifier outputs are independent, the order of the outputs is meaningless. Thus, when the target is a pedestrian, the likelihood that the classifier output column X will have k positive outputs can be calculated as:

Without knowing whether the target is a pedestrian or not a pedestrian, the likelihood that the classifier output column X will have k positive outputs is shown in equation (13).

Therefore, when the classifier output string X has k positive outputs of N outputs, the posterior probability that the target corresponding to X is a pedestrian may be expressed as Equation (14). In this case, it can be assumed that the probability of being a pedestrian is equal to that of a non-pedestrian, as long as the generality is not lost (for example, P (Ped) = P (Non) = 0.5).

The posterior probability defined in that way will immediately follow the classifier output during the initial period of time X is output. However, after some amount of classifier outputs, the posterior probability acts like a hysteresis filter to reduce turbulent noise.

5. Experimental Results

A. Performance of GFB-RBF SVM Pedestrian Recognition System

To prove the suitability of GFB-based features in pedestrian recognition, a pedestrian classifier using GFB-based features and RBF SVM is constructed. Two classifier parameters, C and σ, were optimized by genetic algorithms associated with the DCX database. Of the three training datasets, the first and second datasets were used for training and the third dataset was used for cross validation. Cross-validation error was used as a fitting function of the genetic algorithm.

3 is an exemplary diagram illustrating a receiver operating characteristic (ROC) of the GFB-RBF SVM and the LRF-quadratic SVM.

GFB based features show almost the same performance as LRF based features. Table 1 compares the feature dimension and the number of support vectors with optimal performance.

LRF-quadratic SVM GFB-RBF SVM Feature dimension 1,792 648 Support vector no 5,160 2,379

Because the system uses smaller feature vectors and smaller number of support vectors, the GFB-RBF SVM is competitive compared to the LRF-quadratic SVM in terms of efficiency or complexity.

B. Performance Improvement by ξα-Performance Evaluator

To evaluate the suitability of the γα-evaluator as a fitting function of the genetic algorithm and the effect of using the entire training data set for learning, three optimizations were constructed as shown in Table 2.

Case Training Dataset Validation Criterion Verification Dataset Learning Dataset Validation Dataset One 1, 2 3 Cross-validation T1, T2 2 1, 2 - ξα-estimator T1, T2 3 1, 2, 3 - ξα-estimator T1, T2

Case 1 used cross validation, while Case 2 and Case 3 used ξα-evaluators. Case 2 used only Datasets 1 and 2 (Learning Dataset) for learning, while Case 3 used all three datasets for learning. These three cases used the same Verification Dataset T1 and T2. Case 1 is the optimization made in the previous section, that is, the optimization of GFB-RBF SVM by cross validation.

4A to 4B are exemplary views showing optimization results in three cases.

In FIG. 4A, the ROCs of the optimal results of Case 1 and Case 2 were compared. Here, the dotted line represents the case 1 and the solid line represents the case 2. Optimization with ξα-evaluator can achieve almost the same performance as cross-validation based optimization. In Figure 4b we compare the ROC of the optimal results of case 3 and LRF-quadratic SVM.

Table 3 summarizes the three cases and the optimization results of the LRF-quadratic SVM.

Case Feature dimension Support vector no. Performance (%) P (T | Ped) P (F | Non) LRF-quadratic SVM 1,792 5,160 - - One 648 2,379 79.68 94.58 2 648 1,906 79.91 93.95 3 648 2,830 84.05 95.38

The GFB-RBF SVM is trained using three datasets and the ξα-evaluator overcomes the LRF-quadratic SVM, but its feature vector dimension is 1/3 and its number of support vectors is about half. It is noteworthy that the optimization by the α evaluator can utilize all training datasets for learning and obtain competitive results without any additional manipulations such as Cascade and Bostrap. Its further manipulation is simply to add a data set that is used for cross validation of the training samples.

