Is That Rain? Understanding Effects on Visual Odometry Performance for Autonomous UAVs and Efficient DNN-based Rain Classification at the Edge

Andrea Albanese, Yanran Wang, Davide Brunelli, David Boyle

Abstract

Autonomous Unmanned Aerial Vehicles (UAVs) increasingly rely on Visual Odometry (VO) for navigation; however, adverse weather conditions like rain significantly degrade sensor reliability. This paper quantifies the impact of various rain intensities and angles on VO performance and proposes a lightweight, deep-learning-based classification framework for real-time weather awareness. By introducing a comprehensive dataset of 335,000 images, the study demonstrates that specific rain patterns—particularly slanting rain—can cause navigation errors up to 2.5m RMSE. Furthermore, we identify MobileNetV3 Small as an optimal edge-ready model, achieving 90% classification accuracy with minimal latency, enabling drones to adapt flight strategies in real-time.

Introduction

Most autonomous drones utilize Visual Odometry (VO) to estimate their position and movement by analyzing consecutive camera frames. While highly effective in clear conditions, VO systems are vulnerable to environmental disturbances. Rain, in particular, introduces blur, reduces contrast, and creates physical droplets on the camera lens that obscure visual landmarks.

Existing research into rainy navigation has largely focused on autonomous ground vehicles, where cameras are protected by windshields and wipers. UAVs, however, often have exposed lenses, making them more susceptible to direct water accumulation. This study addresses the research gap by systematically quantifying how different rain intensities affect drone tracking and providing a computational path for drones to become "weather-aware."

Experimental Methodology

To study these effects, the researchers developed a custom, water-resistant hardware rig powered by an Intel NUC 11 and an Intel RealSense D435i depth camera. The VINS-Fusion algorithm was employed for stereo-based visual odometry.

Experiments were conducted in a controlled laboratory environment using a mechanical sprayer to simulate seven distinct conditions:

  1. Clear skies
  2. Vertical Rain (Low, Medium, Heavy)
  3. Slanting Rain (Low, Medium, Heavy)

Slanting rain was specifically designed to simulate the forward motion of a drone flying into precipitation. To enable real-time classification of these conditions, the team evaluated three "edge-ready" Deep Neural Networks (DNNs) known for low computational overhead: MobileNetV2, MobileNetV3 Small, and SqueezeNet.

The Rain-UAV Dataset

A core contribution of this work is the release of a massive, open-access dataset comprising approximately 335,000 real images (~12GB compressed). Unlike previous datasets that provide binary "rain/no rain" labels, this dataset is categorized into the seven specific intensity and angle classes mentioned above. Each class contains roughly 48,000 images, capturing the unique perspective of water droplets directly on a camera lens—a critical factor for UAV-specific computer vision tasks.

Quantifying Navigation Degradation

The study revealed that the physical angle of rain is a more significant factor in VO failure than simple intensity.

  • The Impact of Slanting Rain: Slanting rain is far more destructive than vertical rain. Because slanting rain hits the lens directly, it creates larger, more disruptive aggregations of water. In heavy slanting rain, the system experienced tracking errors (RMSE) of up to 2.5 meters. For a UAV operating in proximity to infrastructure, such an error is the difference between a successful mission and a catastrophic collision.
  • Restoring Time: The researchers introduced the concept of "restoring time"—the duration required for a VO system to recover its accuracy after rain stops. The study found that it takes approximately 33 seconds for the error to drop back below a safe 30cm threshold as the lens clears and the algorithm stabilizes.

Efficient Edge Classification Results

For a drone to adapt to weather, it must classify the environment without exhausting its limited battery and processing power. The results of the DNN evaluation are as follows:

  • Top Performer: MobileNetV3 Small emerged as the most efficient model, achieving 90% accuracy across the seven rain classes.
  • Speed and Latency: On the Intel NUC hardware, the model reached a frame rate of 93 FPS with a latency of only 10.7 ms. This allows the classifier to run concurrently with navigation tasks without impacting flight control loops.
  • Memory Footprint: The model requires only 1.28 MB of memory, making it highly suitable for resource-constrained Internet of Drones (IoD) devices.

Conclusion and Future Work

This research provides a quantitative foundation for developing "weather-aware" autonomous systems. By detecting rain intensity and angle in real-time, a drone's flight controller can automatically increase safety margins, switch to more robust navigation strategies (such as GPS-heavy modes), or initiate emergency landings when conditions become too hazardous.

While this study utilized a controlled indoor environment to ensure precise labeling, future work will focus on variable real-world lighting and diverse precipitation types like sleet or snow. Ultimately, a network of weather-aware drones could serve as a distributed sensing platform, providing hyper-local, real-time weather forecasting for smart city infrastructure.

Summary updated

BibTeX
@article{albanese2025is,
  author = {Andrea Albanese and Yanran Wang and Davide Brunelli and David Boyle},
  title = {{Is That Rain? Understanding Effects on Visual Odometry Performance for Autonomous UAVs and Efficient DNN-based Rain Classification at the Edge}},
  journal = {IEEE Internet of Things Journal},
  year = {2025},
  doi = {10.1109/jiot.2025.3547425},
}
Figure