Introduction
In aerial surveillance and search-and-rescue operations, occlusion remains a significant hurdle for traditional computer vision systems. A target vehicle may be clearly visible from one angle but entirely obscured by a tree or building from another. While standard "passive" object detection attempts to interpret whatever static image it receives, Active Object Detection (AOD) empowers the observer—in this case, an Unmanned Aerial Vehicle (UAV)—to act as an intelligent agent. Instead of remaining static, the drone autonomously determines if moving to a new viewpoint will yield a better observation.
Despite its potential, AOD research has historically focused on indoor robotics. The lack of high-quality, large-scale datasets specifically designed for the "eye-in-the-sky" perspective has hindered the development of Deep Reinforcement Learning (DRL) policies for outdoor UAV navigation. This paper introduces UEVAVD, a dedicated dataset for UAV active vision, and IBE-MAP, a novel reinforcement learning method designed to navigate complex environments efficiently.
The IBE-MAP Framework
The researchers propose IBE-MAP (Inductive Biases Enhanced Multistep Action Prediction), which incorporates two critical inductive biases to improve learning speed and performance:
- Spatial Bias (Scene Pre-decomposition): To reduce visual noise, the system utilizes the Segment Anything Model (SAM) to mask out irrelevant background elements like distant terrain. By focusing strictly on the target and its immediate obstacles, the agent generalizes more effectively to unfamiliar environments.
- Partial Observability Bias (Temporal Memory): Since a single frame rarely provides a complete picture of a complex scene, the model employs a Gated Recurrent Unit (GRU). This allows the drone to aggregate a sequence of observations, "remembering" past views to inform its next movement decision.
The decision-making core is powered by a Dueling Deep Q-learning Network (DuelingDQN), which independently predicts the optimal direction of movement and the necessary magnitude (e.g., moving two units to the left).
The UEVAVD Dataset
The UEVAVD (UAV’s Eye View Active Vision Dataset) was developed using Unreal Engine and the AirSim plugin to simulate high-fidelity urban and woodland environments. Key features of the dataset include:
- Standardized Targets: Five distinct vehicle types—Hatchback, Pickup, Sportscar, Sedan, and SUV—are used with consistent textures to ensure the model learns geometry and shape rather than color patterns.
- Scale and Diversity: The dataset comprises 90,750 images across 150 unique contexts, including variations in lighting, background, and occlusion levels.
- Continuous Simulation: Unlike passive datasets like VisDrone, UEVAVD provides densely sampled multi-view images, enabling the simulation of continuous, autonomous flight paths.
Key Results and Insights
Experimental results demonstrate that IBE-MAP significantly outperforms standard baseline models, particularly in Zero-Shot Generalization (ZSG). The method showed the smallest "GenGap" (the performance delta between training and testing environments), proving that scene decomposition helps the agent ignore environmental distractions and focus on the spatial relationship between targets and obstacles.
Furthermore, the study highlighted a critical trade-off between accuracy and efficiency. By setting a classification confidence threshold (found to be optimal at 0.2), the agent can stop moving once a view is "sufficient," thereby minimizing flight time and battery consumption. In qualitative tests, IBE-MAP consistently guided UAVs to recognizable views using fewer steps and more direct paths than previous methods.
Conclusion and Future Work
UEVAVD provides a foundational resource for the burgeoning field of UAV active vision. While the current results in simulation are promising, future work will focus on bridging the "simulation-to-real" gap by accounting for real-world sensor noise and randomized target orientations. By treating the drone as an active participant in the detection process, IBE-MAP moves us closer to truly autonomous aerial surveillance systems.
Summary updated