Some researchers have shown that HOG-based features are superior to LRF-based features using DCX, MIT and INRIA databases, but HOG-based features require more complexity than GFB-based features. The feature dimension of the HOG based feature is 8,064.

C. Application of configured pedestrian recognition system to real situation

This section reports experimental results when a pedestrian classifier trained with the DCX database was applied to another database without retraining. It is noteworthy that this is the first attempt to investigate the possibility of constructing a pedestrian recognition system with an open database without retraining. Once this approach is proven to be appropriate, it is not necessary to spend a lot of time constructing the database, whatever changes are made in the imaging system.

In the experiment, two test databases were used, which are conventional. The pedestrian recognition system used in this section is 3 in Table 3, which was trained with three training data sets and a ξα-evaluator.

5 is an exemplary diagram illustrating three types of input image sampling methods.

The specific database consisted of 924 pedestrian samples. Each sample has a resolution of 64 x 128. As shown in Figs. 5A to 5C, three kinds of input image sampling methods were tested. In the third, two points were entered to point to the left and right ends of the body. Based on these points, the vertical center line was aligned. The other two points were entered to point to the top and bottom of the body as shown in (c). The resulting input image was sampled while maintaining the same aspect ratio as the DCX database. Such an input image sampling method is to capture pedestrian images in a manner similar to the original database. Once the input image was sampled, it was resized to the same size as the DCX database using Bicubic Interpolation.

6A and 6B are exemplary views illustrating a pedestrian image database.

Urban street images were collected and pedestrian and non-pedestrian images were extracted according to the second and third input image sampling methods. The number of street images captured in the city was 10,398 and the resolution was 640 × 480. In FIG. 6A, samples of urban street images captured while driving a specific area were 1,237 and 928 for the total number of pedestrian and non-pedestrian images, respectively. The most important difference between a typical database and a DCX database is the non-pedestrian image selection criteria. That is, the DCX database captures random images for non-pedestrians, but a typical non-pedestrian image database is selected when detected by scanning a laser radar. 6B shows a sample of an image of a non-pedestrian selected to be laser radar detected.

Here, four different experimental methods were used to evaluate the recognition rates P (T | Ped) associated with three databases. Table 4 summarizes the evaluation results.

Experimentalal

Database Methods One 2 3 4 Histogram equalization Yes No Sampling method Original, Fig. 5 (a) Bounding Box, Fig. 5 (b) Center aligned, aspect ratio maintained, Fig. 5 (c) DCX 84.05 - - - MIT 46.54 69.37 79.47 21.97 MANDO - 69.59 80.42 21.97

Experimental Methods 1-3 used histogram equalization, but Experimental Method 4 did not use histogram equalization. Experimental results 1 and 2 used the input image sampling method of FIGS. 5A and 5B, respectively. Experimental methods 3 and 4 used the input image sampling method of (c). Using the input image sampling method and histogram equalization in (c), the recognition rates of the two databases reached about 80% without retraining. This is only 4% lower than the DCX database's recognition rate.

In addition, the pedestrian recognition system performed better when crossing, as opposed to when the pedestrian was at the front or rear. That is, P (T | Ped) when crossing was 85.22%, and P (T | Ped) when at front or back was 78.02%. This is positive considering that the number of accidents involving pedestrians and bicycles crossing the roadway is three to four times larger than the number of accidents involving pedestrians and bicycles heading in the same direction as the roadway.

7A-7B are exemplary views showing traverse pedestrian image sequences that are correctly recognized.

In FIG. 7B, the pedestrian crossing the road at high speed dynamically changes its appearance and the lighting condition changes drastically depending on the position of the pedestrian, but the entire image sequence consisting of 29 images is correctly recognized.

