Adaptive Tracking: Dynamic Frame Rate for Drone-Based Target Following

In this work, we focus on enhancing target tracking and automatic following using a drone. This task involves a continuous capture-analyze-capture loop, where the analytics results directly influence the drone’s navigation, and must address temporary target loss to ensure reliable following. Current industry methods typically require processing every frame captured by the drone, leading to significant computational overhead. We propose AdaTrack (Adaptive Tracking), an adaptive frame rate approach for drone-based target following. AdaTrack dynamically adjusts the frame rate: reducing it when tracking is straightforward and increasing it when the target is lost or misidentified-indicated by the similarity between consecutive bounding boxes-to quickly re-locate the target when it reappears. Additionally, we implement a control mechanism using multiple-armed bandit (MAB) to generate control commands based on the tracking results. We validate the effectiveness of AdaTrack through field experiments, demonstrating the drone’s ability to follow different targets, such as cars and cyclists, while requiring only a low processing frame rate.