Computer VisionDeep LearningMachine LearningProject

Developing a Self-driving Car – Day #7 & #8

The problem regarding a faster region proposal seems to be solved. The SVM + EA combination did not work fast enough. So I decided quite early to try solving my problem on an other way. I took a look into a bunch of quite old papers, where people tried to detect and recognize traffic signs by shape and color information. Many of these papers aimed for reducing wrong detection as good as possible. However this of course means a lot of computitional overhead for our NN. So I decided to go for a really simple color based thresholding. It works surprisingly great and runs in ~20ms for a 1280×720 image. This is even faster than I thought (I did not use any kind of multithreading, even though this would be quite straight forward to add). Basically the algorithm searches for redish, blueish, greenish and yellowish pixel colors. I used a cluster algorithm (DBSCAN) to merge these pixels into areas. After clustering, I apply Non-Maximum-Suppression on bounding boxes of the clustered areas to remove as much useless bounding boxes as possible. I also remove every bounding box which is smaller than 15px in width or 15px in height or which is bigger than 150px in width or 150px in height. I also removed all bounding boxes which occur at the very bottom of the image (so if boundingbox.y > 5/6*image_height then I remove this box). The very basic algorithm was inspired by Broggi et al. 1

These are some results, where white rectangles denote ROIs.

Before After

The exact algorithm for extracting redish, etc. pixel looks like this (using OpenCV 3 and C++):

But one big problem remains: Recognizing/finding greyish/whitish signs is troublesome. This is quite hard, because clouds, the street, etc. are greyish/whitish, too. So this will be the next thing to do!

See ya!

Broggi A, Cerri P, Medici P, Porta P, Ghisio G. Real time road signs recognition. In: Intelligent Vehicles Symposium, 2007 IEEE (p. 981-986); 2007.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.