D. Post-Probability Based Post Processing

Since the pedestrian recognition system of Case 3 described above in Table 3 was used for post-processing, it is assumed that P (T | Ped) is 84.05% and P (F | Non) is 95.38%. A total of 1,212 pedestrian images in the sequence were post-processed. That is, the posterior probabilities for each image were calculated and distinguished by a threshold equal to 50%. The number of sequences was 91. The length of the sequences was 2 to 41 and the average length was 13.32. In contrast, the 916 non-pedestrian images in the sequence were post-processed. The number of sequences was 113. The length of the sequences was 2 to 33 and the average length was 8.11.

8A to 8C are exemplary views showing the post-processing results.

In each of Figs. 8A to 8C, the first column is the input image column. The second column is the Instant Classifier output. The third column is the post-processing result. 0 is a threshold for distinguishing the output of the classifier, which is indicated by a dashed line. The threshold for distinguishing the posterior probability is 50% and is indicated by the dotted line. 8A and 8B, it can be observed that post-processing eliminates fluctuating output noise, which is clearly caused by the changing appearance of pedestrians.

In each case, pedestrian awareness increased by 30.77% and 27.78%, respectively. In contrast, FIG. 8C shows an example of a non-pedestrian case. Despite post processing, the error rate does not change. Experimental results show that the classification of non-pedestrian targets is more competitive than that of pedestrians. In other words, once a pedestrian is classified as a pedestrian, his output tends to be a pedestrian in all sequences. In contrast, once a pedestrian is misclassified as a pedestrian, his output tends to be a pedestrian in all sequences. This is natural because the normal non-pedestrian is a fixed body and does not change its appearance in the image sequence unless the camera position is changed.

After workup, P (T | Ped) improved from 80.20% to 89.69% and P (F | Non) lowered from 84.17% to 83.19% or the positive error rate rose from 15.83% to 16.81%. The pedestrian recognition rate rose to 9.49%, while the non-pedestrian recognition rate decreased by only 0.98%.

E. Review

In this section, the practical suitability of the proposed pedestrian recognition system was examined by comparing similar systems recently published.

Gavrila proposed a unique probability approach to Hierarchical, Exemplar-Based Shape Matching. The approach uses a Template Tree to effectively represent and match samples of various shapes. The tree is generated offline by the Bottom-up Clustering Approach using Stochastic Optimization. Online matching involves the simultaneous coarse-to-fine approach to the template tree and transformation parameters. The correct detection rate is 80%.

G. Ogawa, et al . Proposed a comprehensive risk recognition system, which is called a risk guidance system with learning capabilities. They present a system as an example of automatically learning through genetic programming to obtain image processing of pedestrians and vehicles taken by onboard camera systems. The correct recognition rate for pedestrians was 81% and the false positive rate was 19.6%.

XB Cao, et al ., Have constructed a single camera-based vision system that detects pedestrians and predicts their location, direction and distance. The system's pedestrian classifier used two consecutive images and consisted of a statistically learned cascade classifier and a decomposed SVM classifier. The cascade classifier consists of 10 filters including seven filters in series using the Appearance Feature and three filters in parallel using the Motion Feature. The system used a small template tree to predict the direction of pedestrian movement. The false positive rate was 0.013%, while the resulting detection rate was 88.8%.

Comparing these recent publications, the system according to the present invention appears competitive in terms of recognition performance and complexity. In other words, the recognition rate is 89.69%, the positive error rate is 16.81%, the low feature dimension is 648, the number of small support vectors is 2.380, and the simple post-processing is based on the post-probability. Since non-pedestrian targets tend not to move, a rather high false positive rate can be solved by a risk assessment that takes into account driving conditions and relative target movement.

6. Conclusion

The present invention is advantageous in providing the following four facts.

1) GFB-RBF SVM is competitive in pedestrian recognition.

2) Performance evaluators can replace cross-validation. In this case, the dataset used for cross-validation can be used for learning and increase the resulting performance. Joachims' ξα-evaluator proves suitable for this use.

3) If the input image is processed by histogram equalization and sampled by the same method used for the learning database, the pedestrian recognition system built into the database can be used in practical applications without having to retrain with the newly acquired database. This not only increases the importance of published databases, but also reduces the repetition of tedious learning procedures after any changes have been made to the imaging system.

