A lightweight multi-task deep learning framework for uav detection and tracking

Abstract

In this work, we propose a custom Multi-Task Learning (MTL) model for real-time UAV detection and tracking, designed to jointly perform classification and bounding box regression. The proposed model was evaluated against state-of-the-art detectors, including YOLOv8 and Faster R-CNN ResNet-50. While YOLOv8 achieved fast inference with strong accuracy, and Faster R-CNN demonstrated high precision in complex scenes, our MTL model outperformed both in classification accuracy and bounding box precision. Specifically, the MTL model achieved a classification accuracy of 98.53%, a bounding box MAE of 0.0256, and an MSE of 0.0027, demonstrating its effectiveness in multi-output learning. To enable tracking, we integrated a Kalman Filter, which maintained consistent ob- ject identities across frames . These results highlight the robustness and efficiency of the proposed MTL-based pipeline for UAV detection and tracking in real-time surveillance applications.