The hottest vehicle detection method based on feat

  • Detail

A feature-based vehicle detection method

1. Introduction

in the field of intelligent vehicle research, vehicle detection technology is a key technology of active safety systems such as FCW (forward collimation warning), ACC (adaptive cruise control), stop & go cruise control. It is used to accurately identify the vehicle ahead and obtain relevant parameters, providing driving environment information for various active safety systems

machine vision is an effective sensor for vehicle detection. Many researchers have proposed a variety of vehicle detection methods for machine vision. (1) The method based on visual flow is used to detect vehicles []. When the visual flow method is applied to the vehicle detection of the vehicle camera, because the vehicle is moving in real time, the accuracy of the analysis of the vehicle motion is not high, and the amount of calculation is large, so it is not easy to ensure the real-time. (2) Vehicle detection method based on stereo vision [3]. The disadvantages of this method are high cost, complex algorithm, sensitive to the motion of the vehicle itself, and prone to drift in the calibration of the camera. (3) Use the training based method to detect vehicles []. This kind of method has high robustness, but its disadvantage is that it is difficult to select the sample data for training and accurately locate the vehicle position

the author adopts the feature-based vehicle detection method, which has a relatively fast algorithm execution speed, can meet the needs of monocular vision system, and does not need to use the selected vehicle feature set to train the sample data like the training based method. This method directly extracts the features used for vehicle detection from the image, so as to judge whether the vehicle exists and determine the location of the vehicle. The analysis of the image shows that the shadow under the vehicle on the road is a very stable feature to judge whether the vehicle exists, so the shadow under the vehicle and the edge of the vehicle are the main features of detection. Relying on a single sensor (such as machine vision or radar) to complete vehicle detection often faces the problems of high false detection rate and missed detection rate [7]. In order to improve the accuracy of vehicle detection and the speed of the algorithm, the obstacle data detected by radar is fused in the algorithm

II. Feature based vehicle detection algorithm

the structure of vehicle detection algorithm is shown in Figure 1

the algorithm is mainly composed of four modules. The preprocessing module is responsible for processing the original image captured by the camera, so as to obtain the images required for subsequent recognition, including the shadow image under the vehicle, the extracted horizontal edge image and the vertical edge image. The tracking module tracks these targets using the original image data and the image data obtained after preprocessing according to the vehicle recognition results of the previous round. If it is confirmed that the target still exists, it will be classified as the vehicle detected in this round; If the target has disappeared, delete it from the tracking file. Using vehicle tracking technology can reduce the execution time of the algorithm. The radar detection module compares the radar data with the vehicle position detected by the tracking module. The radar detection module is responsible for confirming the newly detected target in the image. If the target is confirmed, it will be added to the current round of detection results. The visual detection module complements the image detection. For the vehicle area detected by the tracking module and the radar detection module, the visual module will no longer repeat the inspection, so the image area that needs to be detected is greatly reduced. In order to reduce the false detection rate of the visual module, strict confirmation conditions are adopted in this part

(I) adaptive double threshold image preprocessing

it can be seen from the structure diagram of vehicle detection algorithm in Figure 1 that the preprocessing of the original image is the basis of subsequent vehicle detection. The original image contains a lot of useless information and noise. In order to extract the required vehicle edge information and the shadow information cast by the vehicle on the road from the original image, the original image must be preprocessed

firstly, Sobel operator is used to enhance the edge of the original image, and the form of Sobel operator used is shown in Figure 2

gold system adopts adaptive threshold to binarize the image [8]

where m (C, R) represents m around the pixel point (C, R) in the image after edge enhancement × M is the maximum gray value in the size region, and K is a constant. In gold system, m=7, k=2

the adaptive threshold used by gold system can treat different parts of the image differently, so as to get a better binarization result of the whole image. But its cost is to increase the amount of calculation to perform statistical analysis, which reduces the real-time processing, and its stability is not easy to be guaranteed in a complex driving environment. Therefore, the author improves the adaptive threshold used by gold system and proposes an adaptive double threshold method to recognize the shadow of vehicles on the road

at each time of image preprocessing, select two adjacent 10 on the image × For a square with 10 pixels, calculate the average pixel gray values of the two square areas respectively, and record them as V1 and V2. These two square areas are selected in the middle and lower part of the image and adjacent to each other to ensure that the selected area reflects the average gray level of the pavement in the image. Because it is often disturbed by more noise to determine the position of the shadow under the vehicle by simply relying on the recognition of the horizontal edge, two thresholds are set when binarizing the shadow image: the adaptive threshold Ttreat for the image after horizontal edge enhancement by Sobel operator and the adaptive threshold torigin for the original image

where ctreat is the weighting coefficient for calculating torigin; Corigin is the weighting coefficient for calculating torigin. Where ctreat=0.25, corigin= 0.5. Only when the original gray value of a pixel is less than torigin and the gray value of the point after horizontal edge enhancement by Sobel operator is greater than Ttreat, the point is considered as a shadow candidate

the advantages of using adaptive double threshold image preprocessing method are:

