Early warning of small civilian UAVs is a critical issue in the field of public security and is a challenging task in object detection. Video object detection methods based on deep convolution neural networks have attracted increasing attention over recent years. The object detection algorithm designed for static images can hardly be used directly in video object detection, due to motion blur, out-of-focus, etc. Traditional methods mainly apply temporal information to address this problem. In this paper, we proposed a method based on optical flow for UAV video object detection, employing optical flow information and the object appearance visual features in different stages. We establish a motion enhancement module to utilize motion features of moving objects while suppressing complex backgrounds. In addition, direction information is used to associate detection boxes of adjacent frames, improving the stability of the detection. The results of the experiments on the Anti-UAV2021 Challenge dataset show that the average precision of our model is 89.8%, significantly higher than that of FGFA, SELSA, and YOLOv5 models. The experimental results also demonstrate that the motion enhancement module and direction post-processing module can improve the recall rate of small UAVs under complex backgrounds.
Optical Flow-Guided Deep Convolutional Neural Networks for UAV Detection in Infrared Videos
2022-11-11
671035 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Deep Convolutional Neural Network For Human Detection And Tracking In FLIR Videos (AIAA 2016-1412)
British Library Conference Proceedings | 2016
|Smile detection in the wild with deep convolutional neural networks
British Library Online Contents | 2017
|