4) Post-probability-based postprocessing improves the recognition rate, albeit slightly increasing the positive error rate. This is believed to result from other features of pedestrians and non-pedestrians. In other words, as the pedestrian continues to change its appearance, its classifier output tends to flow in contrast to fixed non-pedestrians. Thus, post-probability-based postprocessing slightly increases false positives while increasing pedestrian recognition.

9 is a block diagram schematically showing the electronic configuration of the pedestrian protection apparatus using the posterior probability according to an embodiment of the present invention.

A pedestrian protection device using A Posteriori Probability according to an embodiment of the present invention is a device for recognizing and protecting a pedestrian in the vicinity of a vehicle, such as an active Pedestrian Protection System (APPS). It can be implemented as.

In addition, the pedestrian protection device using the post-probability according to an embodiment of the present invention to optimize the pedestrian classifier based on the performance evaluator to classify the pedestrians in the image by extracting features from the image taken around the vehicle, the optimized pedestrians Pedestrian recognizer 910 that performs post-processing to calculate A Posteriori Probability using previous outputs of the classifier, and detects and generates pedestrian information based on the calculated post-probability. ; A driving trajectory measurement unit 920 which receives the behavior information of the vehicle from a sensor detecting the behavior of the vehicle, calculates and transmits the driving trajectory of the vehicle; A collision avoidance control unit 930 for receiving a driving trajectory from the driving trajectory measuring unit and receiving pedestrian information from a pedestrian recognizer, evaluating the risk of collision between the vehicle and the pedestrian, and generating and transmitting a braking command; And an active braking unit 940 for braking the vehicle according to the braking command when receiving the braking command from the collision avoidance control unit 930.

Here, the pedestrian recognizer 910 detects a target around the vehicle and determines whether the target is a pedestrian or an obstacle, and if the target is a pedestrian, the pedestrian information (for example, location and speed) , Whether the pedestrian is a pedestrian).

In addition, the driving trajectory measuring unit 920 calculates the driving trajectory of the vehicle and generates driving trajectory information.

In addition, the Risk Assessment & Collision Avoidance (CA) / Collision Mitigation (CM) by Braking Control Logic (930) combines pedestrian information and driving trajectory information to determine whether the vehicle is in danger of colliding with the pedestrian. Evaluate the risks of collisions between vehicles and pedestrians and classify the risks by class to determine a braking order to avoid collisions or mitigate the impact of a collision, or to determine whether a risk is above a preset threshold risk Determine. Here, the braking command can be determined and generated to adjust the degree of braking of the vehicle.

In addition, the active braking unit 940 includes a wheel speed sensor that detects a wheel speed indicating a rotation speed of a steering wheel, a yaw rate sensor that detects yaw rate indicating an angular speed of a vehicle body, and a steering angle of the vehicle. Receives and outputs vehicle behavior information, such as information on whistle, yaw rate, steering angle, etc., from sensors that detect vehicle behavior, such as a steering angle sensor that detects the vehicle, and transmits it to the driving trajectory measurement unit 920, and the collision avoidance controller 930. When the braking command is received from the vehicle, the vehicle is braked using the braking device according to the braking command.

10 is a block diagram schematically illustrating an electronic configuration of a pedestrian recognition apparatus using post probability according to an embodiment of the present invention.

The pedestrian recognition device using the posterior probability according to an embodiment of the present invention may be implemented as a pedestrian recognizer 910 in FIG. 9, hereinafter, referred to as a pedestrian recognizer 910 for convenience of description.

Pedestrian recognizer 910 according to an embodiment of the present invention is a device for recognizing the pedestrian in the vicinity of the vehicle, the pedestrian classifier 1010 for classifying the pedestrian in the image by extracting features from the image taken around the vehicle; A classifier optimizer 1020 that optimizes a pedestrian classifier based on the performance evaluator; And a post-processor 1030 that performs post-processing to calculate A Posteriori Probability using the previous outputs of the optimized pedestrian classifier and detects pedestrians according to the calculated post-probability. do.

