Proceedings Article | 26 May 2023
KEYWORDS: Target detection, Optical flow, Video, Feature extraction, Optical networks, Unmanned aerial vehicles, Visualization, Video processing
Motion target detection is a prerequisite for road monitoring, motion target tracking, instance segmentation and other tasks. UAV video images are easily affected by some unavoidable factors in the acquisition process, such as wind interference and own motion during the shooting process can lead to image background changes, target scale changes and intermittent motion, making the motion target detection task more challenging. To address the problems of poor accuracy of existing UAV video motion target detection methods based on deep optical flow networks and the limitation of target detection performance in complex scenes due to the complex and diverse features of UAV video data, this paper proposes a new UAV video motion target detection method based on optical flow networks. Firstly, a convolutional structure reparameterization method is used in the coding part to further fuse detailed and semantic information to improve the feature expression capability of video images; secondly, the self-attentive global motion feature enhancement module proposed in this paper is introduced to improve the network's ability to extract global information and better combine contextual information to achieve more accurate optical flow estimation; finally, the optical flow threshold segmentation is used to obtain different motion target detection results for different scenes by optical flow threshold segmentation. In this paper, three sets of low-altitude UAV video data from different scenes are selected for experiments on the public dataset AU-AIR2019, and the experimental results prove that the proposed method can achieve better motion target detection results in single-target, multi-target and occluded target scenes, and it is better than the current mainstream optical flow networks: FlowNet1, PWC-Net, HD3, PWC-Net and HD3 on the public dataset FlyingChairs. PWC-Net, HD3, GMA metrics EPE (end point error, EPE) on the public dataset FlyingChairs, and improves the RAFT by 0.10 over the benchmark network in this paper, which effectively improves the accuracy of UAV video motion target detection by deep optical flow networks.
