Traffic Sign Recognition (OpenCV)
Hello, I decided to share my experience of using OpenCV (Intel) in solving a practical problem.
The goal was to follow the next step - automatic recognition of road signs on the video, preferably in real time. Unfortunately, there is very little information in Russian on a similar problem, and even fewer examples.
In general, having fumbled on the Internet, it was decided to use the OpenCV library, which had the necessary mat. device and as it turned out later, quite high performance.
So the stages of solving the problem.
The first stage is getting the image from the camera. When developing, I used a network camera, web-cameras did not have a good enough resolution at hand. OpenCV provides a fairly simple way to work with webcams. (HighGui module)
So with the help of such a code, we constantly get current shots from the camera.
')
The second stage - convert the resulting image. At the first stage, I made recognition of only certain groups of characters. (Forbidding, warning) As a result of some experiments, a filter was obtained, after which we obtain the following image.
The third stage is the application of the threshold filter. OpenCV contains a number of transformations and treatments, I adopted the cvThreshold parameters for this filter selected by practical consideration.
The fourth stage is defining the boundaries of the contours. Again, it was not necessary to think up a bicycle with 3 wheels, a well-optimized function from the cvCanny library - a border detector - was taken. The resulting contour is outlined in the original image.
Thus, I received a list of contours, then, using a previously trained neural network, I classify the resulting contours. If the contour was not related to one previously designated type, it was taken to the trash and not shown.
Such a scheme of work allowed us to create a ready-made application capable of processing somewhere around 3-5 frames per second (depending on the frame load). For the test, a laptop was used - Core 2 Duo T5450, a network camera with a resolution of 1024x768 seems. In tests, the system showed recognition of about 90% of the characters, the only BUT. Sitema will not be able to recognize the sign if it is bent, closed by a billboard and similar things (which unfortunately occurs quite often. I tested it on an ideal stretch of road).
The result was a completely workable application, and most importantly, I got the skills to work with openCV and knowledge in the field of video recognition. I did not find any analogues of this library.
ps at night, this approach is also quite applicable, with some variations.
The goal was to follow the next step - automatic recognition of road signs on the video, preferably in real time. Unfortunately, there is very little information in Russian on a similar problem, and even fewer examples.
In general, having fumbled on the Internet, it was decided to use the OpenCV library, which had the necessary mat. device and as it turned out later, quite high performance.
So the stages of solving the problem.
The first stage is getting the image from the camera. When developing, I used a network camera, web-cameras did not have a good enough resolution at hand. OpenCV provides a fairly simple way to work with webcams. (HighGui module)
cvNamedWindow("web-cam", CV_WINDOW_AUTOSIZE); //
this->capture = cvCreateCameraCapture(0);
cvSetCaptureProperty(this->capture,CV_CAP_PROP_FRAME_WIDTH_HEIGHT,640480); // 640x480
while (1)
{
// -
if(!cvGrabFrame(this->capture)) //
break;
this->frame = cvRetrieveFrame(this->capture); // ( IplImage)
if(!frame)
break;
cvShowImage("web-cam",this->frame); //
cvWaitKey(10);
if(this->fwork == false)
break;
}
So with the help of such a code, we constantly get current shots from the camera.
')
The second stage - convert the resulting image. At the first stage, I made recognition of only certain groups of characters. (Forbidding, warning) As a result of some experiments, a filter was obtained, after which we obtain the following image.
The third stage is the application of the threshold filter. OpenCV contains a number of transformations and treatments, I adopted the cvThreshold parameters for this filter selected by practical consideration.
The fourth stage is defining the boundaries of the contours. Again, it was not necessary to think up a bicycle with 3 wheels, a well-optimized function from the cvCanny library - a border detector - was taken. The resulting contour is outlined in the original image.
Thus, I received a list of contours, then, using a previously trained neural network, I classify the resulting contours. If the contour was not related to one previously designated type, it was taken to the trash and not shown.
Such a scheme of work allowed us to create a ready-made application capable of processing somewhere around 3-5 frames per second (depending on the frame load). For the test, a laptop was used - Core 2 Duo T5450, a network camera with a resolution of 1024x768 seems. In tests, the system showed recognition of about 90% of the characters, the only BUT. Sitema will not be able to recognize the sign if it is bent, closed by a billboard and similar things (which unfortunately occurs quite often. I tested it on an ideal stretch of road).
The result was a completely workable application, and most importantly, I got the skills to work with openCV and knowledge in the field of video recognition. I did not find any analogues of this library.
ps at night, this approach is also quite applicable, with some variations.
Source: https://habr.com/ru/post/61048/
All Articles