Here, the pedestrian classifier 1010 may be implemented with a Gabor Filter Bank-Radial Basis Function Support Vector Machine (GFB-RBF SVM), and a histogram from an image of surroundings of a vehicle. Characterization is performed by performing histogram equalization, segmenting the image into a plurality of sub-images, convolving the sub-images with gabor filters, and calculating statistical feature values. Can be extracted.

In addition, the classifier optimizer 1020 may optimize the pedestrian classifier 1010 using a genetic algorithm (GA) that uses a ξα-Estimator as a fitness function.

In addition, the post-processor 1030 may detect as a pedestrian when the post-probability is greater than or equal to a threshold value, and the post-processor may intentionally have a high value (that is, a preset value of the probability P (F | Non) of recognizing the obstacle as an obstacle). Post-probability can be calculated.

In addition, the pedestrian classifier 1010 may acquire a continuous image of the same object through tracking using a radar or tracking using an object contour.

11 is a flowchart illustrating a pedestrian recognition method using post-probability according to an embodiment of the present invention.

The pedestrian recognizer according to the present invention, that is, the pedestrian recognizer 910 optimizes the pedestrian classifier 910 based on the performance evaluator (S1110), calculates a posterior probability using previous output values of the pedestrian classifier 910 (S1120). ), The pedestrian is detected according to the posterior probability (S1130).

As described above, the pedestrian recognition device using the posterior probability according to the embodiment of the present invention may improve the final performance of the pedestrian recognizer by associating the outputs of the pedestrian recognizer that previously used only one still image with the posterior probability. Thus, by using the obstacle recognition rate P (F | Non), which is the probability of recognizing the obstacle as an obstacle, using an obstacle recognition rate higher than the actual obstacle recognition rate (that is, using a value for the learning database that is much higher than that for the experimental database), the posterior probability You can make it more sensitive to this T output (that is, the correct result output). That is, as the second term of the denominator of Equation 14 decreases, the influence on the posterior probability of the T output of the pedestrian recognizer increases.

In this case, the obstacle recognition rate may decrease slightly, while the pedestrian recognition rate may increase considerably (in the experiment, the obstacle recognition rate increased by 9.49% instead of only 0.98%). Such a phenomenon is not a unique characteristic of A Posteriori Filtering, but a characteristic of a pedestrian recognizer. In other words, the pedestrian recognizer output is likely to change between T and F because the pedestrian is moving and the Appearance is constantly changing. On the other hand, since the obstacle is generally a rigid object, the appearance is less likely to change, so the pedestrian recognizer output is more likely to output only one T or F value.

FIG. 12 exemplarily shows a histogram of the frequency at which a pedestrian recognizer output is changed in a sequence.

As shown, in the case of obstacles, the frequency of change of the recognizer output is concentrated at 0 to 2 times, and in the case of pedestrians, the change of the output of the recognizer is more than two times. Therefore, when post-filtering is applied, as shown in FIGS. 8A to 8C, the performance of pedestrian images whose output is changed is greatly improved, and the performance of obstacle images with constant output is not large.

The present invention actively utilizes the fact that the change in the output of the pedestrian recognizer is large in the case of the pedestrian image, but the change in the output of the pedestrian recognizer is small in the case of the obstacle image.

The foregoing description is merely illustrative of the technical idea of the present invention, and various changes and modifications may be made by those skilled in the art without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical idea of the present invention but to describe the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the equivalent scope should be interpreted as being included in the scope of the present invention.

1 is an exemplary diagram showing a designed GFB in the frequency domain consisting of 24 Gabo filters.

