Improved Target Detection in UAV Photographic Images Using YOLOv7-Tiny

Abstract

The purpose of this paper is to explore the effective monitoring and countermeasures of low-altitude UAVs through multi-sensor coordination so as to escort the sustainable development of a “low-altitude economy”. The core work of this paper centers on multi-source imaging sensing, precise positioning, identification, and behavioral feature extraction of low-altitude UAV targets. Different types of sensors, including visual sensors, radar sensors, sound sensors, etc., are integrated to build a multi-source sensing system, which realizes all-round and multi-angle monitoring of low-altitude UAVs. The improved YOLOv7-Tiny model achieves accurate detection of UAV targets based on this basis. In order to further improve the intelligence level of monitoring and countermeasures, the actuator-evaluator framework of reinforcement learning algorithms is introduced to construct a reinforcement learning framework of “multi-source perception-intelligent cognition-assisted decision-making”. The maximum detection accuracy of the YOLOv7-Tiny-NET model is 0.837, and the model size of the YOLOv7-Tiny-NET model is reduced by 3.52MB and 37.8 f/s increases the detection speed compared with SAG-YOLOv5s. The maximum success rate of the autonomous decision-making algorithm of UAV can be up to 78%~88% when making autonomous decisions on dynamic target tasks. Through the accurate monitoring and intelligent countermeasures of low-altitude drones, it can effectively prevent unmanned aircraft from flying illegally, protect personal privacy, and maintain public safety, thus promoting the sustainable development of a “low-altitude economy” on a healthy and orderly track.