Vision-Based Turn Signal Recognition for Automated Flagging

Flaggers work closely to open traffic lanes which makes them a high-risk profession. They might get hit by distracted, speeding, or intoxicated drivers, leading to severe or even fatal injuries. To address this safety issue, the concept of Automated Flagging System Devices (AFSDs) was proposed. In an AFSD, one of the essential components is to recognize the turn signals of the vehicles in the lanes. However, most existing studies for vehicle turn signal recognition focused on autonomous driving to help autonomous vehicles interpret the intents of the vehicles in front. Little work has been conducted for AFSDs to guide traffic. This paper presents a novel method for recognizing vehicle turn signals using the video camera on an AFSD at the roadside. The method first relies on object detection and tracking models to locate the vehicles and headlights in the video frames. Then, the detected headlights are matched to the vehicles. For each vehicle, the luminance features of its headlights are further extracted. Based on these luminance features over frames, the flashing patterns of the headlights could be recognized to determine the vehicle turn signal. The method was tested with real traffic videos, showing that its overall recognition accuracy could reach 80.32%.