TSCTNet: Traffic Signal and Countdown Timer Detection Network for Autonomous Vehicles

Abstract

The revolution in the field of autonomous vehicles and driver assistance systems has brought with it the need for the development of robust support systems such as traffic signal and countdown timer detection in order to facilitate faster and safer real-world adoption. Various deep learning models and computer vision techniques have extensively been applied to develop systems for the detection of traffic signals. However, the task of detecting an auxiliary timer value along with a traffic signal has relatively been unexplored. In this article, a novel framework, TSCTNet (Traffic Signal and Countdown Timer Detection Network), is proposed that leverages state-of-the-art Mask R-CNN model for object detection followed by efficient image processing techniques for detecting a traffic signal. Subsequently, the detected signals are used to extract a local subimage, which is then processed and fed to a RetinaNet model for detecting the countdown timer value. The framework has been evaluated on a custom dataset (CFTS) of dash-cam videos and achieves a high average precision value for detecting the traffic signal as well as the countdown timer value. The proposed system was found to be performant in various scenarios and can assist in paving the way for the advancement of autonomous vehicle infrastructure.