Motion-aware adaptive fragmentation and transformer-based temporal modeling for lane-change recognition in Unmanned Aerial Vehicle traffic videos

Abstract

Lane-change recognition in traffic videos captured by unmanned aerial vehicles is challenging due to small vehicle size, global camera motion, and the computational cost of dense video processing. We propose a motion-aware adaptive fragmentation framework that combines homography-based camera-motion compensation, optical-flow-guided motion-saliency estimation, and sparse spatiotemporal Transformer modeling. Motion-salient regions are represented using fine-grained patches, whereas regions with lower motion activity are processed using larger patches to reduce redundant spatial representation. The framework was evaluated on a lane-change dataset derived from tracking annotations of the Unmanned Aerial Vehicle Benchmark using an automatic trajectory-based label-generation procedure. The resulting dataset contains four hundred and one thousand three hundred  forty-five eight-frame temporal windows extracted from thirty aerial video sequences and one thousand three hundred sixteen vehicle tracks, including fifty-four thousand seven hundred and seventy-five lane-change and three hundred forty-six thousand, five hundred seventy normal-driving samples. Experimental results demonstrate an accuracy of zero point seven nine, a binary F1-score of zero point seven seven, an area under the receiver operating characteristic curve of zero point nine three, and an area under the precision-recall curve of zero point nine one. The proposed adaptive representation limits the number of spatial tokens to a maximum of sixty-four per vehicle-centered frame, corresponding to a seventy-five percent reduction in the maximum spatial token budget relative to the two hundred fifty-six-token dense uniform-patch tubelet baseline. The results indicate that motion-aware adaptive fragmentation can preserve competitive lane-change recognition performance while substantially reducing the spatial token sequence processed by the Transformer. This provides a promising basis for computationally efficient aerial traffic monitoring, while complete assessment of edge-deployment efficiency requires future end-to-end measurements that also account for homography estimation and optical-flow computation.