A Bangladeshi Road Sign Detection & Recognition System Based on Template matching & Convolutional Neural Network

Abstract

Road Signs are very important but a neglected topic in the discipline of driving vehicles in Bangladesh. As a highly populous country chaos on the road has taken a disastrous shape over the past decade as the number of vehicles on the road had went up and road safety became a crying need as the lives lost in road accidents are constantly on the rise. In case of Road Signs in Bangladesh the descriptive text for it is given more emphasis than the sign itself. Whereas, in the developed countries the signs are given more importance, because Signs are supposed to be universally accepted in case of communication. Hence, the Road Signs in Bangladesh are put quite arbitrarily, especially in the metropolitan areas the signs are credited by Metropolitan Police, or sponsored by companies, they look more like posters or banners than a legitimate Road Sign and it becomes very difficult to locate them. As we are witnessing the development of Autonomous Driving Systems all over the world by companies like Tesla, it is becoming more and more popular in demand. The detection & recognition of Road Signs play a crucial role in it. In Bangladesh Road Signs are ignored most of the time & text based road signs are given the priority. While human brain is capable enough to detect text based Road Signs, a smart system can struggle. Nonetheless, two types of measures could be taken into account in this case, one is through text recognition based and the other is detecting the sign in the poster-like signs and then run it through a neural network based classification to identify it. The later solution was taken into account for this paper, as all the Road Signs (both poster-like & basics) contain the basic signs for directions, they may or may not contain descriptive text along with them, but the road signs are always present in larger or smaller scales. The solution was built around Deep Learning and Image Processing techniques for object detection, segmentation and classification. Supervised Learning and Transfer Learning were used in terms of technical development. Text based recognition systems are not ideal for metropolitan areas because there are so many posters and banners attached to the poles and around the Road Signs.