基于解耦时空特征融合的小目标无人机检测OA

In response to the issues of missed and false detections when detecting small-sized drones at long distances in video,this paper proposes a Decoupled Spatiotemporal Feature Fusion Detection Network(DS2F-DN)which is built on the YOWOv2(You Only Watch Once version 2)framework and employs a differential resolution preprocessing,wherein high-definition key frames and low-resolution frame sequences are separately input into two-dimensional(2D)and three-dimensional(3D)convolutional branches for processing.This approach allows for the comprehensive utilization of spatial information from images and temporal information from video frame sequences.Specifically,we design a Wavelet Transform Decoupling Module(WTDM)within the 2D convolutional branch to decompose the high-and low-frequency information contained in the spatial features.In the 3D convolutional branch,we adopt transposed convolution to achieve multi-scaling of spatiotemporal features,and design a multi-scale feature interaction strategy to enhance the representation ability of spatiotemporal features.Furthermore,regarding the 2D spatial decoupled features output by WTDM and the spatiotemporally compressed features from the 3D convolutional branch,we imple-ment a parallel pointwise convolution structure in the channel encoder to achieve dimensionality reduction and feature fusion.Additionally,we introduce a scaled dot-product self-attention mechanism to improve the model's spatiotemporal modeling capabilities.Experimental results show that the DS2F-DN can achieve a detection accuracy of 71.51%mAP with a frame inference latency of 23.4 ms(approximately 42.7 FPS)on the drone video dataset,Drone-detection-dataset,which outperforms existing methods in overall performance and achieves high-precision real-time detection of small drone targets.