2 illustrates an example in which an input image is divided into nine overlapping auxiliary images;

3 is an exemplary diagram illustrating a receiver operating characteristic (ROC) of a GFB-RBF SVM and an LRF-quadratic SVM;

4A to 4B are exemplary views showing optimization results in three cases;

5 is an exemplary diagram illustrating three types of input image sampling methods;

6A and 6B are exemplary views illustrating a pedestrian image database.

7A-7B illustrate exemplary traverse pedestrian image sequences correctly recognized;

8a to 8c is an exemplary view showing a post-processing result,

9 is a block diagram schematically illustrating an electronic configuration of a pedestrian protection apparatus using a posterior probability according to an embodiment of the present invention;

10 is a block diagram schematically illustrating an electronic configuration of a pedestrian recognition apparatus using post probability according to an embodiment of the present invention.

11 is a flowchart illustrating a pedestrian recognition method using post-probability according to an embodiment of the present invention.

FIG. 12 exemplarily shows a histogram of the frequency at which a pedestrian recognizer output is changed in a sequence.

<Description of Symbols for Main Parts of Drawings>

910: pedestrian recognition unit 920: driving trajectory measurement unit

930: collision avoidance control unit 940: active control unit

1010: Pedestrian Classifier 1020: Classifier Optimizer

1030: post-processor

Claims (8)

In the device for recognizing pedestrians in the vicinity of the vehicle, A pedestrian classifier configured to classify the pedestrian in the image by extracting a feature from an image of the surrounding image of the vehicle; A classifier optimizer that optimizes the pedestrian classifier based on a performance evaluator; And Post-processing for calculating A Posteriori Probability using the previous outputs of the optimized pedestrian classifier and for detecting the pedestrian according to the calculated post-probability. Pedestrian recognition device comprising a. According to claim 1, The pedestrian classifier, Gabor Filter Bank-Radial Basis Function Support Vector Machine (GFB-RBF SVM) Gabor Filter Bank-Radial Basis Function Support Vector Machine. According to claim 1, The pedestrian classifier, Perform histogram equalization on the image, segment the image into a plurality of sub-images, and convolution the sub-image with gabor filters And extracting the feature by calculating a feature value. The method of claim 1, wherein the classifier optimizer, A pedestrian recognition apparatus characterized by optimizing the pedestrian classifier using a genetic algorithm (GA) using ξα-Estimator as a fitness function. The method of claim 1, wherein the post-processor is And detecting the pedestrian as the pedestrian when the posterior probability is greater than or equal to a threshold. The method of claim 1, wherein the post-processor is And setting the probability of recognizing the obstacle as an obstacle to a predetermined value or more to calculate the posterior probability. In the device to recognize and protect the pedestrians in the vicinity of the vehicle, A feature is extracted from the image of the surroundings of the vehicle, and the pedestrian classifier that classifies the pedestrians from the image is optimized based on a performance evaluator, and the posterior probability is calculated using previous outputs of the optimized pedestrian classifier. A pedestrian recognizer which performs post-processing to calculate a value, detects the pedestrian according to the calculated post probability, and generates and transmits pedestrian information; A driving trajectory measuring unit configured to receive the driving information of the vehicle from a plurality of sensors detecting the behavior of the vehicle, calculate and transmit a driving trajectory of the vehicle; A collision avoidance control unit that receives the driving trajectory from the driving trajectory measuring unit and evaluates a risk of collision between the vehicle and the pedestrian and generates and transmits a braking command when receiving the pedestrian information from the pedestrian recognizer; And An active braking unit which brakes the vehicle according to the braking command when the braking command is received from the collision avoidance control unit; Pedestrian protection device comprising a. In the method of recognizing a pedestrian in the vicinity of the vehicle, Classifying the pedestrian in the image by extracting a feature from an image photographing the circumference of the vehicle; A classifier optimizer optimizing the pedestrian classifier based on a performance evaluator; And A post-processor performing post-processing to calculate A Posteriori Probability using the previous outputs of the optimized pedestrian classifier and detecting the pedestrian according to the calculated post-probability. Pedestrian recognition method comprising a.

Images