(1) by calculating V1 and V2 values for each image, it can adapt to the requirements of image binarization under different lighting conditions

(2) compared with the adaptive threshold adopted by the gold system, the amount of calculation is significantly reduced, and only two 10s need to be calculated for each image × The average pixel gray value of the square area of 10 pixels does not need to count the gray level of its adjacent area for each pixel

(3) because two square areas are selected and their average pixel gray values are calculated, and the small value is taken as the adaptive threshold, the interference of factors such as lane lines on the road can be eliminated, so that the average gray value finally used for segmentation threshold calculation can better reflect the road conditions

(4) using the current severe situation of soil pollution prevention and control in China, the double threshold effectively suppresses the noise in the shadow image

statistics of the average execution time of the algorithm using the adaptive threshold of gold system and the adaptive double threshold method proposed by the author are shown in Table 1 (image preprocessing only)

from table 1, it can be seen that using the adaptive double threshold method in this paper for image preprocessing will greatly improve the efficiency of the algorithm

the effect of extracting vehicle shadows using adaptive double threshold method is shown in Figure 3

if only a single threshold is used to binarize the shadow image, the effect is shown in Figure 4

it can be seen from Figure 4 that using a single threshold to binarize the shadow image will retain a lot of noise in the image, which is not conducive to subsequent recognition and processing

if the adaptive threshold is not used, but the fixed threshold is used to binarize the shadow image under strong light conditions, because the selected fixed threshold is suitable for weak light conditions, when the light becomes strong, the shadow information cannot be extracted from the image at all, so the fixed threshold method cannot meet the use requirements under different light conditions. The adaptive double threshold method proposed by the author is used to solve the above problems, which can effectively retain the shadow information of vehicles under various lighting conditions and suppress noise to the greatest extent

in order to obtain a binary image with only horizontal and vertical edges, an adaptive threshold is also used. The segmentation threshold for the binarization of horizontal edge image thori and the segmentation threshold tvert for the binarization of vertical edge image are

respectively, where Chori is the weighting coefficient for calculating the segmentation threshold of horizontal edge; Cvert is the weighting coefficient for calculating the vertical edge segmentation threshold. Chori=2, cvert=1.5

(II) energy density verification for vehicle tracking

after image preprocessing, the image containing horizontal edge information, vertical edge information and shadow information is obtained. On this basis, the tracking module will track and detect the data detected in the previous round to determine whether the target still exists

the tracking algorithm first plans the area of the lower boundary recognition of the current round according to the lower boundary position of the vehicle recognized in the previous round, and recognizes the lower boundary of the target vehicle in this area of the vehicle shadow image. On the basis of completing the lower boundary detection of this round, combined with the recognition results of the previous round of left and right boundaries, the left and right boundaries of this round of vehicles are recognized. In practical applications, the tracking and detection of the left and right boundaries of vehicles are often disturbed by roadside railings, trees, etc., especially when the vehicle in front is driving on the side road, because the vertical edge of the roadside interference is also very obvious, the detection results of the vertical boundary of the vehicle will deviate from the accurate position. The energy density verification of the boundary recognition result is added to the tracking algorithm in this paper, that is, when the energy density (the frequency of shadow points) in a certain area on the right of the left boundary recognition result or in a certain area on the left of the right boundary recognition result is lower than a certain value, the vehicle vertical edge detection is re run according to the confidence of the two boundaries. Define the energy density D as shown in formula (6)

where NTotal is the number of pixels in the area used to verify the energy density (1/4 of the vehicle width is taken in the text), and nshade is the number of shadow points in the area. When D is less than the fixed threshold ddensity, vehicle vertical edge detection will be restarted (ddensity=0.5 in the text)

after adding the energy density verification, when the identification of the vertical boundary of the vehicle deviates from the accurate value, Can be found quickly, he said: "As a human being and corrected, several continuous images using the energy density method to correct the vertical boundary deviation are intercepted, as shown in Figure 5.

the energy density change of the left boundary of the left vehicle in Figure 5 is shown in Table 2.

in the first three figures of Figure 5, due to the interference of the roadside railing, the recognition error of the left boundary of the leftmost vehicle gradually increases. When it is less than the threshold value set by the energy density verification, the vertical edge is detected The program is restarted, and the deviation of left boundary recognition in the fourth image is corrected. After the left and right boundary tracking detection of the vehicle is completed, the upper boundary of the vehicle should be tracked and detected. Finally, the shape verification of the vehicle should be completed to see whether the tracked vehicle contour is within the empirical shape range

(III) variable model used to plan the range of vehicle detection

in order to improve the accuracy of vehicle detection and the speed of the algorithm, the radar detection module is added to the vehicle recognition algorithm. After completing the tracking of the vehicles detected in the previous round, the module is mainly negative

Copyright © 2011 JIN SHI