Real-time vehicle detection with a single camera using shadow segmentation and temporal verification

Vehicle detection with a monocular video sensor offers significant potential in a wide area of promising automotive applications. In this work the shadows in daylight scenes are used to detect vehicles in low-quality camera images and determine their positions and velocities in relation to the own vehicle. We propose a new algorithm that segments shadows in a top-down-view. In contrast to other works no specific thresholds or strong shadow edges are needed. Instead we binarize the image at multiple thresholds and analyze the results to generate vehicle position hypotheses. These hypotheses are verified with a RANSAC-based temporal verification approach. The accuracy of our method is evaluated using monocular real world sequences and ground truth data. Our algorithm works robustly under various lighting conditions in real-time at full video frame rate.