Segment-based UAV Tracking via Shape-Aware Kalman Filter

• Young Woong Jun,
• Sang-Chul Kim,

Vol. 20, No. 3, March 31, 2026
10.3837/tiis.2026.03.016, Download Paper (Free):

• Deep Learning
• UAV Segment Tracking
• Kalman filter
• instance segmentation
• Multi-modal

Abstract

This study proposes a segmentation-based framework for real-time unmanned aerial vehicle (UAV) tracking in complex visual environments. Conventional tracking approaches rely on bounding boxes and centroid-based Kalman filters, which are typically ineffective under occlusion, small object scale, and nonlinear UAV motion. In contrast, the proposed method uses pixel-wise segmentation masks to achieve precise object localization and integrates a shape-aware Kalman filter that models both the motion and geometric properties of UAVs. The state vector is extended to include the area, aspect ratio, and orientation, thereby enabling robust identity preservation during rapid movement and partial visibility. Fast-SCNN is used as a lightweight segmentation backbone to balance real-time inference and spatial accuracy. We benchmark the proposed framework on the Anti-UAV dataset, demonstrating significant reductions in identity switches by up to 84% while maintaining competitive overall tracking performance (multiple object tracking accuracy (MOTA) improvement of 3.7%) compared to state-of-the-art detection-based trackers (YOLOv3, Faster R-CNN, and Tracktor). The proposed framework has strong potential for UAV monitoring applications requiring fine-grained object representation, stable identity tracking, and computational efficiency in real-world scenarios.

Statistics

Cite this article