Early Detection of Incoming Traffic for Automatic Traffic Light Signaling during Roadblock using Vanishing Point-Guided Object Detection and Tracking

Traffic controller manages traffic flow during roadblock. However, the traffic controller has a risk of traffic accidents and tends to be exposed to weather conditions such as extreme sunlight, rain, and snow. Automatic traffic lights equipped with video cameras and connected to a computer become an alternative for traffic control. However, perspective projection influences the captured scene from the camera. It makes incoming traffic that is far from the camera looks small. Early detection of this incoming traffic is important because raising the stop signal when the traffic is already too close to the traffic light may lead to an accident. Deploying deep learning in detecting this incoming traffic deals with a problem caused by the small resolution of the vehicle image. Therefore, we proposed vanishing point-guided object detection and tracking to detect the incoming traffic as early as possible. We combined background subtraction for object detection and the Kalman filter for motion estimation. In addition, the detection and tracking are guided by the vanishing point position obtained from the Region of Interest. We evaluated the proposed method's performance to Regions with Convolutional Neural Networks object detector. Performance results show that the proposed method can detect the incoming traffic earlier than detection using R-CNN. Therefore, detection from the proposed method can support a better judgment for an automatic traffic light.