A Frequency-Domain Model for Detecting Fleeting Objects in Aerial Drone Imagery
This article introduces DroneWorldNet, a high-throughput signal-processing model designed to detect transient and stable objects in aerial drone imagery with inference latency. DroneWorldNet integrates discrete wavelet decomposition and a finite-embedding Fourier transform (FEFT) to extract frequency-domain features from image clip vectors, enabling robust classification of fleeting phenomena such as pedestrians, vehicles, drones, and other mobile entities. By leveraging the parallelism of modern GPUs, DroneWorldNet achieves real-time performance, making it suitable for deployment in edge-cloud architectures supporting autonomous surveillance, urban mobility, and disaster response.
We apply DroneWorldNet to the Dataset for Object Detection in Aerial Images (DOTA), a large-scale benchmark comprising thousands of annotated aerial scenes captured by UAVs across diverse environments. Each image clip is treated as a temporal stack of observations, where spatial and motion cues are embedded across frames. These clips are transformed into frequency-domain tensors using a combination of one-dimensional wavelet decomposition and FEFT, capturing both localized spatial features and global periodicity. This dual representation allows the model to detect both persistent and ephemeral objects, even under conditions of occlusion, low resolution, or irregular sampling.
The DroneWorldNet pipeline begins with spatial clustering of image patches using a density-based approach akin to DBSCAN, grouping temporally adjacent frames into coherent sequences. These sequences are preprocessed to normalize brightness, contrast, and motion blur, and then encoded into tensors that reflect the temporal evolution of each scene. The wavelet decomposition suppresses noise and highlights localized changes, while FEFT extracts periodic and harmonic structures that may indicate transit-like behavior or repetitive motion. These tensors are then passed through a convolutional neural network (CNN) with fully connected layers, which outputs one of four predictions: null (no object), transient (brief appearance), stable (persistent presence), or transit (periodic occlusion or movement).
To train DroneWorldNet, we simulate synthetic aerial sequences using generative models that replicate pedestrian and vehicle motion under varying lighting, altitude, and occlusion conditions. These synthetic clips are augmented with real DOTA like annotations to ensure generalization across urban scenes.
This methodology showcases the potential of frequency-domain analysis for aerial object detection, offering a scalable alternative to frame-by-frame tracking or phase-folding methods, which are often computationally prohibitive at scale. DroneWorldNet’s architecture is modular and adaptable: it can be retrained as a binary classifier for specific object types (e.g., emergency vehicles), or extended to regression tasks such as trajectory estimation or velocity prediction. Its ability to handle irregular sampling and variable sequence lengths makes it particularly well-suited for UAV deployments where cadence and resolution fluctuate due to flight dynamics or environmental constraints.
DroneWorldNet demonstrates that frequency-domain representations—when combined with deep learning—can effectively detect and classify fleeting objects in aerial imagery. This approach opens new avenues for time-domain analysis in geospatial workflows, enabling rapid anomaly detection, traffic monitoring, and situational awareness in complex environments. Future work will explore integration with onboard sensors and real-time feedback loops, extending DroneWorldNet’s capabilities to active tracking and autonomous decision-making in aerial platforms.
