Intelligent Highway Lane Center Identification from Surveillance Camera Video

Surveillance cameras are commonly used along US highways by State Departments of Transportation (DOT) to monitor the traffic status. State-of-the-art automatic traffic status monitoring requires that the cameras be preset at a predefined zoom level and viewing direction. However, in deployed systems these cameras have zoom, pan, and tilt features. The DOT operators will change the zoom level and camera viewing direction to gather traffic information for analysis. Therefore, the road and lane locations on the camera image cannot be at preset values. It is desirable to develop an intelligent system to automatically detect road and lanes using highway surveillance video with any zoom level and viewing direction. This paper describes a novel lane identification framework based on the Deep Learning starting point: YOLOv4 vehicle detection. This new technique does not use any painted lane markings. We identify the vehicles on the road and then aggregate the detected vehicles at one horizontal line on the image. This salient horizontal line is where the YOLOv4 gives the highest vehicle detection confidence scores. Most vehicles stay within a lane most of the time and change lanes only occasionally. This assumption is used to find the lane center locations according to the relative number of vehicles passing the salient horizontal line. Our method provides robust results in all weather conditions with a lane detection F1-Score above 0.85 compared to human labelled ground truth.