Traffic Sign detection w/o deep learning algortihms (low-medium level image process).
The desired program’s purpose is the detection of the traffic signs in several images which are obtained under different circumstances. Such as different lightning, angle, and resolution etc. But the program cannot use a high-level image process or deep learning algorithms. It must be with low and mid-level image process.
First step is data featuring. This means that I analyzed the traffic signs that program will detect, and I have identified a few distinctive features. After getting these features I had the queuing up these features for the best result.
I took this program in two phases, which are shape and color detection. The main reason why I choose these two features is these two features is the best distinctive features for detecting traffic signs. But while coding, I realized filtering by area is much more effective than filtering by shape because filtering by shape causes much more false positive and false negative results. Of course, filtering by area is also causes false detection but not as much as filtering by shape. And, with filtering by area can detect some small traffic signs which shape detection could not detect. Finally, I would like to point out that filtering by area comes after filtering by color in this program. In details,
First step of this algorithm is calling 2 different color detection method. In these methods first steps is, changing the color model of the input image from BGR to HSV. Then initializing lower and upper boundaries for color detection with respect of HSV ranges in OpenCV. While blue color detection using one boundary, the red color detection is using two different boundaries, because of OpenCV range. After initializing the boundaries, we are creating a mask. In color detection we usually use bitwise method to detect color in our original image but in this situation, I did not use the bitwise method because my main purpose is not detecting colors. I used these masks to detect shapes by area.
The filtering by area method gets a parameter to get contours and filter by area. So, these masks will be the parameter of this method. In this method I used some built-in functions such as arclength, contourArea, approxPolyDP etc. Before explaining these methods, I also want to point out that I took this method in two phases too. While I was using only one if condition, I realized I could not detect small signs in an image. So, I used two if conditions to handle that problem.
To detect smaller signs the second if condition is runs. But there is a huge difference between these two conditions, in the second conditions I used a morphological operation to make bigger that small signs. And of course, if I wanted to make bigger that contours there are two morphological operations, which are closing and dilation. I preferred to use dilation instead of closing because closing operations dilation is not enough to detect these small signs. So, I used dilation to make these contours to detection.
Difference between normal contour and dilated contour. With this method program can detect that small signs easily.
Now I can detect some small signs but there is a trade-off, while I can detect small signs there might be also more false positives. Finally, after these if conditions I used a method that is like recursive method. I initialized an another for loop which has same parameters with the first one, but the difference is before these loops that I mentioned above. After this method, the final phase is calling these detection methods to detect traffic signs.
The headlight of car is detected as a traffic sign from program which we can say that is false-positive. And also I would like to say that as I mentioned above there is two conditions in filter by area method and if sign detected with first condition the boundary is green and if sign detected with second condition the boundary is purple. Hence, we can see that trade off in this example.
Final step of this program is that reading an image, then calling two different color detection methods after combining the returned masks send this combined mask to filter by area method.
As an example. In this image we can say the two signs that in front is detected with bigger area condition (first condition) and the behind ones are detected with smaller area condition (second condition) which also use dilation morphological operation.
To understand the effectiveness of this program we can use F1 score calculation, which includes confusion matrix elements such as false-positive, true-negative etc.
To calculate that F1 score we have to had %100 results. And we can get this result with manually detect these signs as human.
And the formula of the F1 Score is: 2 x (precision x recall) / (precision + recall) So, the final calculation is that:
Precision: 0,866 Recall: 0,866 F1-Score: 0.86599