Introduction
Deep learning has significantly advanced robotic perception, yet state-of-the-art semantic segmentation models remain computationally prohibitive for Micro Aerial Vehicles (MAVs). These platforms operate under stringent Size, Weight, and Power (SWaP) constraints, typically relying on embedded modules like the NVIDIA Jetson rather than high-end desktop GPUs. Furthermore, standard datasets like Cityscapes do not represent the unique environments MAVs encounter, such as dark industrial infrastructure or mid-air drone-to-drone tracking. This work introduces MAVNet, an ultra-lightweight micro-network designed to bridge the gap between high-fidelity segmentation and real-time embedded performance.
Proposed Architecture: MAVNet
MAVNet is inspired by the ERFNet architecture but optimized for extreme efficiency, utilizing approximately 400x fewer parameters. The core of the network is the Depth-wise Feature Aggregation Block (DWFab). This block leverages depth-wise separable convolutions to minimize computational overhead while employing dilated convolutions to expand the receptive field. This allows the network to capture spatial context—crucial for identifying objects of varying scales—without increasing parameter count.
To further enhance speed, MAVNet employs aggressive downsampling in its initial layers, reducing memory consumption early in the pipeline. The decoder is intentionally kept shallow, using nearest-neighbor upsampling to reconstruct the segmentation map while avoiding the "checkerboard" artifacts common in more complex deconvolutional layers. To address class imbalance—where a target drone may occupy a negligible fraction of the total pixels—the model is trained using Focal Loss, which prioritizes learning from difficult, sparsely represented pixels.
Datasets and Benchmarks
The efficacy of MAVNet was evaluated on two novel, publicly available benchmarks introduced by the authors:
- MAV Segmentation Dataset: Focused on indoor drone-to-drone tracking, featuring cluttered backgrounds and significant variance in target size.
- Penstock Dataset: Comprising imagery from the interior of hydroelectric dam penstocks, requiring the identification of rivets, corrosion, and water under poor illumination and noisy conditions.
Results and Discussion
MAVNet demonstrates remarkable efficiency, utilizing only 4,300 parameters—roughly 1,800x fewer than UNet and 400x fewer than ERFNet. Despite this reduction, it achieves competitive performance. On a Jetson TX2, MAVNet operates at 4.4 FPS, while reaching 9 FPS on the more modern Jetson Xavier and 48 FPS on a desktop GPU.
In terms of accuracy, MAVNet achieved a Mean IoU of 36.24% on the Penstock dataset, outperforming significantly larger models. For tracking tasks, the authors emphasize Centroid Distance as a more practical metric than Intersection over Union (IoU) for localization; MAVNet achieved a superior centroid distance of 3.36 pixels, making it highly reliable for actual flight maneuvers.
While highly effective at localization, the model exhibits a higher false-positive rate compared to deeper architectures, likely due to its limited capacity for nuanced feature extraction in complex scenes.
Conclusion
MAVNet proves that for specialized MAV tasks, architectural intelligence can compensate for a lack of raw parameter scale. By optimizing for receptive field and utilizing efficient convolution blocks, MAVNet enables real-time semantic perception on SWaP-constrained hardware. Future work will explore further optimization via tools like NVIDIA TensorRT to enable even more agile autonomous flight.
Summary updated