Certain investigations on recognition of traffic signs using machine learning techniques

Abstract

In India, road transport is the major mode of transportation of goods and passengers. During the past three decades, many advanced technologies including Intelligent Transport System (ITS) have been developed and adapted for improving road safety and pollution control. However, the challenges associated with the traffic sign detection using ITS are still attracting many researches to address the complexity in capturing and processing of the signs especially in low light/ night time. In this work, various techniques have been proposed for enhancing the performance of recognition rate through the efficient methods of shape model extraction, segmentation and feature extraction. The major objective of the work is attained through the following specific objectives (i) implementation of Active Appearance Model (AAM) for enabling good shape detection rate, (ii) extracting correct keypoints through the introduction of Harris detector, Maximally Stable Extremal Regions (MSER) detector, and dense detector, and (iii) enhancing the accuracy of the Traffic Sign Detection (TSD) with lower computational time by using MTANN, ATDANN and CVLVQNN. In the first work, shape model generation using Active Appearance Model (AAM) is done. Next, the features at some key points are extracted in the train images and converted into feature descriptors, which are high dimensional vectors. The features are extracted at some key points, which are obtained using the Harris-Laplace salient point detector. It uses a Harris corner detector and subsequently the Laplace operator for scale selection. The edge is detected using Sobel operator. After the detection, the segmentation is applied using the proposed Adaptive Fuzzy Clustering (AFC) technique. Finally, traffic sign detection is performed via the use of Massive Training Artificial Neural Network (MTANN). In the second work, maximally stable extremal region based on the concept of thresholding for an efficient extraction of image key point under feature extraction process after shape conversion using Active Appearance Model (AAM). Those images are segmented by using Fuzzy Hidden Markov Model (FHMM) using Bayesian network approach for detection of traffic sign. Final stage is classification, which is done by using Adaboost Training based Deep Artificial Neural Network (ATDANN) to improve the performance and reduce the error rate during recognition of traffic sign. newline