Detection, localization and distance measurement of traffic sign plates with deep learning

Traffic accidents are responsible for millions of life losses and countless financial impacts. Despite the advancement in the automobile industry, the drivers remain ultimately responsible for the vast majority of accidents. Automated driver assistance systems have been researched for more than two decades to improve the road experience and save the lives of both drivers and pedestrians. With all possible technologies in hand, these systems are still very much preliminary if compared to similar classes. In recent years, an extensive research effort has been dedicated to Traffic Sign Detection (TSD) systems, as traffic signs compose an essential road use guide, and assist in avoiding dangerous situations. For these systems to be useful, they should operate in an online fashion, which requires an affordable computational complexity as well as enhanced image detection. One of the possible solutions is the use of deep learning to maintain high-quality detection. Training deep learning machines is neither fast nor are they adequate on their own for online processing. Yet, they are a potential candidate for a very efficient image detection for TSD systems. To benefit from the computational power of deep learning machines, and in the meantime, improve their training time complexity, this research proposes the use of Principal Component Analysis (PCA) and Faster R-CNN to detect, recognize traffic signs, as well as estimating the distance between them and the driver. PCA is used as a blind-source separation mechanism to efficiently denoise the processed images at a low cost. While Faster R-CNN maintains a very efficient, and computationally less complex, deep learning mechanism applicable for TSD. With two datasets acquired from different databases, the performance of the proposed method is evaluated, and an accuracy of 0.99 was obtained. Besides, left and right traffic signs could also be detected with almost the same accuracy.