MACHINE LEARNING MODEL FOR TRAFFIC SIGN RECOGNITION

Abstract

Autonomous (driverless) cars are increasingly becoming popular, hence calling for robust Traffic Sign Recognition (TSR) systems to ensure road safety. A report by World Health Organization Road Safety Report of 2018, shows that failure to distinguish and recognize traffic signs is among the leading causes of accidents. Existing TSR systems are adversely affected by environmental conditions, partial occlusion of traffic sign, illumination, colour deterioration because of their exposure to different rays including Ultra-Violet (UV), physical deformation, variations in pictogram designs and weather conditions among others. The study was guided by the following main objective; to develop a robust model using machine learning for recognition of traffic signs. Deductive research approaches, was used to achieve the following specific objectives: to analyses existing techniques in TSR, to design an advanced feature extraction technique for robust TSR and develop a machine learning model for recognition of traffic signs. The new Local Directional Histogram of Oriented Gradient (LD-HOG) feature extractor is the main contribution of this work. A more resilient and discriminative descriptor is produced by combining the directional resilience and noise invariance of the Local Directional Pattern (LDP) with the potent gradient magnitude representation of HOG. The German Traffic Sign Detection Benchmark (GTSRB) dataset, was used to extract features. With an average F1-score of 96.5% using an SVM classifier, LD-HOG outperformed HOG by 4.2% and LDP by 7.8%. By helping to create more accurate and dependable advanced driver-assistance systems, the study will benefit a variety of stakeholders and road users, including drivers, passengers, and legislators., and policy makers.