Unsupervised video leap segmentation for fast detection of salient segment transformations in mobile sequences

Multiple-frame segmentation, also referred to as video segmentation, is an important step in many video analysis applications for identifying and tracking specific features as they move through a scene. In a mobile, resource-constrained environment such as an intelligent vehicle system, video segmentation can be utilized in preprocessing to reduce image information and increase processing efficiency for high-level scene understanding applications. We introduce video leap segmentation, a highly efficient multiple-frame segmentation approach for use on embedded and mobile platforms where processing speed is critical. The proposed method is demonstrated to successfully track segments across spatial and temporal bounds, generating fast, stable segmentations of images from captured moving-camera video sequences. Video leap segmentation is applied to the task of rough salient segment transformation detection for alerting potential drivers of critical scene changes that may affect steering decisions. Trial results demonstrate that with little added computation, video leap segmentation can be utilized for salient region detection in traffic scenes with high accuracy, correctly detecting 80% of salient segment transformations in trial scenes with less than 5% false positives. Reducing high-level processing to salient areas using the proposed approach has the potential to significantly improve the processing efficiency of scene interpretation applications in intelligent vehicle systems.