Real-time Aerial Detection and Reasoning on Embedded-UAVs
This work addresses the problem of enabling autonomous surveying and activity reporting for UAV applications, representing a domain-specific incremental improvement.
The paper tackles real-time pedestrian detection and activity recognition from UAV aerial imagery by developing a unified pipeline architecture that operates efficiently on embedded systems, achieving high accuracy and fast inference speed as demonstrated through empirical deployment in real-world environments.
We present a unified pipeline architecture for a real-time detection system on an embedded system for UAVs. Neural architectures have been the industry standard for computer vision. However, most existing works focus solely on concatenating deeper layers to achieve higher accuracy with run-time performance as the trade-off. This pipeline of networks can exploit the domain-specific knowledge on aerial pedestrian detection and activity recognition for the emerging UAV applications of autonomous surveying and activity reporting. In particular, our pipeline architectures operate in a time-sensitive manner, have high accuracy in detecting pedestrians from various aerial orientations, use a novel attention map for multi-activities recognition, and jointly refine its detection with temporal information. Numerically, we demonstrate our model's accuracy and fast inference speed on embedded systems. We empirically deployed our prototype hardware with full live feeds in a real-world open-field environment.