Unified approach for detecting traffic signs and potholes on Indian roads

Abstract

The Driver Alerting System relies on automatic identification and recognition of traffic signs and potholes. According to recent research, no single approach addresses both pothole detection and traffic sign recognition on roads. The majority of research on pothole detection and traffic sign recognition is based on deep learning techniques such as Convolutional Neural Network, Long Short Time Memory, and others. Moreover, most of the works concentrate on foreign roads where the road condition will be much more different than the Indian roads. In this work, a unified model for recognizing the traffic signs and potholes on Indian Roads is developed. The optimum features related to road traffic signs are extracted and matched using the Hybrid Features From Accelerated Segment Test and Random Sample Consensus algorithms. Features from accelerated segment test is a corner detection technique with high computational efficiency, and the random sample consensus algorithm is applied to discard the mismatching points. To detect potholes, the improved Canny Edge detector and bio-inspired Contour detection method are used. Finally, the Support Vector Machine classifier is used to classify the potholes and traffic signs. The discovered potholes' sizes are then calculated using the bounding box regression model. According to experimental results, the suggested unified model leave behind existing models in terms of accuracy, sensitivity, specificity, Matthews correlation coefficient, and F1-Score values.

Keywords

• Pothole detection
• Traffic sign detection
• FAST-RANSAC
• Support vector machine
• Pothole size
• Bio-inspired contour detection