- The paper introduces a global-to-local detection approach that combines motion and appearance features to effectively detect MAVs in complex environments.
- It employs a Kalman filter-based adaptive search region to refine detections and handle occlusions and small target sizes.
- Experiments on the novel ARD-MAV dataset and other benchmarks demonstrate superior accuracy and computational efficiency for onboard UAV systems.
Global-Local MAV Detection for Challenging Conditions
The paper, "Global-Local MAV Detection under Challenging Conditions based on Appearance and Motion," addresses the problem of micro aerial vehicle (MAV) detection using visual inputs under challenging environments. This task is crucial for applications such as collaborative UAV operations and security systems. The authors propose a global-local detection approach that fuses motion and appearance cues, aiming to enhance detection accuracy and efficiency while remaining computationally feasible for onboard UAV systems.
In the context of existing challenges, traditional detection methods based solely on appearance or motion features have limitations, especially in scenarios with complex backgrounds, small target sizes, and limited computational resources. The proposed method consists of a global detector to perform an initial search and a local detector that refines detection within an adaptive search region. The global detector integrates a motion-based detection module that is triggered when appearance features are insufficient. This combination allows the system to maintain high detection rates even under difficult conditions.
A novel dataset, ARD-MAV, was created to train and verify the effectiveness of the proposed system. It contains diverse scenarios such as complex backgrounds and occlusions, with significantly smaller average target sizes compared to prior datasets. The experiments, conducted on ARD-MAV and other challenging datasets like NPS-Drones and Drone-vs-Bird, demonstrate that the proposed method outperforms state-of-the-art solutions in both detection accuracy and computational efficiency. The system achieves near real-time frame rates on platforms like NVIDIA Jetson NX Xavier, underscoring its practical applicability.
Key contributions of the paper include:
- The development of a global-to-local detection method that significantly enhances detection accuracy and efficiency for MAVs under challenging conditions.
- A robust design of motion-based and appearance-based classifiers capable of filtering out alignment inaccuracies in non-planar scenes.
- The introduction of a Kalman filter-based adaptive search region to bolster detection robustness against occlusions and detection misses.
- The creation of the ARD-MAV dataset, which pushes forward the benchmark for MAV detection in terms of difficulty, encompassing scenarios with the smallest average object sizes in the field.
The implications of this research are multifaceted. Practically, it provides a viable MAV detection solution for real-world deployments on aerial vehicles, potentially enhancing autonomous UAV operations and aerial surveillance systems. Theoretically, it underscores the synergistic potential of combining motion and appearance features in visual detection tasks and highlights the importance of adaptive region strategies in dynamic tracking environments. As UAV technologies and their applications expand, future developments could integrate more advanced feature fusion techniques and optimization methods to further refine detection capabilities and efficiency. This work sets a foundation for robust UAV detection systems capable of operating effectively in a variety of challenging environments.