Vision-based detection and tracking of vehicles to the rear with perspective correction in low-light conditions

In this study, the authors present a video-processing system that utilises a camera to detect and track vehicles to the rear at night. Vehicle detection is a pivotal component of camera-based advanced driver assistance systems (ADAS) such as collision warning, blind-spot monitoring and overtaking vehicle detection. When driving in dark conditions, vehicles to the rear are primarily visible by their headlamps. This system implements a low, static camera exposure to ensure headlamps appear distinct and not as enlarged bloomed regions. We further describe a method to identify vehicle headlamp pairs using region-growing thresholding and cross-correlation bilateral symmetry analysis. Images of vehicles at different yaw angles to the camera image plane, such as those turning, or engaging road bends, suffer from perspective distortion, which interferes with the symmetry between lamps. We perform a perspective transformation to correct for this, ensuring consistent detection performance throughout all road manoeuvres. False positives resulting from multiple vehicle situations are considered and addressed. Detected target vehicles are tracked using a Kalman filter which is updated by inter-frame cross-correlation.