부산대학교 도서관

A few characteristics of advanced commercial drones have attracted increasing attention in recent years. Because of their ability to carry payloads bypassing ground security, there is a greater possibility of small drones being exploited for illegitimate operations. The drone tracking and surveillance is critical to prevent security breaches like these. Small drones and birds have similar appearances in complicated backgrounds and make it difficult to recognize drone in surveillance videos. Manual surveillance and drone detection are very complicated and tedious process. Therefore, an automated method is required to distinguish between drone and birds. In this paper, focus measure operators (FMOs)-based method is developed for drone detection. The five FMO parameters are computed on each frame of every video. Feature ranking is done to identify the significant feature for drone identification followed by classification which is performed using random forest classifier. Efficacy of the proposed method is evaluated on Drone-vs-Bird Detection Challenge at IEEE AVSS2021 dataset which is provided by Workshop on Small-Drone Surveillance, Detection and Counteraction Techniques (WOSDETC) supported by the SafeShore Consortium. The proposed method achieved average accuracy 94.15% with sensitivity 96.69% for identifying drones with drone present (DP) versus neither drone nor bird present (NDNBP) (two class), accuracy 93.60% with sensitivity 96.20% DP versus both bird and drone present (BBDP) versus NDNBP (three class), accuracy 92% with sensitivity 95% DP versus bird present versus BBDP versus NDNBP (four class) for moving camera recordings. Average accuracy for two class, three class, and four class classification system with moving as well as stationary camera is 95.18%, 93.20%, and 94%, respectively, which are significantly high and can be utilized for drone identification.