Histogram of Oriented Gradients Based Reduced Feature for Traffic Sign Recognition

Abstract

Vehicles equipped with automatic traffic sign recognition (TSR) capability has recently attained a lot of attention among computer vision community. There has been a need for a robust TSR which can provide the current state of traffic signs on the road which will in turn assists drivers, autonomous vehicles, and mobile robots. An efficient and effective TSR could facilitate drivers in taking necessary decisions in advance which will ultimately reduce many fatal road accidents and deaths which takes place due to lack of road sign awareness among the drivers. Although there has been a lot of research in this direction, the design of a robust and an efficient TSR system is still an unsolved problem. Therefore, in this work, we have proposed an efficient method of TSR using Histogram of Oriented Gradients (HOG). We have tried to exploit the usefulness of HOG features for TSR by selecting a proper set of parameters which results in generating more discriminative features. Further, dimensionality reduction using Principal Component Analysis (PCA) has also been carried out to reduce the number of redundant features. Finally, the reduced features have been classified using a Kernel Extreme Learning Machine (K-ELM) classifier. Numerous experiments have been performed using GTSRB traffic sign recognition benchmark dataset and results show the effectiveness of the proposed approach.