YOLO-MOTF: Motion-temporal fusion for dynamic object detection with a moving camera for assistive wheelchairs

Abstract

Dynamic object detection is fundamental to advancing vision-based navigation systems, particularly in environments where the camera itself is in motion. Despite progress in detection algorithms, existing approaches often struggle with challenges such as egomotion, short-term occlusions, temporal discontinuities, and computational cost. This paper presents YOLO-MOTF, a novel knowledge-based model that integrates spatial features with motion cues, especially for operation under moving camera conditions. The framework incorporates a hybrid motion compensation strategy to suppress camera-induced distortions and an occlusion handling buffer to preserve object trajectories through discontinuities. Additionally, a motion attention gating mechanism selectively reinforces moving object predictions by intersecting fused motion masks with semantic outputs. The proposed system achieves an F1 score of 88.6% and a 93% reduction in flow processing compared to dense flow methods, underscoring its robustness and efficiency in dynamic environments. Beyond theoretical contributions, the model demonstrates direct applicability in real-world knowledge-based decision systems, including healthcare applications such as assistive wheelchair navigation, as well as assistive robotics, autonomous navigation, and surveillance.