Drone-Assisted Behavior Recognition via Key Frame Extraction for Efficient 5G Communication

Hui Li, Tongao Ge, Ruidi Ma, Ying Guo, Guowei Zhao, Jiaming Pei, Lukun Wang, Lingwei Xu

Abstract: Video behavior recognition based on drone + 5G technology has important research value in public security, intelligent transportation, telemetry and remote sensing, etc. However, the high redundancy of videos taken by drones leads to a large amount of calculation for its auxiliary behavior recognition algorithm, the current drone platform has limited computational power and cannot process data well. Therefore, we propose drone-assisted behavior recognition via key frame extraction based on AFCN, V-Stream and S-Stream (AFVS) for efficient 5G communication. Firstly, a key frame extraction algorithm of full convolutional Network and attention mechanism, named AFCN is proposed to quickly extract important information from drone video. Secondly, VS-Stream with dual stream architecture is designed to process both video stream and skeleton stream. In the video stream, we designed the fast and slow path network to obtain the deep features that affect the target behavior, finally achieve high-accuracy behavior recognition. Our experiment results demonstrate that AFVS solves the defects of the drone-assisted behavior recognition algorithm in natural scenes, and greatly improves its performance.